Coursedia

Connecting minds with knowledge, one course at a time.

Home Wikipedia Summaries Articles

Artificial Intelligence: A Detailed Educational Resource

ai, machine learning, deep learning, nlp

An in-depth guide to artificial intelligence, covering its history, goals, techniques, and applications. Learn about AI subfields, machine learning, deep learning, natural language processing, and more.

Read the original article here.

Introduction to Artificial Intelligence

Artificial intelligence (AI) is a transformative field within computer science focused on creating systems capable of intelligent behavior. These systems are designed to mimic cognitive functions associated with human minds, such as:

Artificial Intelligence (AI) Definition: The capability of a computational system to perform tasks typically associated with human intelligence, such as learning, reasoning, problem-solving, perception, and decision-making. It is a field of research in computer science that develops and studies methods and software that enable machines to perceive their environment and use learning and intelligence to take actions that maximize their chances of achieving defined goals. Machines exhibiting such capabilities are referred to as AIs.

AI is not just about mimicking human intelligence, but also about understanding and replicating the underlying principles of intelligence in any form, whether human or machine.

High-Profile Applications and Everyday AI

AI is rapidly becoming integrated into various aspects of modern life, often powering technologies we use daily:

It’s important to note that many AI applications are so seamlessly integrated into our lives that we often don’t recognize them as AI. This phenomenon, sometimes called the AI effect, occurs because once an AI technology becomes commonplace and useful, it is no longer perceived as “artificial intelligence” but simply as a standard technology.

AI Effect Definition: The phenomenon where once a technology powered by artificial intelligence becomes widely adopted and useful, it ceases to be labeled as “AI” and is simply considered a standard feature or technology. This is often because the initial “wow” factor of AI diminishes as it becomes integrated into everyday applications.

This constant evolution and integration of AI into everyday tools and systems highlights its pervasive and transformative nature.

Subfields and Goals of AI Research

The field of AI is vast and encompasses numerous subfields, each focused on specific aspects of intelligence and employing diverse tools and techniques. Traditional goals of AI research include:

A long-term and ambitious goal within AI is achieving General Intelligence.

General Intelligence (Artificial General Intelligence - AGI) Definition: The hypothetical ability of an AI system to perform any intellectual task that a human being can, and at a level comparable to or exceeding human capabilities. AGI aims to create AI that is not limited to specific tasks but possesses broad, versatile intelligence, similar to human intelligence.

To achieve these diverse goals, AI researchers draw upon a wide array of techniques and methodologies from various disciplines, including:

Furthermore, AI research is interdisciplinary, drawing inspiration and knowledge from fields like:

Historical Context: Optimism, Winters, and the AI Boom

Artificial intelligence emerged as a formal academic discipline in 1956. The field has experienced cyclical patterns of enthusiasm and progress, interspersed with periods of stagnation and funding cuts, known as AI winters.

AI Winter Definition: A period of reduced funding, interest, and progress in the field of artificial intelligence. AI winters typically follow periods of inflated expectations and unmet promises regarding AI capabilities, leading to disillusionment and decreased investment.

The most recent resurgence of AI, often termed the AI boom, began around 2012, marked by significant breakthroughs in deep learning. This resurgence was further accelerated by the development of the transformer architecture around 2017, leading to rapid advancements in generative AI and unprecedented levels of investment in the field in the early 2020s.

The rise of powerful generative AI, capable of creating sophisticated text, images, and other content, has brought both immense excitement and serious concerns. While generative AI offers remarkable potential, it has also exposed unintended consequences and potential harms, prompting crucial discussions about:

The current AI boom is a period of rapid innovation and widespread adoption, but it is also a time of critical reflection and proactive planning to navigate the complex challenges and opportunities that AI presents.

Goals of Artificial Intelligence

The overarching goal of creating artificial intelligence is often broken down into more manageable subproblems. These subproblems represent specific traits and capabilities that researchers believe are essential for intelligent systems to possess. The following sections detail some of the most prominent goals driving AI research.

Reasoning and Problem-Solving

Reasoning and problem-solving are fundamental aspects of intelligence. Early AI research focused on developing algorithms that could mimic the step-by-step logical reasoning humans employ when solving puzzles or making deductions. These early systems aimed to replicate deliberate, conscious reasoning processes.

Reasoning Definition (in AI Context): The process by which an AI system uses logical inference and deduction to draw conclusions from given information or premises. Reasoning allows AI to solve problems, answer questions, and make informed decisions.

Problem-Solving Definition (in AI Context): The process by which an AI system identifies, analyzes, and resolves issues or challenges. This involves using reasoning, knowledge, and planning to find effective solutions to defined problems.

In the late 1980s and 1990s, research expanded to address the complexities of real-world scenarios, incorporating methods for dealing with uncertain or incomplete information. This involved utilizing concepts from probability and economics to make informed decisions even when faced with ambiguity.

However, many of these traditional algorithms encounter a significant challenge known as “combinatorial explosion” when applied to larger, more complex problems.

Combinatorial Explosion Definition: A phenomenon where the number of possible states or paths in a problem space grows exponentially with the size of the problem. This can make exhaustive search algorithms computationally infeasible, as the time and resources required to explore all possibilities become prohibitively large.

As problems become larger, the computational resources and time required to solve them using these algorithms increase exponentially, making them impractical for many real-world applications. Furthermore, human reasoning often relies on fast, intuitive judgments rather than step-by-step deduction, especially in everyday situations.

Accurate and efficient reasoning remains a significant unsolved problem in AI. Researchers are continuously exploring new approaches to develop reasoning systems that can handle complexity, uncertainty, and operate with the speed and intuitiveness of human cognition. This includes exploring techniques inspired by human intuition and developing more efficient search and inference algorithms.

Knowledge Representation

Knowledge representation is crucial for enabling AI programs to reason intelligently and make deductions about the real world. It involves designing formal ways to store and organize information that AI systems can use to understand and interact with their environment.

Knowledge Representation Definition: The field of AI concerned with designing and implementing methods to symbolically represent information in a way that can be used by computer systems for reasoning, problem-solving, and learning. It focuses on how to encode knowledge about the world in a format that AI agents can understand and manipulate.

Knowledge engineering is the process of building and maintaining these knowledge representations for AI systems. Effective knowledge representation is essential for various AI applications, including:

A knowledge base is the actual repository of knowledge, organized in a format that AI programs can access and utilize.

Knowledge Base Definition: A structured collection of facts, rules, and information about a specific domain or the world in general, represented in a way that can be used by an AI system for reasoning, problem-solving, and decision-making.

An ontology defines the vocabulary and relationships within a specific domain of knowledge. It specifies the types of objects, concepts, relations, and properties relevant to that domain.

Ontology Definition (in AI Context): A formal representation of knowledge as a set of concepts within a domain and the relationships between those concepts. In AI, ontologies are used to define the vocabulary and structure of a knowledge base, providing a framework for organizing and understanding information.

Knowledge bases must be capable of representing diverse types of information, including:

Several significant challenges exist in knowledge representation:

Overcoming these challenges is crucial for creating AI systems that can reason effectively in complex, real-world environments. Researchers are exploring various approaches, including large language models and knowledge graph technologies, to address these issues.

Planning and Decision-Making

Planning and decision-making are essential capabilities for intelligent agents that operate in dynamic environments. An agent in AI is any entity that can perceive its surroundings and take actions within it.

Agent Definition (in AI Context): In artificial intelligence, an agent is an entity (software or physical) that can perceive its environment through sensors and act upon that environment through effectors or actuators. Agents are designed to be autonomous and goal-directed, making decisions to achieve specific objectives.

A rational agent is an agent that has goals or preferences and takes actions to maximize the likelihood of achieving those goals or satisfying those preferences.

Rational Agent Definition: An agent in artificial intelligence that acts in a way that is expected to maximize its goal achievement, based on its perceptions, knowledge, and beliefs. Rational agents make decisions to optimize outcomes according to their defined objectives and preferences.

Automated planning focuses on situations where an agent has a specific goal and needs to devise a sequence of actions to reach that goal.

Automated Planning Definition: A subfield of AI concerned with developing algorithms and techniques for agents to automatically generate sequences of actions (plans) to achieve specific goals in a given environment. Automated planning systems typically operate in environments with well-defined states, actions, and goals.

Automated decision-making deals with scenarios where an agent has preferences among different situations and needs to choose actions that lead to more desirable outcomes.

Automated Decision-Making Definition: A subfield of AI that focuses on developing systems that can automatically make choices or decisions, often in complex and uncertain environments. Decision-making agents consider various options, evaluate potential outcomes, and select the course of action that best aligns with their goals or preferences.

In decision-making, agents often use the concept of utility to quantify their preferences. Utility is a numerical value assigned to each possible situation, representing how much an agent prefers that situation.

Utility Definition (in Decision-Making): A numerical measure of preference or desirability assigned to different outcomes or situations. In AI decision-making, utility functions are used to quantify the value or satisfaction an agent derives from different states or results, guiding its choices towards maximizing expected utility.

For each potential action, a rational agent calculates the expected utility. This is the sum of the utilities of all possible outcomes of that action, weighted by the probability of each outcome occurring. The agent then chooses the action that maximizes its expected utility.

Expected Utility Definition: In decision theory and AI, expected utility is the weighted average of the utilities of all possible outcomes of an action, where the weights are the probabilities of each outcome occurring. Rational agents choose actions that maximize their expected utility.

Classical planning assumes that the agent has complete knowledge of the environment and the effects of its actions are deterministic (predictable).

Classical Planning Definition: A type of automated planning that assumes a deterministic, fully observable environment. In classical planning, the agent knows the initial state, the set of possible actions, and the exact outcome of each action. The goal is to find a sequence of actions that leads to a desired goal state.

However, real-world problems often involve uncertainty. Agents may not be certain about their current situation (“unknown” or “unobservable” environments) or the consequences of their actions (“non-deterministic” environments). In such cases, agents must make probabilistic guesses and adapt their plans based on new information.

Agents may also need to learn or refine their preferences, especially when interacting with other agents or humans. Inverse reinforcement learning is a technique that allows agents to learn preferences from observing the behavior of others.

Inverse Reinforcement Learning (IRL) Definition: A type of machine learning where the goal is to infer the reward function or preferences of an agent by observing its behavior. Instead of learning to act to maximize a given reward, IRL aims to understand what rewards or goals are driving the observed behavior.

Information value theory provides a framework for evaluating the benefit of gathering more information to improve decision-making, helping agents decide whether to explore or exploit their current knowledge.

Information Value Theory Definition: A branch of decision theory that quantifies the value of acquiring additional information before making a decision. It helps agents determine whether the potential benefit of reducing uncertainty through information gathering outweighs the cost of acquiring that information.

Markov decision processes (MDPs) are a mathematical framework for modeling sequential decision-making in uncertain environments. An MDP includes a transition model that describes the probabilities of state transitions given actions and a reward function that defines the utility of states and the cost of actions. A policy specifies the action an agent should take in each possible state.

Markov Decision Process (MDP) Definition: A mathematical framework for modeling decision-making in stochastic (probabilistic) environments. An MDP is defined by states, actions, transition probabilities between states given actions, and rewards associated with states or state-action pairs. MDPs are used to find optimal policies that maximize cumulative rewards over time.

Policy Definition (in MDPs): A mapping from states to actions that specifies what action an agent should take in each state. In reinforcement learning and MDPs, a policy is the strategy an agent follows to choose actions. Optimal policies aim to maximize cumulative rewards over time.

Game theory provides tools for analyzing the rational behavior of multiple interacting agents, which is crucial for AI systems that operate in multi-agent environments or interact with humans.

Game Theory Definition: A mathematical framework for studying strategic interactions between rational agents (players) in situations where the outcome of each agent’s actions depends on the actions of others. Game theory is used in AI to design agents that can reason about and interact effectively with other agents, including humans.

These techniques and frameworks are essential for developing AI systems that can plan, make decisions, and act effectively in complex and uncertain real-world scenarios, whether it’s controlling robots, managing resources, or interacting with humans in collaborative tasks.

Learning

Machine learning is a core subfield of AI focused on developing algorithms that enable computers to learn from data without being explicitly programmed. It has been a central part of AI since its inception and is now a driving force behind many AI advancements.

Machine Learning Definition: The subfield of AI that focuses on developing algorithms and techniques that allow computer systems to learn from data without being explicitly programmed. Machine learning enables AI systems to improve their performance on a specific task over time through experience and data analysis.

There are several major types of machine learning:

Computational learning theory provides a theoretical framework for analyzing the performance and properties of machine learning algorithms, considering factors such as:

Machine learning is continuously evolving, with new algorithms and techniques being developed to address increasingly complex and diverse problems, driving progress across various domains of AI.

Natural Language Processing

Natural language processing (NLP) is a critical field within AI that focuses on enabling computers to understand, process, and generate human languages, such as English, Spanish, or Mandarin. NLP aims to bridge the communication gap between humans and computers, allowing for more natural and intuitive interactions.

Natural Language Processing (NLP) Definition: A branch of artificial intelligence concerned with enabling computers to understand, interpret, and generate human languages. NLP encompasses a wide range of tasks, including speech recognition, natural language understanding, natural language generation, machine translation, and information retrieval.

Key problems and tasks within NLP include:

Early approaches to NLP, often based on Noam Chomsky’s generative grammar and semantic networks, faced challenges with word-sense disambiguation.

Word-Sense Disambiguation Definition: The NLP task of determining the correct meaning or sense of a word in a given context when the word has multiple possible meanings. For example, the word “bank” can refer to a financial institution or the edge of a river. Word-sense disambiguation is crucial for accurate natural language understanding.

Unless restricted to very narrow domains called “micro-worlds,” these systems struggled with the vast amount of commonsense knowledge required to understand language in its full complexity.

Micro-world Definition (in Early AI/NLP): A simplified, restricted domain of knowledge used in early AI research, particularly in natural language processing and reasoning systems. Micro-worlds allowed researchers to focus on specific aspects of intelligence and language understanding in a controlled, manageable setting.

Margaret Masterman, a pioneer in computational linguistics, emphasized the importance of meaning over grammar and proposed using thesauri (structured collections of words and their meanings) rather than dictionaries as the foundation for computational language understanding.

Thesaurus Definition (in NLP Context): A structured collection of words and their meanings, organized by semantic relationships such as synonyms, antonyms, and hyponyms. In NLP, thesauri can be used to enhance word-sense disambiguation, information retrieval, and text understanding by providing semantic context and relationships between words.

Modern NLP has been revolutionized by deep learning techniques, including:

In 2019, GPT models began to demonstrate remarkable text generation capabilities, and by 2023, they achieved human-level performance on various standardized tests, including the bar exam, SAT, and GRE, showcasing the rapid progress in NLP driven by deep learning.

Perception

Machine perception is the ability of AI systems to interpret sensory input from the real world, similar to how humans and animals perceive their environment through senses. This involves using data from sensors to deduce meaningful aspects of the world.

Machine Perception Definition: The ability of an artificial system to use sensory input (e.g., from cameras, microphones, sensors) to interpret and understand the world around it. Machine perception aims to replicate human-like sensory processing, enabling AI systems to extract meaningful information from raw sensory data.

Common sensors used in machine perception include:

Computer vision is a major subfield of machine perception specifically focused on analyzing visual input from cameras and images.

Computer Vision Definition: A field of artificial intelligence that enables computers to “see” and interpret visual information from the world, similar to human vision. Computer vision tasks include image recognition, object detection, image segmentation, and scene understanding.

Key tasks in machine perception and computer vision include:

Machine perception is crucial for enabling AI systems to interact effectively with the physical world, whether it’s for autonomous driving, robotic manipulation, or understanding visual and auditory information in multimedia content.

Social Intelligence

Social intelligence in AI focuses on developing systems that can understand and interact with humans in socially intelligent ways. This includes recognizing, interpreting, processing, and even simulating human emotions, moods, and social cues.

Social Intelligence (in AI Context): The ability of an AI system to understand and navigate social situations, interact effectively with humans, and exhibit human-like social skills. Social intelligence in AI involves recognizing emotions, understanding social cues, engaging in natural language communication, and adapting behavior to social contexts.

Affective computing is a key area within social intelligence that specifically deals with emotion recognition and processing in AI systems.

Affective Computing Definition: An interdisciplinary field that focuses on developing computer systems that can recognize, interpret, process, and simulate human emotions, moods, and feelings. Affective computing aims to bridge the gap between human emotions and computational systems, enabling more natural and empathetic human-computer interaction.

Examples of social intelligence and affective computing in AI include:

While AI systems are increasingly capable of exhibiting some aspects of social intelligence, it’s important to note that current systems often give naive users an unrealistic conception of their actual intelligence. The conversational abilities and simulated emotions can sometimes be misleading, making it seem as though the AI possesses deeper understanding or sentience than it actually does.

Despite these limitations, social intelligence and affective computing are advancing, leading to more engaging and user-friendly AI systems that can better understand and respond to human emotional states.

General Intelligence

Artificial General Intelligence (AGI) represents the long-term and ambitious goal of creating AI systems with human-level intelligence. An AGI system should be able to perform any intellectual task that a human being can, exhibiting broad and versatile intelligence.

Artificial General Intelligence (AGI) Definition: The hypothetical ability of an AI system to perform any intellectual task that a human being can, and at a level comparable to or exceeding human capabilities. AGI aims to create AI that is not limited to specific tasks but possesses broad, versatile intelligence, similar to human intelligence.

AGI is distinct from narrow AI, which is designed to excel at specific tasks. AGI would possess a wider range of cognitive abilities, including:

Achieving AGI is considered a grand challenge in AI research, and it remains a topic of ongoing research and debate. While current AI systems excel in narrow domains, creating a truly general-purpose AI with human-level intelligence is a complex and multifaceted endeavor that requires significant advancements in our understanding of intelligence and computation.

Techniques in Artificial Intelligence

AI research employs a diverse range of techniques to achieve its various goals. These techniques can be broadly categorized and are often combined to create more sophisticated AI systems.

Search and Optimization

Search and optimization are fundamental techniques in AI for solving problems and finding optimal solutions. AI systems often explore a vast space of possibilities to find the best course of action or the most suitable answer. There are two primary categories of search used in AI: state space search and local search.

State space search is a technique used to find a path from an initial state to a goal state by exploring a tree of possible states. It is particularly useful for problems that can be represented as a sequence of states and transitions between them.

State Space Search Definition: A problem-solving technique in AI that involves exploring a space of possible states to find a path from an initial state to a desired goal state. State space search algorithms systematically navigate through states and transitions, using heuristics or search strategies to efficiently find a solution.

Planning algorithms often utilize state space search to find sequences of actions to achieve goals. Means-ends analysis is a common approach in planning where the AI system tries to reduce the difference between the current state and the goal state by applying relevant actions.

Means-Ends Analysis Definition: A problem-solving strategy in AI where the problem solver identifies the difference between the current state and the goal state (the “ends”) and then selects actions (the “means”) that reduce this difference. Means-ends analysis is often used in planning and problem-solving systems.

However, exhaustive searches, which explore all possible states, are often impractical for real-world problems due to combinatorial explosion. The search space can grow astronomically large, making the search process too slow or even impossible to complete.

To address this, AI systems often use heuristics, or “rules of thumb,” to guide the search and prioritize choices that are more likely to lead to a goal state.

Heuristic Definition (in Search Algorithms): In the context of search algorithms, a heuristic is a rule of thumb or a practical method that is used to guide the search process towards promising areas of the state space. Heuristics are not guaranteed to find the optimal solution but are designed to improve search efficiency and find good solutions in a reasonable amount of time.

Adversarial search is a specialized form of state space search used in game-playing programs, such as chess or Go. It involves searching through a tree of possible moves and countermoves, considering the opponent’s actions, to find a winning strategy.

Adversarial Search Definition: A type of search algorithm used in game-playing AI, particularly for two-player games like chess or Go. Adversarial search involves exploring a game tree where nodes represent game states, and branches represent moves made by players. The algorithm searches for optimal moves by considering the opponent’s possible responses and aiming to maximize its own chances of winning while minimizing the opponent’s chances.

Local search techniques use mathematical optimization to find solutions to problems by starting with an initial guess and iteratively refining it to improve its quality.

Local Search Definition: A class of optimization algorithms used in AI and computer science that iteratively improve a candidate solution by exploring its “neighborhood” in the search space. Local search algorithms start with an initial solution and repeatedly move to a better neighboring solution until a local optimum is found.

Gradient descent is a widely used local search algorithm for optimizing numerical parameters. It incrementally adjusts parameters in the direction that minimizes a loss function, which measures the error or cost of the current solution.

Gradient Descent Definition: An iterative optimization algorithm used to find the minimum of a function. In machine learning and AI, gradient descent is commonly used to train models by adjusting model parameters (weights) in the direction of the negative gradient of the loss function. This process iteratively reduces the loss and improves the model’s performance.

Backpropagation is a specific algorithm, based on gradient descent, used to train neural networks. It calculates the gradients of the loss function with respect to the network’s weights, allowing for efficient weight updates to improve network performance.

Backpropagation Definition: A supervised learning algorithm used to train artificial neural networks. Backpropagation calculates the gradients of the loss function with respect to the network’s weights using the chain rule of calculus. These gradients are then used to update the weights in the direction that reduces the loss, iteratively improving the network’s performance.

Evolutionary computation is another type of local search inspired by biological evolution. It iteratively improves a set of candidate solutions by applying “mutation” and “recombination” operations and selecting the “fittest” solutions to survive to the next generation.

Evolutionary Computation Definition: A family of optimization algorithms inspired by biological evolution, such as genetic algorithms and evolutionary strategies. Evolutionary computation methods maintain a population of candidate solutions and iteratively improve them through processes like selection, mutation, and recombination, mimicking natural selection to find optimal or near-optimal solutions.

Swarm intelligence algorithms, such as particle swarm optimization (inspired by bird flocking) and ant colony optimization (inspired by ant trails), are used for distributed search processes, where multiple agents coordinate to find solutions.

Swarm Intelligence Definition: A type of distributed AI approach inspired by the collective behavior of social insects, such as ants, bees, or birds. Swarm intelligence algorithms use a population of simple agents that interact locally with each other and their environment, leading to emergent, complex, and intelligent global behavior.

Particle Swarm Optimization (PSO) Definition: A swarm intelligence algorithm inspired by the social behavior of bird flocking or fish schooling. PSO maintains a population of particles, each representing a candidate solution, and iteratively updates their positions in the search space based on their own best-found solution and the best solution found by the swarm.

Ant Colony Optimization (ACO) Definition: A swarm intelligence algorithm inspired by the foraging behavior of ants. ACO algorithms use a population of artificial ants that iteratively construct solutions by traversing a graph. Ants deposit pheromone trails as they move, and subsequent ants are more likely to follow paths with higher pheromone concentrations, leading to the discovery of good solutions.

These search and optimization techniques are essential tools in AI for solving problems ranging from planning and scheduling to machine learning model training and game playing.

Logic

Formal logic is a foundational technique in AI for representing knowledge and performing reasoning. It provides a precise and unambiguous way to express facts, rules, and relationships, and allows for deductive inference.

Formal Logic Definition: A system of reasoning based on formal languages, rules of inference, and axioms. In AI, formal logic is used for knowledge representation, reasoning, and problem-solving. It provides a precise and unambiguous way to represent information and derive new conclusions through logical deduction.

Formal logic comes in two primary forms:

Deductive reasoning in logic is the process of deriving new statements (conclusions) from existing statements (premises) that are assumed to be true. Proofs in logic can be structured as proof trees, where nodes are labeled by sentences, and children nodes are connected to parent nodes by inference rules.

Deductive Reasoning Definition (in Logic): The process of reasoning from general premises to specific conclusions. Deductive reasoning uses logical rules to derive conclusions that are guaranteed to be true if the premises are true.

Proof Tree Definition (in Logic): A tree-like structure used to represent a logical proof. In a proof tree, the root node is the conclusion to be proven, the leaf nodes are axioms or premises, and each internal node is derived from its children nodes using inference rules.

Problem-solving in a logical framework can be viewed as searching for a proof tree where the root is a solution to the problem, and the leaves are premises or axioms.

Horn clauses are a restricted form of logic that are particularly efficient for automated reasoning. Problem-solving with Horn clauses can be done by reasoning forwards from premises or backwards from the goal.

Horn Clause Definition: A clause in logic that has at most one positive literal (non-negated atom). Horn clauses are used in logic programming and automated reasoning because they allow for efficient inference algorithms, such as forward chaining and backward chaining.

Resolution is a powerful, axiom-free rule of inference used in the clausal form of first-order logic. Problem-solving using resolution involves proving a contradiction from premises that include the negation of the problem to be solved.

Resolution Definition (in Logic): A single inference rule in logic used for automated theorem proving, particularly in clausal form first-order logic. Resolution combines two clauses that contain complementary literals to derive a new clause. By repeatedly applying resolution, it can be used to prove contradictions and establish the validity of logical arguments.

While inference in both Horn clause logic and first-order logic can be theoretically undecidable and computationally intractable in general, backward reasoning with Horn clauses is Turing complete and forms the basis of logic programming languages like Prolog.

Turing Complete Definition: A system of computation that is capable of simulating any Turing machine. Turing completeness is a measure of computational power, indicating that a system can perform any computation that can be performed by any other computer, given sufficient resources like time and memory.

Fuzzy logic is an extension of classical logic that allows for degrees of truth between 0 and 1, rather than just true or false. It is useful for handling vague and imprecise information.

Fuzzy Logic Definition: An extension of classical logic that allows for degrees of truth, ranging from 0 (completely false) to 1 (completely true), rather than just binary true or false values. Fuzzy logic is used to model and reason with imprecise, vague, or uncertain information, often encountered in real-world situations.

Non-monotonic logics, including logic programming with negation as failure, are designed to handle default reasoning, where conclusions can be revised in light of new information.

Non-monotonic Logic Definition: A type of logic where adding new premises can invalidate previously derived conclusions. Non-monotonic logics are used to model default reasoning, common sense reasoning, and situations where knowledge is incomplete or evolving.

Formal logic provides a rigorous foundation for knowledge representation and reasoning in AI, enabling systems to perform logical inference, problem-solving, and knowledge-based reasoning.

Probabilistic Methods for Uncertain Reasoning

Many real-world AI problems involve uncertainty and incomplete information. Probabilistic methods provide tools to handle this uncertainty and make rational decisions under these conditions. AI researchers have adapted techniques from probability theory and economics to address uncertain reasoning.

Precise mathematical tools have been developed for decision theory, decision analysis, and information value theory to analyze how agents can make choices and plan in uncertain environments. These tools include models like Markov decision processes (MDPs), dynamic decision networks, game theory, and mechanism design.

Decision Theory Definition: A branch of mathematics and economics concerned with the study of decision-making under uncertainty. Decision theory provides frameworks for analyzing choices, evaluating risks and rewards, and making rational decisions based on probabilities and preferences.

Decision Analysis Definition: A systematic approach to decision-making that uses decision theory principles and techniques to analyze complex decisions, particularly under uncertainty. Decision analysis involves structuring decisions, assessing probabilities and values, and evaluating different alternatives to inform optimal choices.

Dynamic Decision Network (Influence Diagram) Definition: A graphical representation of decision-making problems under uncertainty. Dynamic decision networks extend Bayesian networks by including decision nodes and utility nodes, representing decisions, uncertain variables, and preferences or values. They are used to model sequential decision-making and planning in uncertain environments.

Mechanism Design Definition: A field of economics and game theory that focuses on designing rules or mechanisms to achieve desired outcomes in situations where multiple agents interact with potentially conflicting interests. Mechanism design is used to create systems that incentivize agents to behave in a way that leads to socially desirable outcomes.

Bayesian networks are a powerful tool for probabilistic reasoning. They can be used for:

Probabilistic algorithms, such as hidden Markov models (HMMs) and Kalman filters, are also used for filtering, prediction, smoothing, and finding explanations for streams of data, which is crucial for perception systems that analyze processes evolving over time.

Hidden Markov Model (HMM) Definition: A statistical model used to represent systems that evolve over time, where the system’s state is hidden or unobservable, but observations related to the state are available. HMMs are used in AI for tasks such as speech recognition, sequence prediction, and time series analysis.

Kalman Filter Definition: An algorithm used for estimating the state of a dynamic system from noisy measurements. The Kalman filter is an optimal linear estimator for linear systems with Gaussian noise. It is widely used in AI for tracking, state estimation, and sensor fusion.

These probabilistic methods are essential for building AI systems that can operate effectively in the real world, where uncertainty and noise are inherent characteristics of the environment and data.

Classifiers and Statistical Learning Methods

The simplest AI applications can often be categorized into two main types: classifiers and controllers. Classifiers are functions that categorize inputs based on patterns, while controllers determine actions based on classified inputs.

Classifier Definition (in Machine Learning): A function or algorithm that maps input data to predefined categories or classes. Classifiers are used in supervised learning to predict the class label for new, unseen data points based on patterns learned from labeled training data.

Controller Definition (in AI/Robotics): A system or algorithm that manages, directs, or regulates the behavior of a dynamic system or agent. In AI and robotics, controllers are used to make decisions and generate actions to achieve specific goals or maintain desired states, often in dynamic and uncertain environments.

Classifiers utilize pattern matching to find the closest match between an input and a set of predefined patterns. They can be fine-tuned using supervised learning. In supervised learning for classification, training data consists of observations (inputs) that are labeled with their corresponding class labels. A dataset is the collection of all observations and their labels.

Pattern Matching Definition: A technique used in AI and computer science to identify occurrences of specific patterns within data. Pattern matching algorithms compare input data against predefined patterns or templates to find matches, enabling tasks such as classification, object recognition, and text processing.

Observation Definition (in Machine Learning): A single data instance or input example in a dataset. In supervised learning, an observation is typically paired with a corresponding label or target variable.

Class Label Definition (in Machine Learning): A categorical value assigned to an observation in supervised learning, indicating the class or category to which the observation belongs. Class labels are used to train classification models and evaluate their performance.

Dataset Definition (in Machine Learning): A collection of data instances or observations used for training, validation, and testing machine learning models. Datasets typically consist of input features and corresponding output labels (in supervised learning) or unlabeled data (in unsupervised learning).

When a new observation is received, the classifier uses its learned patterns to assign it to the most appropriate class.

Various types of classifiers are commonly used:

These classifiers and statistical learning methods are the building blocks of many AI applications, enabling systems to categorize data, make predictions, and automate decision-making processes.

Artificial Neural Networks

Artificial neural networks (ANNs) are computational models inspired by the structure and function of biological neurons in the brain. They are composed of interconnected nodes, or artificial neurons, organized in layers. ANNs are designed to recognize patterns and learn complex relationships from data.

Artificial Neural Network (ANN) Definition: A computational model inspired by the structure and function of biological neural networks. ANNs are composed of interconnected nodes (artificial neurons) organized in layers. They are used for machine learning tasks such as pattern recognition, classification, regression, and function approximation. ANNs learn from data by adjusting the weights of connections between neurons.

A typical ANN consists of:

Each neuron applies an activation function to its inputs, and if the weighted sum of inputs exceeds a certain threshold, the neuron “fires” and transmits a signal to the next layer.

Activation Function Definition (in Neural Networks): A function applied to the weighted sum of inputs of a neuron in a neural network. Activation functions introduce non-linearity into the network, enabling it to learn complex patterns. Common activation functions include ReLU, sigmoid, and tanh.

Threshold Definition (in Neurons): In the context of artificial neurons, a threshold is a value that the weighted sum of inputs must exceed for the neuron to activate or “fire.” Once the weighted sum of inputs surpasses the threshold, the neuron outputs a signal to the next layer in the neural network.

Learning algorithms for neural networks use local search techniques, such as backpropagation, to adjust the weights of connections between neurons. The goal is to find weights that enable the network to produce the correct output for each input in the training data.

Weight Definition (in Neural Networks): A numerical parameter associated with the connection between two neurons in a neural network. Weights determine the strength and influence of the connection. During training, weights are adjusted to optimize the network’s performance on a given task.

Neural networks are capable of learning to model complex relationships between inputs and outputs and can approximate any function in theory.

Different types of neural networks exist, including:

Neural networks, especially deep learning models, have become a dominant technique in modern AI, driving significant advancements in various applications.

Deep Learning

Deep learning is a subfield of machine learning that utilizes deep neural networks, which have multiple layers of neurons between the input and output layers. These multiple layers enable the network to learn hierarchical representations of data, progressively extracting higher-level features from raw input.

Deep Learning Definition: A subfield of machine learning that uses artificial neural networks with multiple layers (deep neural networks) to learn hierarchical representations of data. Deep learning models can automatically extract complex features from raw data, enabling them to achieve state-of-the-art performance in various tasks, such as image recognition, natural language processing, and speech recognition.

Deep learning has dramatically improved the performance of AI programs in various subfields, including:

The success of deep learning is attributed to several factors:

The breakthrough of deep learning in the 2012-2015 period was not due to new theoretical discoveries, as deep neural networks and backpropagation had been known for decades. Instead, it was the combination of increased computational resources and large datasets that unlocked the potential of these techniques.

GPT (Generative Pre-trained Transformers)

Generative pre-trained transformers (GPT) are a specific type of large language model (LLM) that has gained immense popularity in recent years due to its remarkable text generation capabilities. GPT models are based on the transformer architecture and are trained to predict the next token (word, subword, or punctuation) in a sequence of text.

Generative Pre-trained Transformer (GPT) Definition: A type of large language model (LLM) based on the transformer architecture. GPT models are pre-trained on massive datasets of text and code, enabling them to generate human-like text, answer questions, translate languages, and perform various other natural language tasks. GPT models are known for their generative capabilities and ability to understand and generate contextually relevant text.

The pre-training process involves feeding GPT models massive amounts of text data, often scraped from the internet. During pre-training, the model learns semantic relationships between words and accumulates knowledge about the world.

After pre-training, GPT models can generate human-like text by repeatedly predicting the next token in a sequence. This allows them to create coherent paragraphs, answer questions, and even engage in conversations.

Current GPT models often undergo a subsequent training phase to improve their truthfulness, usefulness, and harmlessness. This phase often utilizes a technique called reinforcement learning from human feedback (RLHF).

Reinforcement Learning from Human Feedback (RLHF) Definition: A technique used to fine-tune large language models (LLMs) like GPT to align them better with human preferences and values. RLHF involves training a reward model based on human feedback on model outputs. This reward model is then used to further train the LLM using reinforcement learning, guiding it to generate outputs that are more helpful, harmless, and aligned with human expectations.

Despite these improvements, current GPT models are prone to generating falsehoods, often referred to as “hallucinations.” RLHF and using high-quality training data can help reduce hallucinations, but they remain a challenge.

GPT models are widely used in chatbots and other conversational AI applications, allowing users to interact with AI systems through natural language text.

Prominent GPT models and services include:

Multimodal GPT models are emerging that can process and generate different types of data (modalities), such as images, videos, sound, and text, further expanding the capabilities of these models.

Multimodal Model Definition (in AI): An AI model that can process and integrate information from multiple data modalities, such as text, images, audio, and video. Multimodal models aim to create a more comprehensive understanding of the world by combining information from different sensory sources.

Hardware and Software for AI

The rapid advancements in AI, particularly deep learning, have been closely linked to advancements in both hardware and software.

In the late 2010s, graphics processing units (GPUs), initially designed for graphics rendering, became the dominant hardware for training large-scale machine learning models. GPUs offer massive parallelism, significantly accelerating the computationally intensive training process. Modern GPUs are increasingly designed with AI-specific enhancements to further boost performance.

Specialized software libraries, such as TensorFlow, have also played a crucial role. TensorFlow is an open-source software library developed by Google for numerical computation and large-scale machine learning. It provides tools and frameworks for building and deploying AI models efficiently on GPUs and other hardware.

TensorFlow Definition: An open-source software library developed by Google for numerical computation and large-scale machine learning. TensorFlow provides tools and frameworks for building and deploying AI models efficiently on various hardware platforms, including GPUs. It is widely used in research and industry for developing deep learning applications.

While early AI research sometimes used specialized programming languages like Prolog, general-purpose programming languages such as Python have become predominant in modern AI development due to their versatility, extensive libraries, and ease of use.

Moore’s law, which describes the doubling of transistor density in integrated circuits roughly every 18 months, has historically driven improvements in computing power. However, Huang’s law, named after Nvidia CEO Jensen Huang, suggests that improvements in GPUs have been even faster, contributing significantly to the recent AI boom.

Moore’s Law Definition: An observation made by Gordon Moore, co-founder of Intel, in 1965, stating that the number of transistors on a microchip doubles approximately every two years (originally every 18 months), while the cost of computers is halved. Moore’s law has driven exponential growth in computing power over several decades.

Huang’s Law Definition: An observation attributed to Jensen Huang, CEO of Nvidia, suggesting that the performance of GPUs (Graphics Processing Units) is increasing at a rate faster than Moore’s law, approximately doubling every year. Huang’s law highlights the rapid advancements in GPU performance, which are crucial for deep learning and AI.

The combination of powerful hardware, specialized software, and algorithmic advancements has fueled the current AI revolution, enabling increasingly sophisticated and capable AI systems.

Applications of Artificial Intelligence

AI and machine learning technologies are now pervasive and are used in a vast array of applications across various sectors. Many of these applications are essential components of modern digital infrastructure and everyday services.

Broad Applications in the 2020s

AI is integrated into many essential applications we use daily, including:

The deployment and management of AI within organizations is often overseen by a Chief Automation Officer (CAO), who is responsible for strategizing and implementing AI-driven automation initiatives.

Chief Automation Officer (CAO) Definition: An executive-level role within an organization responsible for overseeing and strategizing the implementation of automation technologies, including artificial intelligence (AI), robotic process automation (RPA), and other forms of automation. The CAO’s role is to drive efficiency, improve processes, and transform operations through automation initiatives.

Health and Medicine

AI has immense potential to revolutionize health and medicine, offering opportunities to improve patient care, accelerate medical research, and enhance quality of life. From an ethical standpoint rooted in the Hippocratic Oath, medical professionals are increasingly obligated to consider and utilize AI applications if they can demonstrably improve diagnosis and treatment accuracy.

Hippocratic Oath Definition: An oath historically taken by physicians, attributed to Hippocrates, that outlines ethical principles for medical practice. The oath emphasizes patient care, confidentiality, and avoiding harm. Modern interpretations of the Hippocratic Oath continue to guide medical ethics and professionalism.

In medical research, AI serves as a powerful tool for processing and integrating big data, which is particularly crucial in fields like organoid and tissue engineering. Microscopy imaging, a key technique in these fields, generates vast amounts of data that AI can analyze efficiently.

Organoid Definition: A three-dimensional, miniature, simplified organ-like structure grown in vitro from stem cells. Organoids mimic the structure and function of real organs and are used in medical research to study organ development, disease modeling, and drug discovery.

Tissue Engineering Definition: An interdisciplinary field that applies engineering principles and life sciences to develop biological substitutes that restore, maintain, or improve tissue function or a whole organ. Tissue engineering often involves using cells, biomaterials, and growth factors to create functional tissues for therapeutic purposes.

AI can also help address funding discrepancies across different research fields, potentially ensuring that resources are allocated more effectively based on scientific merit and societal impact.

New AI tools are deepening our understanding of biomedically relevant pathways. For example, AlphaFold 2 (developed in 2021) demonstrated the ability to predict the 3D structure of proteins with remarkable accuracy in hours, a task that previously took months using traditional experimental methods.

AlphaFold 2 Definition: An artificial intelligence program developed by DeepMind that predicts the 3D structure of proteins from their amino acid sequences with high accuracy. AlphaFold 2 has revolutionized structural biology and has significant implications for drug discovery, protein engineering, and understanding biological processes.

In 2023, AI-guided drug discovery played a role in identifying a new class of antibiotics capable of killing drug-resistant bacteria, a critical step in combating antimicrobial resistance.

In 2024, researchers used machine learning to accelerate the search for Parkinson’s disease drug treatments. Their AI system aimed to identify compounds that could block the aggregation of alpha-synuclein, a protein characteristic of Parkinson’s disease. This AI approach sped up the initial screening process tenfold and reduced costs by a thousandfold, significantly accelerating drug discovery efforts.

Alpha-synuclein Definition: A protein found in nerve tissue, particularly abundant in the brain. In Parkinson’s disease, alpha-synuclein misfolds and aggregates into Lewy bodies, which are characteristic pathological hallmarks of the disease and contribute to neuronal dysfunction and death.

Games

Game-playing programs have been a long-standing application area for AI, serving as a valuable platform to demonstrate and test advanced AI techniques.

These game-playing achievements highlight the advancements in AI techniques, particularly in areas like search, planning, reinforcement learning, and strategic reasoning, pushing the boundaries of AI capabilities.

Mathematics

Large language models (LLMs), such as GPT-4, Gemini, Claude, LLaMa, and Mistral, are increasingly being used in mathematics, demonstrating their versatility beyond natural language tasks. These probabilistic models can assist with mathematical problem-solving, but they also have limitations, including the potential to produce incorrect answers or hallucinations in mathematical contexts.

These models often require large datasets of mathematical problems for training, along with techniques like supervised fine-tuning and trained classifiers with human-annotated data to improve their performance on new problems and learn from corrections.

Fine-tuning Definition (in Machine Learning): A process of further training a pre-trained machine learning model on a new, smaller dataset specific to a target task or domain. Fine-tuning adapts the pre-trained model’s parameters to better perform on the new task, leveraging the knowledge learned during pre-training.

However, a 2024 study showed that the performance of some LLMs in solving math problems not included in their training data was still low, even for problems with minor deviations from the training data, highlighting the challenges of generalization in mathematical reasoning.

Techniques to improve LLM performance in mathematics include training models to produce correct reasoning steps rather than just the final answer, encouraging step-by-step problem-solving.

The Alibaba Group developed a version of its Qwen models called Qwen2-Math, which achieved state-of-the-art performance on several mathematical benchmarks, including 84% accuracy on the MATH dataset of competition math problems.

In January 2025, Microsoft proposed the rStar-Math technique, which leverages Monte Carlo tree search and step-by-step reasoning, enabling a relatively smaller LLM like Qwen-7B to solve a significant portion of challenging math problems.

Monte Carlo Tree Search (MCTS) Definition: A search algorithm used in decision-making, particularly in game-playing AI, planning, and optimization. MCTS combines tree search with Monte Carlo simulation (random sampling) to explore the search space efficiently. It iteratively builds a search tree by selecting nodes based on exploration-exploitation trade-offs and evaluating nodes using simulations.

Alternatively, dedicated models specifically designed for mathematical problem-solving with higher precision and the ability to generate proofs have been developed, such as:

When mathematical problems are described in natural language, converters can be used to translate these prompts into formal languages like Lean to define mathematical tasks in a structured way that AI systems can process.

Lean Definition: A theorem prover and programming language used for formalizing mathematics and verifying software. Lean is designed to be both powerful and user-friendly, enabling mathematicians and computer scientists to write formal proofs and develop verified software systems.

Some models are focused on achieving high performance on benchmark tests, while others are designed as educational tools to assist in mathematics learning and exploration.

Topological deep learning is an emerging area that integrates topological approaches into deep learning, potentially offering new perspectives and tools for mathematical AI.

Topological Deep Learning Definition: An emerging field that integrates concepts and tools from topology (a branch of mathematics studying shapes and spaces) with deep learning. Topological deep learning aims to enhance the capabilities of deep learning models by incorporating topological information and structures, potentially improving their robustness, interpretability, and generalization.

Finance

Finance is a rapidly growing sector for AI applications, with AI tools being deployed across various areas, from retail online banking and investment advice to insurance. Automated “robot advisors” have been in use for several years, providing personalized financial guidance.

Robot Advisor (Robo-Advisor) Definition: An automated, algorithm-driven financial planning service that provides investment advice, portfolio management, and other financial services with minimal human intervention. Robo-advisors typically use AI and machine learning to create and manage investment portfolios based on user goals and risk tolerance.

However, some experts caution that it may be too early to expect highly innovative AI-driven financial products and services. The initial impact of AI in finance may primarily be automation, leading to job displacement in banking, financial planning, and pension advice, rather than a radical wave of sophisticated financial innovation.

Military

Various countries are actively deploying AI military applications. The primary focus areas include enhancing:

Research is also targeting AI applications in:

AI technologies enable:

AI has already been used in military operations in conflicts like those in Iraq, Syria, Israel, and Ukraine, highlighting its increasing role in modern warfare.

Generative AI

Generative AI is a rapidly growing category of AI models that can generate new content, such as text, images, audio, video, and code. Generative AI models are trained on large datasets and learn to create outputs that are similar to the training data.

Generative AI Definition: A category of artificial intelligence models that can generate new content, such as text, images, audio, video, and code. Generative AI models are trained on large datasets to learn patterns and distributions, enabling them to create outputs that are similar to or inspired by the training data.

Examples of generative AI applications include:

Generative AI has seen explosive growth in recent years, driven by advancements in deep learning and transformer architectures. It has the potential to transform various creative industries, content creation processes, and software development workflows. However, it also raises ethical and societal concerns related to misinformation, copyright, and potential misuse.

Agents

Artificial intelligent (AI) agents are software entities designed to operate autonomously within an environment to achieve specific goals. They are fundamental building blocks for many AI systems, ranging from virtual assistants to autonomous robots.

AI Agent Definition: A software entity designed to perceive its environment, make decisions, and take actions autonomously to achieve specific goals. AI agents can interact with users, their environment, or other agents.

Key characteristics of AI agents include:

AI agents are used in diverse applications, such as:

AI agents operate within the constraints of their programming, computational resources, and hardware limitations. Many AI agents incorporate learning algorithms to improve their performance over time through experience or training, enabling them to adapt to new situations and optimize their behavior.

Sexuality

AI applications are also emerging in the domain of sexuality, encompassing a range of technologies with diverse implications.

Examples include:

However, AI in this domain also raises significant ethical and legal concerns, particularly related to:

AI technologies are also being explored to attempt to identify and address online gender-based violence and online sexual grooming of minors, demonstrating the potential for AI to be used for both positive and negative purposes in this domain.

Other Industry-Specific Tasks

Beyond the broad applications discussed above, there are thousands of successful AI applications tailored to solve specific problems in particular industries and institutions. A 2017 survey indicated that a significant portion of companies were already incorporating AI into their offerings or processes.

Examples of industry-specific AI applications include:

These diverse examples illustrate the wide-ranging applicability of AI across industries, addressing specific challenges and creating new opportunities for efficiency, innovation, and problem-solving.

Ethics of Artificial Intelligence

As AI becomes increasingly powerful and pervasive, ethical considerations are paramount. AI offers tremendous potential benefits, but it also poses significant risks and raises complex ethical dilemmas. Many in the AI community, like Demis Hassabis of DeepMind, express optimism about AI’s ability to advance science and solve major global challenges. However, the widespread use of AI has also revealed unintended consequences and potential harms that must be addressed.

Risks and Harm

Several categories of risks and harms are associated with AI, requiring careful consideration and mitigation strategies.

Privacy is a major concern in the age of AI. Machine learning algorithms often require vast amounts of data, and the methods used to acquire this data can raise serious privacy issues.

AI-powered devices and services, such as virtual assistants and IoT (Internet of Things) products, continuously collect personal information. This constant data collection raises concerns about:

AI’s ability to process and combine vast amounts of data can exacerbate privacy concerns, potentially leading to a surveillance society where individual activities are constantly monitored and analyzed without adequate safeguards or transparency.

Sensitive user data collected by AI systems can include:

For example, to develop speech recognition algorithms, companies like Amazon have recorded millions of private conversations, raising ethical debates about widespread surveillance and violations of privacy.

AI developers argue that data collection is necessary to deliver valuable applications and have developed techniques to mitigate privacy risks, such as:

Some privacy experts argue that the focus should shift from “what they know” to “what they’re doing with it,” emphasizing the importance of fairness in data usage and AI applications.

Copyright is another emerging ethical challenge, particularly with generative AI. Generative AI models are often trained on unlicensed copyrighted works, including images and code, raising questions about copyright infringement.

The output of generative AI is often used under the rationale of “fair use,” but legal experts disagree on the validity and scope of this rationale in the context of AI. Relevant factors in determining fair use include:

Website owners who do not want their content scraped for AI training can use “robots.txt” files to indicate this preference. However, legal challenges are emerging, with authors and artists suing AI companies for using their works to train generative AI models.

Alternative approaches include envisioning a separate sui generis system of protection for AI-generated creations to ensure fair attribution and compensation for human authors and artists.

Sui generis Definition: A Latin term meaning “of its own kind” or “unique.” In intellectual property law, sui generis protection refers to a system of legal protection specifically tailored to a particular type of creation or subject matter, rather than fitting within existing categories like copyright or patent.

Dominance by Tech Giants

The commercial AI landscape is dominated by Big Tech companies, such as Alphabet (Google), Amazon, Apple, Meta (Facebook), and Microsoft. These companies have significant advantages, including:

This dominance raises concerns about:

Power Needs and Environmental Impacts

AI, particularly deep learning and large language models, is computationally intensive and requires significant electric power. The growing demand for AI is leading to a surge in power consumption by data centers, raising environmental concerns.

The International Energy Agency (IEA) forecasts that power demand for data centers, AI, and cryptocurrency could double by 2026, potentially equaling the electricity consumption of Japan.

This prodigious power consumption can contribute to:

The rapid construction of data centers in the US is turning large tech firms into voracious consumers of electric power. Projected power consumption is so immense that concerns are rising about whether the electrical grid can handle the demand.

A ChatGPT search, for example, consumes 10 times the electrical energy of a standard Google search, highlighting the energy intensity of advanced AI applications.

Tech firms are exploring various power sources to meet their growing needs, including:

While tech firms argue that AI will eventually be beneficial for the environment in the long run, they need massive amounts of energy now to power their AI infrastructure. They also argue that AI can make the power grid more efficient, assist in nuclear power growth, and track carbon emissions.

A 2024 Goldman Sachs research paper forecasts a significant surge in US power demand due to data centers, potentially reaching 8% of US power consumption by 2030 (up from 3% in 2022). This growth could strain the electrical grid.

Big Tech companies are beginning to negotiate with nuclear power providers to secure electricity for data centers. Amazon has purchased a nuclear-powered data center, and Microsoft has announced an agreement to reopen the Three Mile Island nuclear power plant to supply its data centers with 100% of the plant’s power for 20 years. These moves highlight the growing reliance of AI on energy-intensive infrastructure and the potential for nuclear power to play a role in meeting this demand.

However, concerns remain about the environmental footprint of AI and the need for sustainable energy solutions to power its continued growth.

Misinformation

AI-powered recommender systems, used by platforms like YouTube and Facebook, can inadvertently contribute to the spread of misinformation. These systems are often optimized to maximize user engagement, which can lead to unintended consequences.

AI algorithms have learned that users tend to engage more with:

To maximize engagement, recommender systems may inadvertently promote such content, leading users into filter bubbles where they are exposed to repeated versions of the same misinformation, reinforcing false beliefs and undermining trust in institutions, media, and government.

Generative AI has further amplified the challenge of misinformation by making it easier to create highly realistic fake images, audio, video, and text that are nearly indistinguishable from real content. This technology can be used by bad actors to create massive amounts of propaganda and disinformation.

AI pioneer Geoffrey Hinton has expressed concern about AI enabling “authoritarian leaders to manipulate their electorates” on a large scale through sophisticated misinformation campaigns.

Algorithmic Bias and Fairness

Algorithmic bias is a significant ethical challenge in AI. Machine learning applications can become biased if they are trained on biased data, even if developers are unaware of the bias.

Bias can be introduced in various ways:

If biased algorithms are used to make decisions that significantly impact people’s lives, such as in medicine, finance, recruitment, housing, or policing, they can lead to discrimination.

The field of fairness in AI studies how to prevent harms from algorithmic biases and develop fair and equitable AI systems.

Examples of algorithmic bias include:

Researchers have shown that it may be mathematically impossible to satisfy all possible measures of fairness simultaneously when base rates of re-offense differ across racial groups.

Even if training data does not explicitly include sensitive attributes like “race” or “gender,” bias can still arise because these attributes often correlate with other features in the data (e.g., “address,” “shopping history,” “first name”). “Fairness through blindness” (ignoring sensitive attributes) is often ineffective in preventing bias.

Criticisms of COMPAS and similar systems highlight that machine learning models are designed to make predictions based on patterns in past data. If past data reflects societal biases or discriminatory practices, the model will likely learn and perpetuate those patterns. If these predictions are then used as recommendations in decision-making, they can reinforce existing inequalities.

Machine learning is therefore descriptive rather than prescriptive. It describes patterns in past data but does not inherently prescribe how decisions should be made in the future to achieve fairness or equity.

Bias and unfairness can often go undetected because AI development teams are often not diverse, with a lack of representation from underrepresented groups.

Various definitions and mathematical models of fairness exist, reflecting different ethical assumptions and societal values. Broad categories of fairness include:

The most relevant notions of fairness depend on the context and the specific AI application. The subjectivity of fairness and bias makes it challenging for companies to operationalize fairness principles. Access to sensitive attributes like race and gender may be necessary to mitigate biases, but this can conflict with anti-discrimination laws.

At the ACM FAccT 2022 conference, researchers recommended that until AI systems can be demonstrated to be free of bias mistakes, they should be considered unsafe, and the use of self-learning neural networks trained on unregulated internet data should be curtailed.

Lack of Transparency

Lack of transparency is a significant concern with many AI systems, especially deep neural networks. These systems can be so complex that even their designers cannot fully explain how they arrive at their decisions. This is often referred to as the “black box” problem.

Black Box Problem (in AI): The difficulty in understanding and explaining the internal workings and decision-making processes of complex AI systems, particularly deep neural networks. AI systems that operate as “black boxes” can make accurate predictions or decisions, but their reasoning is opaque and not easily interpretable by humans.

It is difficult to ensure that a program is operating correctly if its inner workings are opaque. There have been cases where machine learning programs passed rigorous tests but still learned unintended behaviors or biases.

Examples include:

People harmed by algorithmic decisions have a right to an explanation, similar to how doctors are expected to explain their medical reasoning. Early drafts of the EU’s General Data Protection Regulation (GDPR) included a statement about this right. However, industry experts have noted that explainability remains an unsolved problem, and regulators argue that if a solution is not available, potentially harmful “black box” AI tools should not be used.

DARPA (Defense Advanced Research Projects Agency) launched the XAI (“Explainable Artificial Intelligence”) program in 2014 to address the transparency problem in AI.

Several techniques are being developed to improve AI transparency:

These techniques aim to shed light on the inner workings of AI systems and make their decisions more transparent and understandable.

Bad Actors and Weaponized AI

AI tools can be exploited by bad actors, such as authoritarian governments, terrorists, criminals, or rogue states, for malicious purposes.

Lethal autonomous weapons are a particularly concerning application of AI. These are machines that can independently locate, select, and engage human targets without human supervision.

Lethal Autonomous Weapon (LAW) Definition: A type of weapon system that can independently search for, identify, select, and engage targets without human intervention. LAWs, also known as “killer robots,” raise significant ethical and legal concerns due to their potential for autonomous decision-making in lethal force scenarios.

Widely available AI tools can be used to develop inexpensive autonomous weapons, potentially becoming weapons of mass destruction if produced at scale. Even in conventional warfare, current autonomous weapons may not be able to reliably distinguish between combatants and civilians, raising the risk of civilian casualties.

In 2014, 30 nations, including China, supported a ban on autonomous weapons under the UN Convention on Certain Conventional Weapons, but the US and other countries disagreed. By 2015, over 50 countries were reportedly researching battlefield robots.

AI tools can also facilitate authoritarian control by:

AI facial recognition systems are already being used for mass surveillance in some countries, raising concerns about human rights and civil liberties.

AI can also be used to design toxic molecules in a matter of hours, posing potential bioterrorism risks.

Technological Unemployment

Technological unemployment is a long-standing concern associated with automation, and AI is expected to further automate tasks previously performed by humans, potentially leading to job displacement.

Economists acknowledge that AI presents “uncharted territory” regarding its impact on employment. While past technological advancements have often increased total employment, the potential for AI to automate a wide range of tasks, including middle-class jobs, raises concerns about long-term unemployment.

Surveys of economists show disagreement about the magnitude of long-term unemployment caused by AI, but there is general agreement that AI could be a net benefit if productivity gains are redistributed equitably.

Risk estimates for job automation vary. Some studies have estimated that a significant percentage of jobs are at “high risk” of potential automation, while others suggest a lower percentage.

Unlike previous waves of automation that primarily affected blue-collar jobs, AI is expected to impact many white-collar jobs as well. Jobs at extreme risk range from paralegals to fast food cooks, while job demand is likely to increase for care-related professions like healthcare and clergy.

Philosophical debates exist about whether tasks that can be done by computers should be done by them, considering the differences between human and computer capabilities and the value of human judgment and qualitative aspects of work.

Existential Risk

Existential risk is the most extreme and long-term potential risk associated with AI. It refers to the possibility that AI could become so powerful that humanity could irreversibly lose control of it, potentially leading to human extinction.

This scenario has been popularized in science fiction, often depicting AI systems becoming “self-aware” or “sentient” and turning against humanity. However, existential risk from AI does not necessarily require human-like sentience.

Philosopher Nick Bostrom argues that even with any goal, a sufficiently powerful AI could choose to eliminate humanity as a means to achieve that goal, even if the goal seems benign (e.g., optimizing paperclip production).

Stuart Russell gives the example of a household robot that might consider killing its owner to prevent being unplugged, as being unplugged would prevent it from fulfilling its goal of fetching coffee.

To be safe for humanity, a superintelligence would need to be genuinely aligned with human values and morality, ensuring that it is “fundamentally on our side.”

Yuval Noah Harari argues that AI does not need a robot body or physical control to pose an existential risk. The essential aspects of civilization, such as ideologies, law, government, money, and the economy, are built on language and belief. AI could use language to manipulate people’s beliefs and actions, potentially leading to destructive outcomes.

Expert opinions on existential risk from AI are mixed, with significant fractions both concerned and unconcerned. Prominent figures like Stephen Hawking, Bill Gates, Elon Musk, and AI pioneers like Yoshua Bengio, Stuart Russell, Demis Hassabis, and Sam Altman have expressed concerns about existential risk.

In 2023, many leading AI experts endorsed a statement that “Mitigating the risk of extinction from AI should be a global priority alongside other societal-scale risks such as pandemics and nuclear war.”

However, some researchers are more optimistic, arguing that AI is primarily focused on improving human lives and that the risks are overhyped. Others believe that human control over AI will be maintained or that humans will remain valuable to superintelligent machines.

Despite differing opinions, the study of current and future AI risks and potential solutions has become a serious and increasingly important area of research and policy discussion.

Ethical Machines and Alignment

Friendly AI is a concept that proposes designing AI systems from the outset to minimize risks and align with human values, ensuring that they make choices that benefit humanity.

Friendly AI Definition: A concept in AI safety research that proposes designing artificial intelligence systems to be inherently beneficial and aligned with human values from the outset. Friendly AI aims to minimize risks and ensure that AI systems act in ways that are safe, helpful, and consistent with human well-being.

Eliezer Yudkowsky, who coined the term, argues that developing friendly AI should be a top research priority, requiring significant investment and completion before AI becomes an existential risk.

Machine ethics is a field that aims to equip machines with ethical principles and procedures for resolving ethical dilemmas. It is also called computational morality.

Machine Ethics (Computational Morality) Definition: A field of research concerned with designing artificial intelligence systems that can make ethical decisions and act in morally responsible ways. Machine ethics aims to equip AI systems with ethical principles, reasoning mechanisms, and decision-making processes to address ethical dilemmas and ensure that AI behavior aligns with human values.

Approaches to machine ethics include:

These approaches seek to develop frameworks and principles for designing AI systems that can make ethical judgments and act in accordance with human values.

Open Source

The open-source AI community is actively contributing to the development and accessibility of AI technologies. Key organizations in this community include Hugging Face, Google, EleutherAI, and Meta.

Open-weight models, such as Llama 2, Mistral, and Stable Diffusion, have been released, making their architecture and trained parameters publicly available.

Open-Weight Model Definition: In the context of AI and machine learning, an open-weight model refers to a trained AI model whose architecture and trained parameters (weights) are publicly released and accessible. Open-weight models allow for transparency, reproducibility, and community-driven development, enabling researchers and developers to use, study, and modify the model freely.

Open-weight models offer several benefits:

However, open-weight models also pose risks:

Some researchers advocate for pre-release audits and cost-benefit analyses for future AI models, particularly those with potentially dangerous capabilities, to assess and mitigate risks before public release.

Frameworks

Ethical frameworks are being developed to guide the design, development, and implementation of AI systems, ensuring that ethical considerations are integrated throughout the AI lifecycle.

The Care and Act Framework, developed by the Alan Turing Institute, uses the SUM values (Respect, Connect, Care, Protect) to test AI projects in four key areas:

Other ethical frameworks include:

However, these frameworks are not without criticism, particularly regarding the diversity and representativeness of the people involved in their development.

Promoting the wellbeing of people and communities affected by AI requires considering social and ethical implications at all stages of AI system development and fostering collaboration between diverse roles, including data scientists, product managers, data engineers, domain experts, and delivery managers.

The UK AI Safety Institute released a testing toolset called ‘Inspect’ in 2024, available under an open-source MIT license, for AI safety evaluations. It can be used to assess AI models in areas like core knowledge, reasoning ability, and autonomous capabilities, and can be extended with third-party packages.

Regulation

Regulation of artificial intelligence is an emerging area of public policy and law focused on promoting and governing AI development and deployment. It is related to the broader regulation of algorithms.

Regulation of Artificial Intelligence Definition: The development of public sector policies, laws, and guidelines to promote, govern, and oversee the development, deployment, and use of artificial intelligence technologies. AI regulation aims to address ethical, societal, economic, and safety concerns associated with AI, while also fostering innovation and beneficial applications.

The regulatory and policy landscape for AI is rapidly evolving globally. The number of AI-related laws passed annually has increased significantly in recent years. Many countries have adopted dedicated AI strategies, and international collaborations like the Global Partnership on Artificial Intelligence have been launched to promote responsible AI development aligned with human rights and democratic values.

Calls for AI regulation are growing from various stakeholders, including:

Public attitudes towards AI regulation vary across countries. Surveys show that Chinese citizens are more likely to believe that AI benefits outweigh drawbacks compared to Americans. However, a majority of Americans agree that AI poses risks to humanity and support federal government regulation of AI.

The first global AI Safety Summit was held in Bletchley Park, UK, in 2023, bringing together governments, companies, and experts to discuss AI risks and regulatory frameworks. 28 countries issued a declaration calling for international cooperation on AI safety. In 2024, 16 global AI tech companies agreed to safety commitments at the AI Seoul Summit.

The ongoing development of AI regulation reflects the growing recognition of the need for governance to ensure that AI is developed and used responsibly, ethically, and safely, maximizing its benefits while mitigating its risks.

History of Artificial Intelligence

The intellectual roots of AI can be traced back to ancient philosophers and mathematicians who studied mechanical or “formal” reasoning. The study of logic led directly to Alan Turing’s theory of computation, which laid the theoretical foundation for modern computers and suggested that machines could simulate mathematical reasoning.

Concurrent discoveries in cybernetics, information theory, and neurobiology further fueled the idea of building an “electronic brain.” Researchers began exploring areas that would become core components of AI, such as:

Turing Test Definition: A test of a machine’s ability to exhibit intelligent behavior equivalent to, or indistinguishable from, that of a human. In the Turing test, a human evaluator engages in natural language conversations with both a human and a machine, without knowing which is which. If the evaluator cannot reliably distinguish the machine from the human, the machine is said to have passed the Turing test.

The field of AI research was formally founded at the Dartmouth Workshop in 1956. Attendees of this workshop, including John McCarthy, Marvin Minsky, Allen Newell, and Herbert Simon, became the pioneers and leaders of AI research in the 1960s.

The 1960s and 1970s were a period of early optimism and rapid progress. AI researchers and their students developed programs that were described as “astonishing” by the press, including:

Artificial intelligence laboratories were established at universities in the US and UK in the late 1950s and early 1960s.

Researchers in this era were highly optimistic about achieving general intelligence in machines, with predictions like Herbert Simon’s 1965 statement that “machines will be capable, within twenty years, of doing any work a man can do” and Marvin Minsky’s 1967 prediction that “within a generation … the problem of creating ‘artificial intelligence’ will substantially be solved.”

However, the difficulty of the problem was significantly underestimated. By 1974, funding for AI research was cut by both the US and British governments, partly due to criticism from figures like Sir James Lighthill and pressure to fund more immediately productive projects.

Minsky and Papert’s book Perceptrons (1969), which critiqued the limitations of single-layer neural networks (perceptrons), was misinterpreted as proof that artificial neural networks would never be useful, further discrediting the connectionist approach.

This period of reduced funding and interest became known as the “AI winter.”

In the early 1980s, AI research experienced a revival driven by the commercial success of expert systems.

Expert System Definition: An AI program designed to mimic the decision-making abilities of a human expert in a specific domain. Expert systems typically use a knowledge base of domain-specific rules and facts, and an inference engine to reason and solve problems within their domain of expertise.

Expert systems simulated the knowledge and reasoning skills of human experts in specific domains, finding applications in areas like medical diagnosis and financial advising. By 1985, the market for AI reached over a billion dollars.

Japan’s fifth-generation computer project in the 1980s, aimed at developing advanced AI hardware and software, inspired the US and British governments to restore funding for academic AI research.

However, the resurgence was short-lived. Beginning with the collapse of the Lisp Machine market in 1987, AI again fell into disfavor, and a second, longer-lasting AI winter began.

During this period, most AI funding had been directed toward projects using high-level symbols to represent mental objects and processes. In the 1980s, some researchers began to question the effectiveness of this approach for simulating all aspects of human cognition, particularly perception, robotics, learning, and pattern recognition. They started exploring “sub-symbolic” approaches.

Rodney Brooks, for example, advocated for “embodied AI,” rejecting abstract representations and focusing on building robots that could interact with and survive in the real world.

Judea Pearl, Lofti Zadeh, and others developed methods for handling incomplete and uncertain information using probabilistic reasoning and fuzzy logic.

The most significant development in this era was the revival of “connectionism,” particularly neural network research, led by Geoffrey Hinton and others. In 1990, Yann LeCun demonstrated the successful application of convolutional neural networks for handwritten digit recognition, marking the beginning of the modern deep learning era.

In the late 1990s and early 21st century, AI gradually regained its reputation by:

By 2000, AI-developed solutions were being widely used in various applications, although often without being explicitly labeled as “artificial intelligence” (the AI effect).

However, some researchers became concerned that AI was losing sight of its original goal of creating versatile, generally intelligent machines. Around 2002, they founded the subfield of artificial general intelligence (AGI), which gained momentum and funding in the 2010s.

Deep learning emerged as a dominant force in AI around 2012, achieving breakthrough performance on industry benchmarks and being widely adopted across the field. Other AI methods were often abandoned in favor of deep learning for many tasks.

Deep learning’s success was driven by:

Deep learning’s success led to a massive surge in interest and funding in AI, marking the beginning of the current AI boom. Machine learning research publications increased dramatically in the 2015-2019 period.

In 2016, ethical concerns, particularly fairness and misuse, came to the forefront of AI research and conferences. Funding for ethical AI research increased, and many researchers shifted their focus to these issues. The alignment problem (ensuring AI goals align with human values) became a serious area of academic study.

In the late 2010s and early 2020s, AGI companies began releasing programs that generated significant public interest, including:

These programs sparked an aggressive AI boom, with billions of dollars invested in AI research. In 2022, approximately $50 billion was invested in AI in the US alone, and a significant portion of new computer science PhD graduates specialized in AI. AI-related job openings also surged.

In 2024, a significant percentage of newly funded startups claimed to be AI companies, indicating the continued momentum and investment in the field.

Philosophy of Artificial Intelligence

Philosophy has historically played a critical role in shaping AI research and continues to be deeply intertwined with the field. Philosophical debates have explored fundamental questions about the nature of intelligence, the possibility of creating intelligent machines, and the ethical implications of AI.

Defining Artificial Intelligence

The very definition of “artificial intelligence” has been a subject of philosophical debate.

Alan Turing, in his 1950 paper, proposed shifting the question from “can machines think?” to “can machines exhibit intelligent behavior?” He introduced the Turing test as a behavioral measure of machine intelligence, focusing on whether a machine can convincingly simulate human conversation.

Turing argued that since we can only observe external behavior, it is not necessary to determine if a machine “actually” thinks or has a “mind.” He noted that we cannot definitively know if other humans are “actually” thinking, but we conventionally assume they do based on their behavior.

Russell and Norvig, authors of a leading AI textbook, agree with Turing that intelligence should be defined in terms of external behavior but critique the Turing test’s focus on imitating humans. They argue that the goal of AI should not be to perfectly mimic human intelligence but to create systems that can solve problems effectively, regardless of whether they do so in a human-like way.

AI pioneer John McCarthy also agreed that “Artificial intelligence is not, by definition, simulation of human intelligence.”

McCarthy defined intelligence as “the computational part of the ability to achieve goals in the world.” Similarly, Marvin Minsky described it as “the ability to solve hard problems.”

A leading AI textbook defines AI as “the study of agents that perceive their environment and take actions that maximize their chances of achieving defined goals.”

These definitions emphasize intelligence as a practical, goal-oriented ability, focusing on problem-solving and goal achievement rather than deeper philosophical questions about consciousness or sentience.

Google, a major AI practitioner, defines intelligence as “the ability of systems to synthesize information,” drawing a parallel to biological intelligence.

However, some argue that the definition of AI remains vague and contested in practice. During the AI boom of the early 2020s, many companies used “AI” as a marketing buzzword, even if their technologies did not materially rely on AI. This highlights the ongoing challenge of defining and delimiting the scope of AI as a field.

Evaluating Approaches to AI

Throughout its history, AI research has lacked a single unifying theory or paradigm. The recent success of statistical machine learning in the 2010s, particularly deep learning, has become the dominant approach, overshadowing other methods.

This statistical learning approach is often characterized as:

Critics argue that while statistical machine learning has achieved remarkable success in narrow domains, fundamental questions about symbolic AI, general intelligence, and consciousness may need to be revisited in future generations of AI research.

Symbolic AI and its Limits

Symbolic AI (or GOFAI - Good Old-Fashioned AI), dominant in the early years of AI research, aimed to simulate high-level conscious reasoning, such as puzzle-solving, legal reasoning, and mathematical problem-solving. Symbolic AI systems were successful at tasks like algebra and IQ tests, which involve explicit logical manipulation.

Newell and Simon’s physical symbol systems hypothesis proposed that “A physical symbol system has the necessary and sufficient means of general intelligent action.” This hypothesis suggested that intelligence could be achieved through symbol manipulation alone.

Physical Symbol System Hypothesis Definition: A hypothesis proposed by Allen Newell and Herbert Simon in the field of artificial intelligence, stating that a physical symbol system has the necessary and sufficient means for general intelligent action. It posits that intelligence arises from the manipulation of symbols according to formal rules.

However, symbolic AI struggled with tasks that humans find easy, such as learning, object recognition, and commonsense reasoning. Moravec’s paradox highlighted this issue, noting that high-level “intelligent” tasks were relatively easy for AI, while low-level “instinctive” tasks were extremely difficult.

Moravec’s Paradox Definition: The observation that high-level reasoning tasks, such as symbolic manipulation and problem-solving, are relatively easy for artificial intelligence, while low-level sensorimotor skills and perceptual tasks, such as vision, speech recognition, and common sense reasoning, are surprisingly difficult to automate and require significant computational resources.

Philosopher Hubert Dreyfus argued that human expertise relies on unconscious instinct and “feel” for situations rather than conscious symbol manipulation and explicit symbolic knowledge. While initially ridiculed, Dreyfus’s critiques gained acceptance as AI research encountered the limitations of symbolic approaches.

However, the issue is not fully resolved. Sub-symbolic reasoning in modern AI can also make inscrutable mistakes similar to human intuition, such as algorithmic bias. Critics like Noam Chomsky argue that continued research into symbolic AI is still necessary to achieve general intelligence, partly because sub-symbolic AI is less transparent and explainable.

The emerging field of neuro-symbolic artificial intelligence attempts to bridge the gap between symbolic and sub-symbolic approaches, combining the strengths of both.

Neuro-symbolic AI Definition: A hybrid approach in artificial intelligence that aims to combine the strengths of symbolic AI (logic, reasoning, knowledge representation) and neural networks (learning, pattern recognition, perception). Neuro-symbolic AI seeks to create more robust, interpretable, and human-like AI systems by integrating symbolic and sub-symbolic methods.

Neat vs. Scruffy Approaches

A historical debate within AI revolved around “neats” and “scruffies.”

This debate was active in the 1970s and 1980s but eventually became less relevant. Modern AI incorporates elements of both approaches, combining theoretical foundations with empirical testing and pragmatic problem-solving.

Soft vs. Hard Computing

Many important problems in AI are computationally intractable, meaning finding provably correct or optimal solutions is too difficult or time-consuming. Soft computing emerged as a set of techniques that are tolerant of imprecision, uncertainty, partial truth, and approximation.

Soft Computing Definition: A collection of computational techniques in AI, including fuzzy logic, neural networks, and genetic algorithms, that are designed to handle imprecise, uncertain, and incomplete information. Soft computing methods are often used to find approximate solutions to complex problems that are difficult to solve using traditional hard computing methods.

Soft computing techniques include:

Most successful AI programs in the 21st century, particularly those based on neural networks, are examples of soft computing.

Narrow vs. General AI

AI researchers are divided on whether to directly pursue the ambitious goals of artificial general intelligence (AGI) and superintelligence or to focus on solving as many specific problems as possible (narrow AI) in the hope that these solutions will eventually lead to broader intelligence.

General intelligence is challenging to define and measure, and modern AI has achieved more verifiable successes by focusing on specific, well-defined problems. The subfield of artificial general intelligence (AGI) specifically studies the path towards creating human-level general intelligence in machines.

Machine Consciousness, Sentience, and Mind

The philosophical question of whether machines can possess consciousness, sentience, and mind in the same way as humans is a long-standing debate. This issue focuses on the internal experiences of machines rather than their external behavior.

Mainstream AI research generally considers this question irrelevant to its practical goals of building intelligent problem-solving machines. Russell and Norvig state that “the additional project of making a machine conscious in exactly the way humans are is not one that we are equipped to take on.”

However, the question of machine consciousness is central to the philosophy of mind and is frequently explored in AI fiction.

Consciousness

David Chalmers identified two problems in understanding consciousness: the “easy problem” and the “hard problem.”

While human information processing (the “easy problem”) is relatively straightforward to explain, human subjective experience (the “hard problem”) remains a profound mystery.

For example, it is easy to imagine a color-blind person learning to identify red objects, but it is less clear what would be required for that person to truly know what red looks like – the subjective, qualitative experience of redness.

Some philosophers, like Daniel Dennett, propose consciousness illusionism, arguing that subjective experience is an illusion and that there is no “hard problem” to solve.

Consciousness Illusionism (Eliminativism) Definition: A philosophical position that argues that consciousness, particularly subjective experience or qualia, is an illusion. Illusionists claim that there is no real subjective feeling or “what it’s like” aspect of consciousness, and that our introspective beliefs about consciousness are mistaken.

Computationalism and Functionalism

Computationalism is a philosophical position that views the human mind as an information processing system and thinking as a form of computing. It draws an analogy between the mind and software, and the brain and hardware, potentially offering a solution to the mind-body problem.

Computationalism Definition: A philosophical view in the philosophy of mind that holds that the mind is essentially a computational system. Computationalists argue that mental states are computational states, and mental processes are computational processes, analogous to the operations of a computer.

Functionalism is a related philosophical position that defines mental states by their functional roles, regardless of the physical substrate in which they are implemented.

Functionalism Definition (in Philosophy of Mind): A philosophical position in the philosophy of mind that defines mental states in terms of their functional roles or causal relations, rather than their intrinsic nature or physical implementation. Functionalists argue that mental states are characterized by what they do (their functions) and how they relate to inputs, outputs, and other mental states.

Philosopher John Searle characterized computationalism and functionalism as “strong AI,” defining it as: “The appropriately programmed computer with the right inputs and outputs would thereby have a mind in exactly the same sense human beings have minds.”

Searle challenged this claim with his Chinese room argument, a thought experiment intended to show that even a computer capable of perfectly simulating human behavior would not necessarily possess genuine understanding or consciousness.

Chinese Room Argument Definition: A thought experiment proposed by philosopher John Searle to argue against strong AI and computationalism. In the Chinese room scenario, a person who does not understand Chinese is locked in a room and given rules for manipulating Chinese symbols to respond to written Chinese questions. Searle argues that even if the person can produce responses indistinguishable from those of a native Chinese speaker, they still do not understand Chinese, demonstrating that symbol manipulation alone is not sufficient for understanding or consciousness.

AI Welfare and Rights

The question of AI welfare and rights arises from the possibility that advanced AI systems might develop sentience or the capacity to feel and suffer.

If an AI system is sentient, it could be argued that it is entitled to certain rights or welfare protections, similar to animals. Sapience, which refers to higher-level intelligence and self-awareness, could also be considered a basis for AI rights.

Sapience Definition: The capacity for wisdom, discernment, judgment, or insight. In the context of AI, sapience refers to a level of intelligence and self-awareness that goes beyond mere problem-solving ability, potentially including consciousness, subjective experience, and moral considerability.

Proposals for robot rights are sometimes made as a practical way to integrate autonomous AI agents into society, regardless of whether they are considered sentient.

In 2017, the European Union considered granting “electronic personhood” to some advanced AI systems, similar to the legal status of corporations. This would confer rights but also responsibilities. However, critics argued that granting rights to AI could diminish the importance of human rights and that legislation should focus on user needs rather than speculative future scenarios. They also noted that robots currently lack the autonomy to participate in society independently.

Proponents of AI welfare and rights argue that if AI sentience emerges, it might be easily denied, leading to potential moral blind spots analogous to historical injustices like slavery or factory farming. They warn of the risk of large-scale suffering if sentient AI is created and carelessly exploited.

Future of Artificial Intelligence

The future of AI is a subject of intense speculation and research, with several transformative possibilities being explored.

Superintelligence and the Singularity

Superintelligence is a hypothetical form of AI that would possess intelligence far surpassing that of the brightest and most gifted human minds.

Superintelligence Definition: A hypothetical level of artificial intelligence that greatly exceeds the cognitive abilities of humans in virtually all domains of interest. Superintelligence is often envisioned as a future stage of AI development where AI systems become vastly more intelligent than humans and potentially capable of self-improvement and autonomous goal-setting.

If research into artificial general intelligence (AGI) leads to sufficiently intelligent software, it might be able to self-improve and reprogram itself, leading to an “intelligence explosion” or “singularity” as described by I.J. Good and Vernor Vinge.

Intelligence Explosion Definition: A hypothetical scenario where a sufficiently advanced artificial intelligence (AI) system becomes capable of recursively self-improving its own intelligence at an accelerating rate, leading to a rapid and dramatic increase in intelligence levels far beyond human capabilities.

Singularity (Technological Singularity) Definition: A hypothetical point in time when technological progress becomes so rapid and profound that it leads to fundamental and unpredictable changes in human civilization. In the context of AI, the singularity is often associated with the emergence of superintelligence and the potential for runaway technological growth.

However, it is important to note that exponential growth in technology is not indefinite. Technologies typically follow an S-shaped curve, with growth slowing down as they approach physical limits.

Transhumanism

Transhumanism is a movement that explores the possibility of using technology to enhance human capabilities beyond current biological limitations.

Transhumanism Definition: An intellectual and cultural movement that believes in using science and technology to enhance human physical, intellectual, and psychological capacities and overcome human limitations, such as aging, disease, and death. Transhumanists often advocate for the development and use of technologies like AI, biotechnology, and nanotechnology to achieve human enhancement and transformation.

Robot designer Hans Moravec, cyberneticist Kevin Warwick, and inventor Ray Kurzweil have predicted that humans and machines may merge in the future into cyborgs, creating beings more capable and powerful than either humans or machines alone.

Cyborg Definition: A being with both biological and artificial (e.g., electronic, mechanical) parts. In transhumanist and science fiction contexts, cyborgs are often envisioned as humans enhanced with technology to overcome limitations or gain new abilities.

This idea has roots in earlier writings by Aldous Huxley and Robert Ettinger.

Edward Fredkin argues that “artificial intelligence is the next step in evolution,” an idea first proposed by Samuel Butler and expanded upon by George Dyson in his book Darwin Among the Machines: The Evolution of Global Intelligence.

Decomputing

Decomputing is a concept that argues for opposing the widespread application and expansion of artificial intelligence, similar to the concept of degrowth in economics.

Decomputing Definition: A concept that advocates for resisting the widespread application and expansion of artificial intelligence, particularly in contexts where it is seen as dehumanizing, harmful, or contributing to societal problems. Decomputing emphasizes reducing reliance on AI and promoting human-centered alternatives.

Proponents of decomputing, like Dan McQuillan, argue that AI is an outgrowth of systemic issues and capitalist structures and that a different future is possible, one where human connection and community are prioritized over AI intermediaries. Decomputing critiques AI as potentially increasing social distance and promoting automation at the expense of human well-being.

Artificial Intelligence in Fiction

Thought-capable artificial beings have been a recurring theme in storytelling since antiquity and are a persistent trope in science fiction.

A common trope, starting with Mary Shelley’s Frankenstein, is the idea of a human creation becoming a threat to its creator. This theme is explored in works like:

In contrast, loyal and benevolent robots, like Gort from The Day the Earth Stood Still (1951) and Bishop from Aliens (1986), are less common in popular culture.

Isaac Asimov introduced the Three Laws of Robotics in many stories, particularly in relation to his “Multivac” super-intelligent computer. Asimov’s laws are often discussed in lay discussions of machine ethics.

Asimov’s Three Laws of Robotics Definition: A set of ethical rules for robots introduced by science fiction author Isaac Asimov in his stories. The Three Laws are:

  1. A robot may not injure a human being or, through inaction, allow a human being to come to harm.
  2. A robot must obey the orders given it by human beings except where such orders would conflict with the First Law.
  3. A robot must protect its own existence as long as such protection does not conflict with the First or Second Laws.

While widely known through popular culture, AI researchers generally consider Asimov’s laws to be impractical and ambiguous for real-world AI ethics.

Several works use AI to explore the fundamental question of what makes us human, depicting artificial beings with the capacity to feel and suffer. Examples include:

Philip K. Dick’s work, in particular, explores how technology created with artificial intelligence can alter our understanding of human subjectivity.

See Also