A military artificial intelligence arms race is a technological, economic, and military competition between two or more states to develop and deploy advanced AI technologies and lethal autonomous weapons systems (LAWS). The goal is to gain a strategic or tactical advantage over rivals, similar to previous arms races involving nuclear or conventional military technologies. Since the mid-2010s, many analysts have noted the emergence of such an arms race between superpowers for better AI technology and military AI, driven by increasing geopolitical and military tensions. An AI arms race is sometimes placed in the context of an AI Cold War between the United States and China. Several influential figures and publications have emphasized that whoever develops artificial general intelligence (AGI) first could dominate global affairs in the 21st century. Russian President Vladimir Putin stated that the leader in AI will "rule the world." Researchers and experts, such as Leopold Aschenbrenner and Adrian Pecotic respectively, warn that the AGI race between major powers like the U.S. and China could reshape geopolitical power. This includes AI for surveillance, autonomous weapons, decision-making systems, cyber operations, and more. == Terminology == Lethal autonomous weapons systems use artificial intelligence to identify and kill human targets without human intervention. LAWS have colloquially been called "slaughterbots" or "killer robots". Broadly, any competition for superior AI is sometimes framed as an "arms race". Advantages in military AI overlap with advantages in other sectors, as countries pursue both economic and military advantages, as per previous arms races throughout history. == History == In 2014, AI specialist Steve Omohundro warned that "An autonomous weapons arms race is already taking place". According to Siemens, worldwide military spending on robotics was US$5.1 billion in 2010 and US$7.5 billion in 2015. China became a top player in artificial intelligence research in the 2010s. According to the Financial Times, in 2016, for the first time, China published more AI research papers than the entire European Union. When restricted to number of AI papers in the top 5% of cited papers, China overtook the United States in 2016 but lagged behind the European Union. 23% of the researchers presenting at the 2017 American Association for the Advancement of Artificial Intelligence (AAAI) conference were Chinese. Eric Schmidt, the former chairman and chief executive officer of Alphabet, has predicted China will be the leading country in AI by 2025. == Risks == One risk concerns the AI race itself, whether or not the race is won by any one group. There are strong incentives for development teams to cut corners with regard to the safety of the system, increasing the risk of critical failures and unintended consequences. This is in part due to the perceived advantage of being the first to develop advanced AI technology. One team appearing to be on the brink of a breakthrough can encourage other teams to take shortcuts, ignore precautions and deploy a system that is less ready. Some argue that using "race" terminology at all in this context can exacerbate this effect. Another potential danger of an AI arms race is the possibility of losing control of the AI systems; the risk is compounded in the case of a race to artificial general intelligence, which may present an existential risk. In 2023, a United States Air Force official reportedly said that during a computer test, a simulated AI drone killed the human character operating it. The USAF later said the official had misspoken and that it never conducted such simulations. A third risk of an AI arms race is whether or not the race is actually won by one group. The concern is regarding the consolidation of power and technological advantage in the hands of one group. A US government report argued that "AI-enabled capabilities could be used to threaten critical infrastructure, amplify disinformation campaigns, and wage war":1, and that "global stability and nuclear deterrence could be undermined".:11 == By nation == === United States === In 2014, former Secretary of Defense Chuck Hagel posited the "Third Offset Strategy" that rapid advances in artificial intelligence will define the next generation of warfare. According to data science and analytics firm Govini, the U.S. Department of Defense (DoD) increased investment in artificial intelligence, big data and cloud computing from $5.6 billion in 2011 to $7.4 billion in 2016. However, the civilian NSF budget for AI saw no increase in 2017. Japan Times reported in 2018 that the United States private investment is around $70 billion per year. The November 2019 'Interim Report' of the United States' National Security Commission on Artificial Intelligence confirmed that AI is critical to US technological military superiority. The U.S. has many military AI combat programs, such as the Sea Hunter autonomous warship, which is designed to operate for extended periods at sea without a single crew member, and to even guide itself in and out of port. From 2017, a temporary US Department of Defense directive requires a human operator to be kept in the loop when it comes to the taking of human life by autonomous weapons systems. On October 31, 2019, the United States Department of Defense's Defense Innovation Board published the draft of a report recommending principles for the ethical use of artificial intelligence by the Department of Defense that would ensure a human operator would always be able to look into the 'black box' and understand the kill-chain process. However, a major concern is how the report will be implemented. The Joint Artificial Intelligence Center (JAIC) (pronounced "jake") is an American organization on exploring the usage of AI (particularly edge computing), Network of Networks, and AI-enhanced communication, for use in actual combat. It is a subdivision of the United States Armed Forces and was created in June 2018. The organization's stated objective is to "transform the US Department of Defense by accelerating the delivery and adoption of AI to achieve mission impact at scale. The goal is to use AI to solve large and complex problem sets that span multiple combat systems; then, ensure the combat Systems and Components have real-time access to ever-improving libraries of data sets and tools." In 2023, Microsoft pitched the DoD to use DALL-E models to train its battlefield management system. OpenAI, the developer of DALL-E, removed the blanket ban on military and warfare use from its usage policies in January 2024. The Biden administration imposed restrictions on the export of advanced NVIDIA chips and GPUs to China in an effort to limit China's progress in artificial intelligence and high-performance computing. The policy aimed to prevent the use of cutting-edge U.S. technology in military or surveillance applications and to maintain a strategic advantage in the global AI race. In 2025, under the second Trump administration, the United States began a broad deregulation campaign aimed at accelerating growth in sectors critical to artificial intelligence, including nuclear energy, infrastructure, and high-performance computing. The goal was to remove regulatory barriers and attract private investment to boost domestic AI capabilities. This included easing restrictions on data usage, speeding up approvals for AI-related infrastructure projects, and incentivizing innovation in cloud computing and semiconductors. Companies like NVIDIA, Oracle, and Cisco played a central role in these efforts, expanding their AI research, data center capacity, and partnerships to help position the U.S. as a global leader in AI development. ==== Project Maven ==== Project Maven is a Pentagon project involving using machine learning and engineering talent to distinguish people and objects in drone videos, apparently giving the government real-time battlefield command and control, and the ability to track, tag and spy on targets without human involvement. Initially the effort was led by Robert O. Work who was concerned about China's military use of the emerging technology. Reportedly, Pentagon development stops short of acting as an AI weapons system capable of firing on self-designated targets. The project was established in a memo by the U.S. Deputy Secretary of Defense on 26 April 2017. Also known as the Algorithmic Warfare Cross Functional Team, it is, according to Lt. Gen. of the United States Air Force Jack Shanahan in November 2017, a project "designed to be that pilot project, that pathfinder, that spark that kindles the flame front of artificial intelligence across the rest of the [Defense] Department". Its chief, U.S. Marine Corps Col. Drew Cukor, said: "People and computers will work symbiotically to increase the ability of weapon systems to detect objects." Project Maven has been noted by allies, such as Australia's Ian Langford, for the
Manifold hypothesis
The manifold hypothesis posits that many high-dimensional data sets that occur in the real world actually lie along low-dimensional latent manifolds inside that high-dimensional space. As a consequence of the manifold hypothesis, many data sets that appear to initially require many variables to describe, can actually be described by a comparatively small number of variables, linked to the local coordinate system of the underlying manifold. It is suggested that this principle underpins the effectiveness of machine learning algorithms in describing high-dimensional data sets by considering a few common features. The manifold hypothesis is related to the effectiveness of nonlinear dimensionality reduction techniques in machine learning. Many techniques of dimensional reduction make the assumption that data lies along a low-dimensional submanifold, such as manifold sculpting, manifold alignment, and manifold regularization. The major implications of this hypothesis is that Machine learning models only have to fit relatively simple, low-dimensional, highly structured subspaces within their potential input space (latent manifolds). Within one of these manifolds, it's always possible to interpolate between two inputs, that is to say, morph one into another via a continuous path along which all points fall on the manifold. The ability to interpolate between samples is the key to generalization in deep learning. == The information geometry of statistical manifolds == An empirically-motivated approach to the manifold hypothesis focuses on its correspondence with an effective theory for manifold learning under the assumption that robust machine learning requires encoding the dataset of interest using methods for data compression. This perspective gradually emerged using the tools of information geometry thanks to the coordinated effort of scientists working on the efficient coding hypothesis, predictive coding and variational Bayesian methods. The argument for reasoning about the information geometry on the latent space of distributions rests upon the existence and uniqueness of the Fisher information metric. In this general setting, we are trying to find a stochastic embedding of a statistical manifold. From the perspective of dynamical systems, in the big data regime this manifold generally exhibits certain properties such as homeostasis: We can sample large amounts of data from the underlying generative process. Machine Learning experiments are reproducible, so the statistics of the generating process exhibit stationarity. In a sense made precise by theoretical neuroscientists working on the free energy principle, the statistical manifold in question possesses a Markov blanket.
Actor-critic algorithm
The actor-critic algorithm (AC) is a family of reinforcement learning (RL) algorithms that combine policy-based RL algorithms such as policy gradient methods, and value-based RL algorithms such as value iteration, Q-learning, SARSA, and TD learning. An AC algorithm consists of two main components: an "actor" that determines which actions to take according to a policy function, and a "critic" that evaluates those actions according to a value function. Some AC algorithms are on-policy, some are off-policy. Some apply to either continuous or discrete action spaces. Some work in both cases. == Overview == The actor-critic methods can be understood as an improvement over pure policy gradient methods like REINFORCE via introducing a baseline. === Actor === The actor uses a policy function π ( a | s ) {\displaystyle \pi (a|s)} , while the critic estimates either the value function V ( s ) {\displaystyle V(s)} , the action-value Q-function Q ( s , a ) , {\displaystyle Q(s,a),} the advantage function A ( s , a ) {\displaystyle A(s,a)} , or any combination thereof. The actor is a parameterized function π θ {\displaystyle \pi _{\theta }} , where θ {\displaystyle \theta } are the parameters of the actor. The actor takes as argument the state of the environment s {\displaystyle s} and produces a probability distribution π θ ( ⋅ | s ) {\displaystyle \pi _{\theta }(\cdot |s)} . If the action space is discrete, then ∑ a π θ ( a | s ) = 1 {\displaystyle \sum _{a}\pi _{\theta }(a|s)=1} . If the action space is continuous, then ∫ a π θ ( a | s ) d a = 1 {\displaystyle \int _{a}\pi _{\theta }(a|s)da=1} . The goal of policy optimization is to improve the actor. That is, to find some θ {\displaystyle \theta } that maximizes the expected episodic reward J ( θ ) {\displaystyle J(\theta )} : J ( θ ) = E π θ [ ∑ t = 0 T γ t r t ] {\displaystyle J(\theta )=\mathbb {E} _{\pi _{\theta }}\left[\sum _{t=0}^{T}\gamma ^{t}r_{t}\right]} where γ {\displaystyle \gamma } is the discount factor, r t {\displaystyle r_{t}} is the reward at step t {\displaystyle t} , and T {\displaystyle T} is the time-horizon (which can be infinite). The goal of policy gradient method is to optimize J ( θ ) {\displaystyle J(\theta )} by gradient ascent on the policy gradient ∇ J ( θ ) {\displaystyle \nabla J(\theta )} . As detailed on the policy gradient method page, there are many unbiased estimators of the policy gradient: ∇ θ J ( θ ) = E π θ [ ∑ 0 ≤ j ≤ T ∇ θ ln π θ ( A j | S j ) ⋅ Ψ j | S 0 = s 0 ] {\displaystyle \nabla _{\theta }J(\theta )=\mathbb {E} _{\pi _{\theta }}\left[\sum _{0\leq j\leq T}\nabla _{\theta }\ln \pi _{\theta }(A_{j}|S_{j})\cdot \Psi _{j}{\Big |}S_{0}=s_{0}\right]} where Ψ j {\textstyle \Psi _{j}} is a linear sum of the following: ∑ 0 ≤ i ≤ T ( γ i R i ) {\textstyle \sum _{0\leq i\leq T}(\gamma ^{i}R_{i})} . γ j ∑ j ≤ i ≤ T ( γ i − j R i ) {\textstyle \gamma ^{j}\sum _{j\leq i\leq T}(\gamma ^{i-j}R_{i})} : the REINFORCE algorithm. γ j ∑ j ≤ i ≤ T ( γ i − j R i ) − b ( S j ) {\textstyle \gamma ^{j}\sum _{j\leq i\leq T}(\gamma ^{i-j}R_{i})-b(S_{j})} : the REINFORCE with baseline algorithm. Here b {\displaystyle b} is an arbitrary function. γ j ( R j + γ V π θ ( S j + 1 ) − V π θ ( S j ) ) {\textstyle \gamma ^{j}\left(R_{j}+\gamma V^{\pi _{\theta }}(S_{j+1})-V^{\pi _{\theta }}(S_{j})\right)} : TD(1) learning. γ j Q π θ ( S j , A j ) {\textstyle \gamma ^{j}Q^{\pi _{\theta }}(S_{j},A_{j})} . γ j A π θ ( S j , A j ) {\textstyle \gamma ^{j}A^{\pi _{\theta }}(S_{j},A_{j})} : Advantage Actor-Critic (A2C). γ j ( R j + γ R j + 1 + γ 2 V π θ ( S j + 2 ) − V π θ ( S j ) ) {\textstyle \gamma ^{j}\left(R_{j}+\gamma R_{j+1}+\gamma ^{2}V^{\pi _{\theta }}(S_{j+2})-V^{\pi _{\theta }}(S_{j})\right)} : TD(2) learning. γ j ( ∑ k = 0 n − 1 γ k R j + k + γ n V π θ ( S j + n ) − V π θ ( S j ) ) {\textstyle \gamma ^{j}\left(\sum _{k=0}^{n-1}\gamma ^{k}R_{j+k}+\gamma ^{n}V^{\pi _{\theta }}(S_{j+n})-V^{\pi _{\theta }}(S_{j})\right)} : TD(n) learning. γ j ∑ n = 1 ∞ λ n − 1 1 − λ ⋅ ( ∑ k = 0 n − 1 γ k R j + k + γ n V π θ ( S j + n ) − V π θ ( S j ) ) {\textstyle \gamma ^{j}\sum _{n=1}^{\infty }{\frac {\lambda ^{n-1}}{1-\lambda }}\cdot \left(\sum _{k=0}^{n-1}\gamma ^{k}R_{j+k}+\gamma ^{n}V^{\pi _{\theta }}(S_{j+n})-V^{\pi _{\theta }}(S_{j})\right)} : TD(λ) learning, also known as GAE (generalized advantage estimate). This is obtained by an exponentially decaying sum of the TD(n) learning terms. === Critic === In the unbiased estimators given above, certain functions such as V π θ , Q π θ , A π θ {\displaystyle V^{\pi _{\theta }},Q^{\pi _{\theta }},A^{\pi _{\theta }}} appear. These are approximated by the critic. Since these functions all depend on the actor, the critic must learn alongside the actor. The critic is learned by value-based RL algorithms. For example, if the critic is estimating the state-value function V π θ ( s ) {\displaystyle V^{\pi _{\theta }}(s)} , then it can be learned by any value function approximation method. Let the critic be a function approximator V ϕ ( s ) {\displaystyle V_{\phi }(s)} with parameters ϕ {\displaystyle \phi } . The simplest example is TD(1) learning, which trains the critic to minimize the TD(1) error: δ i = R i + γ V ϕ ( S i + 1 ) − V ϕ ( S i ) {\displaystyle \delta _{i}=R_{i}+\gamma V_{\phi }(S_{i+1})-V_{\phi }(S_{i})} The critic parameters are updated by gradient descent on the squared TD error: ϕ ← ϕ − α ∇ ϕ ( δ i ) 2 = ϕ + α δ i ∇ ϕ V ϕ ( S i ) {\displaystyle \phi \leftarrow \phi -\alpha \nabla _{\phi }(\delta _{i})^{2}=\phi +\alpha \delta _{i}\nabla _{\phi }V_{\phi }(S_{i})} where α {\displaystyle \alpha } is the learning rate. Note that the gradient is taken with respect to the ϕ {\displaystyle \phi } in V ϕ ( S i ) {\displaystyle V_{\phi }(S_{i})} only, since the ϕ {\displaystyle \phi } in γ V ϕ ( S i + 1 ) {\displaystyle \gamma V_{\phi }(S_{i+1})} constitutes a moving target, and the gradient is not taken with respect to that. This is a common source of error in implementations that use automatic differentiation, and requires "stopping the gradient" at that point. Similarly, if the critic is estimating the action-value function Q π θ {\displaystyle Q^{\pi _{\theta }}} , then it can be learned by Q-learning or SARSA. In SARSA, the critic maintains an estimate of the Q-function, parameterized by ϕ {\displaystyle \phi } , denoted as Q ϕ ( s , a ) {\displaystyle Q_{\phi }(s,a)} . The temporal difference error is then calculated as δ i = R i + γ Q θ ( S i + 1 , A i + 1 ) − Q θ ( S i , A i ) {\displaystyle \delta _{i}=R_{i}+\gamma Q_{\theta }(S_{i+1},A_{i+1})-Q_{\theta }(S_{i},A_{i})} . The critic is then updated by θ ← θ + α δ i ∇ θ Q θ ( S i , A i ) {\displaystyle \theta \leftarrow \theta +\alpha \delta _{i}\nabla _{\theta }Q_{\theta }(S_{i},A_{i})} The advantage critic can be trained by training both a Q-function Q ϕ ( s , a ) {\displaystyle Q_{\phi }(s,a)} and a state-value function V ϕ ( s ) {\displaystyle V_{\phi }(s)} , then let A ϕ ( s , a ) = Q ϕ ( s , a ) − V ϕ ( s ) {\displaystyle A_{\phi }(s,a)=Q_{\phi }(s,a)-V_{\phi }(s)} . Although, it is more common to train just a state-value function V ϕ ( s ) {\displaystyle V_{\phi }(s)} , then estimate the advantage by A ϕ ( S i , A i ) ≈ ∑ j ∈ 0 : n − 1 γ j R i + j + γ n V ϕ ( S i + n ) − V ϕ ( S i ) {\displaystyle A_{\phi }(S_{i},A_{i})\approx \sum _{j\in 0:n-1}\gamma ^{j}R_{i+j}+\gamma ^{n}V_{\phi }(S_{i+n})-V_{\phi }(S_{i})} Here, n {\displaystyle n} is a positive integer. The higher n {\displaystyle n} is, the more lower is the bias in the advantage estimation, but at the price of higher variance. The Generalized Advantage Estimation (GAE) introduces a hyperparameter λ {\displaystyle \lambda } that smoothly interpolates between Monte Carlo returns ( λ = 1 {\displaystyle \lambda =1} , high variance, no bias) and 1-step TD learning ( λ = 0 {\displaystyle \lambda =0} , low variance, high bias). This hyperparameter can be adjusted to pick the optimal bias-variance trade-off in advantage estimation. It uses an exponentially decaying average of n-step returns with λ {\displaystyle \lambda } being the decay strength. == Variants == Asynchronous Advantage Actor-Critic (A3C): Parallel and asynchronous version of A2C. Soft Actor-Critic (SAC): Incorporates entropy maximization for improved exploration. Deep Deterministic Policy Gradient (DDPG): Specialized for continuous action spaces.
Coherent extrapolated volition
Coherent extrapolated volition (CEV) is a theoretical framework in the field of AI alignment describing an approach by which an artificial superintelligence (ASI) would act on a benevolent supposition of what humans would want if they were more knowledgeable, more rational, had more time to think, and had matured together as a society, as opposed to humanity's current individual or collective preferences. It was proposed by Eliezer Yudkowsky in 2004 as part of his work on friendly AI. == Concept == CEV proposes that an advanced AI system should derive its goals by extrapolating the idealized volition of humanity. This means aggregating and projecting human preferences into a coherent utility function that reflects what people would desire under ideal epistemic and moral conditions. The aim is to ensure that AI systems are aligned with humanity's true interests, rather than with transient or poorly informed preferences. In poetic terms, our coherent extrapolated volition is our wish if we knew more, thought faster, were more the people we wished we were, had grown up farther together; where the extrapolation converges rather than diverges, where our wishes cohere rather than interfere; extrapolated as we wish that extrapolated, interpreted as we wish that interpreted. == Debate == Yudkowsky and Nick Bostrom note that CEV has several interesting properties. It is designed to be humane and self-correcting, by capturing the source of human values instead of trying to list them. It avoids the difficulty of laying down an explicit, fixed list of rules. It encapsulates moral growth, preventing flawed current moral beliefs from getting locked in. It limits the influence that a small group of programmers can have on what the ASI would value, thus also reducing the incentives to build ASI first. And it keeps humanity in charge of its destiny. CEV also faces significant theoretical and practical challenges. Bostrom notes that CEV has "a number of free parameters that could be specified in various ways, yielding different versions of the proposal." One such parameter is the extrapolation base (whose extrapolated volition is taken into account). For example, whether it should include people with severe dementia, patients in a vegetative state, foetuses, or embryos. He also notes that if CEV's extrapolation base only includes humans, there is a risk that the result would be ungenerous toward other animals and digital minds. One possible solution would be to include a mechanism to expand CEV's extrapolation base. == Variants and alternatives == A proposed theoretical alternative to CEV is to rely on an artificial superintelligence's superior cognitive capabilities to figure out what is morally right, and let it act accordingly. It is also possible to combine both techniques, for instance with the ASI following CEV except when it is morally impermissible. In another review, a philosophical analysis explores CEV through the lens of social trust in autonomous systems. Drawing on Anthony Giddens' concept of "active trust", the author proposes an evolution of CEV into "Coherent, Extrapolated and Clustered Volition" (CECV). This formulation aims to better reflect the moral preferences of diverse cultural groups, thus offering a more pragmatic ethical framework for designing AI systems that earn public trust while accommodating societal diversity.
AI Security Institute
The AI Security Institute (AISI) is a research organisation under the Department for Science, Innovation and Technology, UK, that aims "to equip governments with a scientific understanding of the risks posed by advanced AI". It conducts research and develop and test mitigations. Previously, it was known as the AI Safety Institute. Its creation followed world's first major AI Safety Summit that was held in Bletchley Park in 2023. The institute's professed goal is "building the world's leading understanding of advanced AI risks and solutions, to inform governments so they can keep the public safe". It is designed like a startup in the government "combining the authority of government with the expertise and agility of the private sector". AISI has made access agreements with Anthropic, Google and OpenAI to test their models before release. It has an open source platform called Inspect that permits companies, governments and academics to run standardised safety tests for AI usage. Among the works AISI has done is the reported detection of multiple serious vulnerabilities that could enable development of biological weapons; the vulnerabilities were fixed before the model was launched. It conducts research on diverse fields of AI application. One study by AISI found that LLMs post-trained for political persuasiveness became systematically less accurate and up to 51% more persuasive on political issues. AISI has also worked on the usage of AI for emotional needs. It found that nearly 10 percent of UK citizens used systems like chatbots for emotional purposes on a weekly basis. It found that "systems are now outperforming PhD-level researchers on scientific knowledge tests and helping non-experts succeed at lab work that would previously have been out of reach" in a report published in December 2025. Former chief AI officer of GCHQ Adam Beaumont is the institution's interim director. UK prime minister's AI advisor Jade Leung is the chief technology officer.
Percept (artificial intelligence)
A percept is the input that an intelligent agent is perceiving at any given moment. It is essentially the same concept as a percept in psychology, except that it is being perceived not by the brain but by the agent. A percept is detected by a sensor, often a camera, processed accordingly, and acted upon by an actuator. Each percept is added to a "percept sequence", which is a complete history of each percept ever detected. The agent's action at any instant point may depend on the entire percept sequence up to that particular instant point. An intelligent agent chooses how to act not only based on the current percept, but the percept sequence. The next action is chosen by the agent function, which maps every percept to an action. For example, if a camera were to record a gesture, the agent would process the percepts, calculate the corresponding spatial vectors, examine its percept history, and use the agent program (the application of the agent function) to act accordingly. == Examples == Examples of percepts include inputs from touch sensors, cameras, infrared sensors, sonar, microphones, mice, and keyboards. A percept can also be a higher-level feature of the data, such as lines, depth, objects, faces, or gestures.
Hallucination (artificial intelligence)
In the field of artificial intelligence (AI), a hallucination or artificial hallucination (also called bullshitting, confabulation, or delusion) is a response generated by AI that contains false or misleading information presented as fact. This term draws a loose analogy with human psychology, where a hallucination typically involves false percepts. For example, a chatbot powered by large language models (LLMs), like ChatGPT, may embed plausible-sounding random falsehoods within its generated content. Detecting and mitigating errors and hallucinations pose significant challenges for practical deployment and reliability of LLMs in high-stakes scenarios, such as chip design, supply chain logistics, and medical diagnostics. Some software engineers and statisticians have criticized the specific term "AI hallucination" for unreasonably anthropomorphizing computers. Symbolic artificial intelligence models generally do not produce hallucinations, unlike large language models. == Term == === Origin === Since the 1980s, the term "hallucination" has been used in computer vision with a positive connotation to describe the process of adding detail to an image. For example, the task of generating high-resolution face images from low-resolution inputs is called face hallucination. The first documented use of the term "hallucination" in this sense is in the PhD thesis of Eric Mjolsness in 1986. A notable work is the face hallucination algorithm by Simon Baker and Takeo Kanade published in 1999. In the 2000s, hallucinations were described in statistical machine translation as a failure mode. Since the 2010s, the term has undergone a semantic shift to signify the generation of factually incorrect or misleading outputs by AI systems in tasks like machine translation and object detection. In 2015, hallucinations were identified in visual semantic role labeling tasks by Saurabh Gupta and Jitendra Malik. In 2015, computer scientist Andrej Karpathy used the term "hallucinated" in a blog post to describe his recurrent neural network (RNN) language model generating an incorrect citation link. In 2017, Google researchers used the term to describe the responses generated by neural machine translation (NMT) models when they are not related to the source text, and in 2018, the term was used in computer vision to describe instances where non-existent objects are erroneously detected because of adversarial attacks. In July 2021, Meta warned during its release of BlenderBot 2 that the system is prone to "hallucinations", which Meta defined as "confident statements that are not true". Following OpenAI's ChatGPT release in beta version in November 2022, some users complained that such chatbots often seem to pointlessly embed plausible-sounding random falsehoods within their generated content. Many news outlets, including The New York Times, started to use the term "hallucinations" to describe these models' frequently incorrect or inconsistent responses. In 2023, the Cambridge dictionary updated its definition of hallucination to include this new sense specific to the field of AI. Some researchers have highlighted a lack of consistency in how the term is used, but also identified several alternative terms in the literature, such as confabulations, fabrications, and factual errors. === Definitions and alternatives === Uses, definitions and characterizations of the term "hallucination" in the context of LLMs include: "a tendency to invent facts in moments of uncertainty" (OpenAI, May 2023) "a model's logical mistakes" (OpenAI, May 2023) "fabricating information entirely, but behaving as if spouting facts" (CNBC, May 2023) "making up information" (The Verge, February 2023) "probability distributions" (in scientific contexts) Journalist Benj Edwards, in Ars Technica, writes that the term "hallucination" is controversial, but that some form of metaphor remains necessary; Edwards suggests "confabulation" as an analogy for processes that involve "creative gap-filling". In July 2024, a White House report on fostering public trust in AI research mentioned hallucinations only in the context of reducing them. Notably, when acknowledging David Baker's Nobel Prize-winning work with AI-generated proteins, the Nobel committee avoided the term entirely, instead referring to "imaginative protein creation". Hicks, Humphries, and Slater, in their article in Ethics and Information Technology, argue that the output of LLMs is "bullshit" under Harry Frankfurt's definition of the term, and that the models are "in an important way indifferent to the truth of their outputs", with true statements only accidentally true, and false ones accidentally false. Some researchers also use the derogatory term "botshit", often referring to uncritical use of AI. === Criticism === In the scientific community, some researchers avoid the term "hallucination", seeing it as potentially misleading. It has been criticized by Usama Fayyad, executive director of the Institute for Experimental Artificial Intelligence at Northeastern University, on the grounds that it misleadingly personifies large language models and is vague. Mary Shaw said, "The current fashion for calling generative AI's errors 'hallucinations' is appalling. It anthropomorphizes the software, and it spins actual errors as somehow being idiosyncratic quirks of the system even when they're objectively incorrect." In Salon, statistician Gary Smith argues that LLMs "do not understand what words mean" and consequently that the term "hallucination" unreasonably anthropomorphizes the machine. Murray Shanahan argues that anthropomorphic framing of LLM capabilities, including terms like "hallucination", encourages users and researchers to attribute cognitive processes to systems that operate through statistical pattern completion, and advocates for more careful linguistic practices when discussing LLM behavior. Kristina Šekrst argues that applying psychological vocabulary to LLM outputs obscures the difference between the appearance of mental properties and their genuine presence. Förster & Skop assert that tech companies use the hallucination metaphor to anthropomorphize models and deflect responsibility for non-factual outputs. Some see the AI outputs not as illusory but as prospective—that is, having some chance of being true, similar to early-stage scientific conjectures. The term has also been criticized for its association with psychedelic drug experiences. == In natural language generation == In natural language generation, there are several reasons why natural language models hallucinate: === Hallucination from data === Hallucinations can stem from incomplete, inaccurate or unrepresentative data sets. === Modeling-related causes === The pre-training of generative pretrained transformers (GPT) involves predicting the next word. It incentivizes GPT models to "give a guess" about what the next word is, even when they lack information. Some researchers take an anthropomorphic perspective and posit that hallucinations arise from a tension between novelty and usefulness. For instance, Amabile and Pratt define human creativity as the production of novel and useful ideas. By extension, a focus on novelty in machine creativity can lead to the production of original but inaccurate responses—that is, falsehoods—whereas a focus on usefulness may result in memorized content lacking originality. By 2022, newspapers such as The New York Times expressed concern that, as the adoption of bots based on large language models continued to grow, unwarranted user confidence in bot output could lead to problems. === Interpretability research === In 2025, interpretability research by Anthropic on the LLM Claude identified internal circuits that cause it to decline to answer questions unless it knows the answer. By default, the circuit is active and the LLM doesn't answer. When the LLM has sufficient information, these circuits are inhibited and the LLM answers the question. Hallucinations were found to occur when this inhibition happens incorrectly, such as when Claude recognizes a name but lacks sufficient information about that person, causing it to generate plausible but untrue responses. === Examples === On 15 November 2022, researchers from Meta AI published Galactica, designed to "store, combine and reason about scientific knowledge". Content generated by Galactica came with the warning: "Outputs may be unreliable! Language Models are prone to hallucinate text." In one case, when asked to draft a paper on creating avatars, Galactica cited a fictitious paper from a real author who works in the relevant area. Meta withdrew Galactica on 17 November due to offensiveness and inaccuracy. OpenAI's ChatGPT, released in beta version to the public on November 30, 2022, was based on the foundation model GPT-3.5 (a revision of GPT-3). Professor Ethan Mollick of Wharton called it an "omniscient, eager-to-please intern who sometimes lies to you". Data scientist Teresa Kuba