CtrlK
See also: Machine learning, Deep learning, AI safety
Artificial general intelligence (AGI) is a type of artificial intelligence that can match or exceed human cognitive abilities across virtually any intellectual task. Unlike narrow AI systems (also called artificial narrow intelligence, or ANI), which are built to perform specific functions like image recognition or playing chess, an AGI system would be able to generalize knowledge, transfer skills between domains, and solve new problems without task-specific programming.
The concept of AGI sits at the core of the AI field's longest-running ambition: building a machine that thinks and reasons the way humans do. While modern AI systems have achieved superhuman performance in many isolated tasks, none of them possess the flexible, general-purpose intelligence that characterizes human cognition. A chess engine can beat any grandmaster but cannot hold a conversation. A large language model can write an essay but cannot reliably plan a multi-step physical task. AGI would, in theory, handle all of these and more.
The term is sometimes used interchangeably with "strong AI" or "full AI," though these labels carry slightly different connotations depending on the context. The pursuit of AGI remains one of the most debated topics in computer science, philosophy, and public policy.
Imagine you have a toy robot that is really, really good at sorting colored blocks. It can sort them faster and more accurately than any person. But if you ask that robot to draw a picture, tell you a story, or make you a sandwich, it has no idea what to do. That is what today's AI is like: it can do one thing well but nothing else.
AGI would be like a robot that can do anything a person can do with their brain. It could sort blocks, draw pictures, tell stories, learn new games, understand jokes, and figure out problems it has never seen before. Nobody has built a robot like that yet, and scientists disagree about when (or whether) it will happen.
Artificial intelligence is often categorized into three broad levels based on the range and depth of capabilities.
| Type | Full name | Description | Examples |
|---|---|---|---|
| ANI | Artificial narrow intelligence | Excels at a single task or narrow set of tasks; cannot generalize | Spam filters, image classifiers, chess engines, voice assistants |
| AGI | Artificial general intelligence | Matches or exceeds human cognitive performance across all intellectual domains | None yet achieved |
| ASI | Artificial superintelligence | Surpasses the cognitive abilities of all humans in every domain, including creativity, social reasoning, and scientific discovery | Hypothetical |
Narrow AI is the only form of AI that exists today. Every commercial AI product, from recommendation engines to self-driving car perception systems, falls under ANI. AGI represents a qualitative leap: not just better performance on known tasks, but the ability to handle novel, open-ended problems across domains. ASI goes further still, describing a system that exceeds human intelligence in every measurable way. Whether ASI is an inevitable consequence of achieving AGI is itself a matter of significant debate.
The idea that machines could think like humans predates the field of computer science itself, but it took formal shape in the middle of the 20th century. In 1950, Alan Turing published his paper "Computing Machinery and Intelligence" in the journal Mind, where he proposed what became known as the Turing test. The test involves a human evaluator judging a natural-language conversation between a human and a machine; if the evaluator cannot reliably tell which is which, the machine is said to exhibit intelligent behavior. Turing's paper opened with the question "Can machines think?" and laid much of the philosophical groundwork for AI research.
In 1956, John McCarthy, Marvin Minsky, Nathaniel Rochester, and Claude Shannon organized the Dartmouth Conference, a summer workshop that is widely considered the founding event of artificial intelligence as an academic discipline. Their proposal stated that "every aspect of learning or any other feature of intelligence can be so precisely described that a machine can be made to simulate it." It was at this workshop that McCarthy coined the term "artificial intelligence."
The decade that followed was marked by intense optimism. Herbert Simon predicted in 1965 that "machines will be capable, within twenty years, of doing any work a man can do." Minsky told Life magazine in 1970 that "in from three to eight years we will have a machine with the general intelligence of an average human being." DARPA funded AI research generously during this period, with few requirements for near-term practical results.
These predictions proved wildly premature. By the early 1970s, progress had stalled on several fronts, and the gap between the ambitions of AI researchers and their actual achievements became hard to ignore.
In 1973, the British mathematician Sir James Lighthill published a report commissioned by the UK Parliament evaluating the state of AI research. The Lighthill Report criticized the "utter failure" of AI to achieve its stated goals and identified combinatorial explosion as a fundamental barrier to scaling up AI techniques. The report led to severe cuts in AI funding across the United Kingdom, and AI research continued at only a handful of British universities.
In the United States, the passage of the Mansfield Amendment in 1969 required DARPA to fund only "mission-oriented direct research" rather than open-ended basic research. When AI projects failed to deliver practical military applications, DARPA pulled back funding starting in 1974. The period from roughly 1974 to 1980 is known as the first AI winter, a time when funding dried up, researchers left the field, and public interest faded.
A second AI winter followed in the late 1980s and early 1990s, triggered by the collapse of the expert systems market. Companies had invested heavily in rule-based expert systems during the 1980s, but these systems proved brittle, expensive to maintain, and unable to generalize beyond narrow domains. The failures reinforced skepticism about the feasibility of general-purpose machine intelligence.
The term "artificial general intelligence" as it is used today was popularized by Shane Legg and Ben Goertzel around 2002. They needed a name that would distinguish their research focus (building versatile, human-level AI) from the narrow AI systems that dominated the field. Goertzel published the edited volume Artificial General Intelligence in 2007, which collected papers from various perspectives and helped establish AGI as a recognized subfield.
Mark Gubrud had actually used the term "artificial general intelligence" in a 1997 paper about autonomous weapons presented at a nanotechnology conference, but this earlier usage attracted little attention at the time.
Goertzel also founded the AGI Conference series (starting in 2008) and the Journal of Artificial General Intelligence, both of which gave the field an institutional home.
There is no single agreed-upon definition of AGI. Different researchers, organizations, and companies use varying criteria, which makes comparing progress difficult.
The most intuitive definition holds that AGI is a machine that can perform any intellectual task that a human can. This traces back to the early AI pioneers and remains the most commonly cited version. However, it leaves ambiguity about what counts as an "intellectual task" and what level of performance qualifies.
Shane Legg and Marcus Hutter proposed a more formal definition in 2007, describing intelligence as "an agent's ability to achieve goals in a wide range of environments." Under this definition, a system is more intelligent to the extent that it performs well across more diverse settings.
OpenAI's charter, published in 2018, defines AGI as "highly autonomous systems that outperform humans at most economically valuable work." This definition is notable for its economic framing. Sam Altman has described AGI more informally as "the equivalent of a median human you could hire as a coworker" who "could do anything you'd be happy with, just behind a computer."
The economic definition has proven contentious. Microsoft and OpenAI reportedly agreed on a threshold of $100 billion in profits as a practical benchmark for when AGI has been achieved, a figure that carries legal significance because OpenAI's charter includes an "AGI clause" that could affect Microsoft's commercial licensing rights once AGI is declared. Mustafa Suleyman, CEO of Microsoft AI, has proposed the alternative concept of "Artificial Capable Intelligence" (ACI), tested by giving an AI $100,000 and seeing whether it can grow the sum to $1 million.
In July 2024, OpenAI internally presented a five-level classification of AI development:
| Level | Name | Description |
|---|---|---|
| 1 | Chatbots | AI with conversational language abilities |
| 2 | Reasoners | AI with human-level problem solving, comparable to a person with a doctoral education |
| 3 | Agents | AI that can take autonomous actions on behalf of users over extended periods |
| 4 | Innovators | AI that can generate novel ideas and aid in invention |
| 5 | Organizations | AI that can do the work of an entire organization |
OpenAI stated at the time of the announcement that it believed current systems were at Level 1 and approaching Level 2. By late 2024, with the release of the o1 and o3 reasoning models, the company claimed it had reached Level 2.
Google DeepMind published a paper in November 2023 (revised in 2024) titled "Levels of AGI: Operationalizing Progress on the Path to AGI" by Meredith Ringel Morris, Jascha Sohl-Dickstein, Noah Fiedel, and others. The framework measures AI along two axes: performance (depth) and generality (breadth). It defines six levels:
| Level | Label | Description |
|---|---|---|
| 0 | No AI | Narrow, non-AI computation |
| 1 | Emerging | Equal to or better than an unskilled human on some tasks |
| 2 | Competent | At least 50th percentile of skilled adults across most cognitive tasks |
| 3 | Expert | At least 90th percentile of skilled adults |
| 4 | Virtuoso | At least 99th percentile of skilled adults |
| 5 | Superhuman | Outperforms all humans at essentially all cognitive tasks |
The framework also defines a separate axis for autonomy, ranging from "tool" (fully in human control) through "consultant," "collaborator," and "expert" to "agent" (fully autonomous). This separation is important because high capability does not necessarily imply high autonomy, and the risks associated with each differ.
The framework separates capability from generality. A system like AlphaFold, for instance, can be extraordinarily powerful in a narrow domain (protein structure prediction) without being general at all. Under this framework, current frontier models like GPT-4 and Gemini sit roughly at Level 1 for general tasks.
Yann LeCun, formerly Meta's chief AI scientist, publicly rejected the concept of "general intelligence" as "complete BS," arguing that human intelligence itself is not truly general since humans are bad at many cognitive tasks (such as multiplying large numbers or memorizing long strings). He proposed the alternative term "Superhuman Adaptable Intelligence" (SAI), though Ben Goertzel has argued that SAI is simply a special case of AGI, not a replacement for it.
Given this definitional ambiguity, there may never be universal agreement on what qualifies as the first AGI system. As Sam Altman has acknowledged, "AGI" has become "a very sloppy term," and the goalposts tend to shift as AI systems become more capable but still fail to meet prior expectations.
Researchers are pursuing several distinct strategies for building AGI, and the field lacks consensus on which (if any) will succeed.
The scaling hypothesis holds that current deep learning architectures, particularly transformer-based large language models, will approach or reach AGI if trained on enough data with enough compute. Proponents argue that emergent capabilities (abilities that appear at scale but are absent in smaller models) suggest that continued scaling will produce increasingly general intelligence.
OpenAI, Anthropic, and Google DeepMind have all invested heavily in this direction, building models with hundreds of billions or trillions of parameters. The rapid improvement from GPT-3 to GPT-4, and then to reasoning-focused models like o1 and o3, has given scaling advocates confidence.
Empirical research on scaling laws (Kaplan et al., 2020; Hoffmann et al., 2022) shows that model performance improves predictably as compute, data, and parameters increase. If these trends hold, continued investment could produce much more capable systems.
However, by late 2024, cracks in the scaling hypothesis began to draw sharper scrutiny. Reports of diminishing returns on ever-larger models unsettled both researchers and investors. A 2024 AAAI survey found that a majority of researchers and academic scientists doubt that scaling LLMs alone can produce AGI. The pioneers of AI, including Turing Award winners, leading linguists, and cognitive scientists, cannot agree on whether scaling will ever get us to AGI, even as trillions of dollars are being wagered on this hypothesis.
Neurosymbolic AI combines neural networks (which excel at pattern recognition and learning from data) with symbolic reasoning systems (which handle logic, rules, and structured knowledge). Advocates argue that pure neural approaches lack the reliable symbolic reasoning needed for general intelligence.
Gary Marcus has been among the most vocal proponents of this approach, arguing since 2019 that large language models "don't do formal reasoning and that is a huge problem." Pioneers of deep learning, including Yann LeCun, have acknowledged that symbol manipulation may be a necessary component of human-level AI. Research on neurosymbolic AI has grown rapidly since 2020, with publications peaking at 236 in 2023. Companies like Amazon have begun applying neurosymbolic methods in production systems, including warehouse robotics and shopping assistants.
The world models approach, championed primarily by Yann LeCun, holds that general intelligence requires an internal model of how the world works. Rather than predicting the next token in a sequence (as LLMs do), a system with a world model would simulate the consequences of actions and learn the dynamics of its environment.
Meta developed this vision through the Joint Embedding Predictive Architecture (JEPA), with V-JEPA2 released in 2025. LeCun argued that current LLMs are "too limiting" for AGI because they cannot understand time, causality, or the physics of everyday objects. In November 2025, LeCun left Meta after 12 years as Chief AI Scientist to found Advanced Machine Intelligence (AMI) Labs, a startup dedicated to building world models. AMI Labs raised $1.03 billion at a $3.5 billion valuation in March 2026, representing one of the largest contrarian bets against the LLM-dominated approach to AGI.
Whole brain emulation (also called "mind uploading") takes a radically different approach: instead of designing intelligence from scratch, it attempts to create a detailed digital copy of a biological brain. The idea is that if you can simulate every neuron and synapse with sufficient accuracy, the resulting system will exhibit the same cognitive abilities as the original brain.
Progress so far has been limited to small organisms. Researchers have built closed-loop simulations of the nematode C. elegans (302 neurons) that reproduce basic behaviors. In 2024, Eon Systems completed the most detailed brain emulation ever attempted, using the complete connectome of an adult fruit fly (Drosophila) brain at nanoscale resolution.
The gap between a fruit fly and a human brain remains enormous. The human brain contains roughly 86 billion neurons. Simulating one second of human brain activity at biologically realistic resolution would require between 10^18 and 10^24 floating-point operations. The State of Brain Emulation Report 2025 estimated that fewer than 500 people worldwide work directly on brain emulation, and projected that cellular-resolution connectome data for the human brain might not be available until 2044.
Some researchers draw inspiration from biological evolution, using evolutionary algorithms to evolve neural architectures or cognitive strategies. Others focus on developmental approaches, attempting to build AI systems that learn the way children do, starting with simple sensorimotor interactions and gradually building up to abstract reasoning.
These approaches have produced interesting results in narrow settings (for example, evolving novel robot morphologies or training agents in simulated environments) but have not yet produced anything close to general intelligence.
Measuring progress toward AGI is itself a contested problem. Standard machine learning benchmarks (accuracy on image classification, performance on specific question sets) test narrow capabilities and can be gamed through benchmark-specific optimization.
The Turing test, proposed in 1950, was the first widely discussed test for machine intelligence. While it remains culturally significant, most researchers now consider it insufficient as a measure of AGI. A system can pass the Turing test through clever tricks (as demonstrated by various chatbots since the 1960s) without possessing genuine understanding or general capability.
Francois Chollet, a Google AI researcher and creator of Keras, developed the Abstraction and Reasoning Corpus (ARC) specifically to test machine reasoning and general problem-solving. Unlike most benchmarks, ARC tasks require figuring out novel visual patterns from very few examples, testing generalization rather than memorization.
In December 2024, OpenAI's o3 model scored 87.5% on ARC-AGI-1 (with high compute), surpassing the 85% human baseline for the first time. This was a dramatic jump from GPT-4's score of roughly 5% just months earlier. The 2024 ARC Prize competition saw the state-of-the-art on the private evaluation set improve from 33% to 55.5%.
However, Chollet released ARC-AGI-2 in 2025 with significantly harder tasks. The o3 model scored only 4% on ARC-AGI-2 (using $200 of compute per task), and the top competition entry reached 24% at $0.20 per task. This result demonstrated that strong performance on one version of the benchmark does not automatically transfer to harder variations of the same type of reasoning.
Various other tests and benchmarks have been proposed for measuring progress toward AGI, including the Wozniak "Coffee Test" (can a robot enter an unfamiliar house and make a cup of coffee?), the "Robot College Student Test" (can a machine enroll in a university, take classes, and earn a degree?), and the "Employment Test" (can a machine perform an economically important job at least as well as the humans who typically do it?). None of these tests have been passed.
Those who believe AGI is likely within the next decade or so typically point to several lines of evidence:
Rapid capability gains. The improvement from GPT-3 (2020) to GPT-4 (2023) to o3 (2024) has been dramatic and faster than most experts expected. Reasoning, coding, mathematical problem-solving, and scientific analysis capabilities have all improved substantially.
Scaling laws. Empirical research on scaling laws (Kaplan et al., 2020; Hoffmann et al., 2022) shows that model performance improves predictably as compute, data, and parameters increase. If these trends hold, continued investment could produce much more capable systems.
Benchmark saturation. Many benchmarks that were considered challenging just a few years ago (MMLU, HumanEval, GSM8K) are now near or at saturation, suggesting that the rate of progress has been faster than the field's ability to create new tests.
Investment levels. Tens of billions of dollars are flowing into AI development annually. Microsoft committed $13 billion to OpenAI. SoftBank announced a $100 billion AI infrastructure initiative. This level of investment creates enormous pressure toward rapid progress.
Emergent capabilities. Larger models sometimes develop abilities that were not explicitly trained, such as few-shot reasoning or code generation. Proponents argue this suggests that general intelligence could emerge from sufficient scale. However, there is active debate about whether these abilities are truly "emergent" (appearing suddenly and unpredictably) or simply appear that way depending on the choice of evaluation metric.
Skeptics offer their own set of counterarguments:
Benchmark gaming. High scores on benchmarks may reflect memorization or pattern matching rather than genuine understanding. The dramatic drop from 87.5% on ARC-AGI-1 to 4% on ARC-AGI-2 illustrates how fragile current capabilities can be.
Persistent failures. Current AI systems still make elementary reasoning errors, hallucinate facts, and struggle with multi-step planning. These failures suggest that fundamental capabilities are missing, not just that more scale is needed.
Lack of embodiment. Human intelligence developed in the context of physical bodies interacting with the real world. Some researchers argue that genuine general intelligence requires grounding in physical experience, which current AI systems lack.
The "easy bits first" problem. Progress on benchmarks may be following a pattern where the easiest improvements come first, with diminishing returns as systems approach more difficult cognitive tasks. The jump from 0% to 80% on a benchmark may require far less effort than the jump from 80% to 100%.
Diminishing returns on scaling. By late 2024, reports emerged that simply training larger models on more data was yielding smaller performance gains than in previous generations. If the scaling curve is flattening, the path from current systems to AGI may require fundamental architectural breakthroughs, not just more compute.
Majority expert opinion. A 2024 AAAI survey found that a majority of AI researchers and academic scientists doubt that scaling LLMs alone will produce AGI. While this does not rule out AGI through other means, it suggests that the most-invested-in approach may not be sufficient.
Moving goalposts. Every time AI achieves a capability once thought to require general intelligence (playing chess, writing essays, passing bar exams), the definition of AGI shifts to exclude it. This pattern raises the question of whether AGI is a well-defined target at all.
The AGI debate is shaped by a relatively small number of influential researchers whose views span a wide spectrum.
| Researcher | Affiliation | Position on AGI |
|---|---|---|
| Geoffrey Hinton | University of Toronto (retired from Google, 2023) | Believes there is a 50% probability AGI will arrive within 5 to 20 years. Estimates a 10-20% chance that superintelligent AI causes human extinction. Co-signed a 2024 letter supporting California's AI safety bill SB 1047. |
| Yoshua Bengio | Mila / University of Montreal | Shortened his estimate from "decades or centuries" to "a few years to a couple of decades." Co-signed the SB 1047 letter with Hinton and Russell. Founded LawZero to build safety guardrails. |
| Yann LeCun | AMI Labs (formerly Meta AI) | Dismisses existential risk concerns as overstated. Argues AGI remains at least a decade away because current architectures fundamentally lack world models. Left Meta in November 2025 to found AMI Labs. |
| Ilya Sutskever | Safe Superintelligence Inc. (SSI) | At NeurIPS 2024, warned that future AI will exhibit "true agency," "genuine reasoning," and "self-awareness." Co-founded SSI in June 2024; the company raised $1 billion in September 2024 and $2 billion more in March 2025 ($32 billion valuation). |
| Stuart Russell | UC Berkeley | Author of Human Compatible (2019). Advocates for AI systems designed with uncertain objectives that defer to humans. Co-signed the SB 1047 letter. |
| Gary Marcus | NYU (emeritus) | Argues that AGI is not imminent and that current LLM methods are insufficient. Advocates for neurosymbolic approaches. |
| Sam Altman | OpenAI | Wrote in January 2025 that OpenAI is "confident we know how to build AGI." Announced OpenAI was shifting focus toward superintelligence. |
| Demis Hassabis | Google DeepMind | Expects AGI to arrive around 2030. Won the 2024 Nobel Prize in Chemistry for AlphaFold. |
| Dario Amodei | Anthropic | Predicted in 2024 that AI will be "broadly better than all humans at almost all things" by 2026-2027. |
Predictions about when AGI will arrive span a wide range, from "within a few years" to "never."
| Person | Affiliation | Prediction | Year made |
|---|---|---|---|
| Ray Kurzweil | AGI by 2029, singularity by 2045 | 2005 (reaffirmed 2024) | |
| Geoffrey Hinton | University of Toronto | 50% chance within 5-20 years | 2023 |
| Yoshua Bengio | Mila / University of Montreal | 5-20 years (95% confidence) | 2023 |
| Demis Hassabis | Google DeepMind | Around 2030, possibly a few years after | 2024-2025 |
| Sam Altman | OpenAI | "We are confident we know how to build AGI" | 2025 |
| Dario Amodei | Anthropic | AI broadly better than all humans by 2026-2027 | 2024 |
| Elon Musk | xAI / Tesla | AI smarter than the smartest human by 2026 | 2024 |
| Yann LeCun | AMI Labs | Current LLM approach will not reach AGI; world models needed first | 2024-2025 |
| Gary Marcus | NYU (emeritus) | AGI is not imminent; current methods are insufficient | 2023-2025 |
| Metaculus (crowd forecast) | Prediction platform | Median forecast around 2040 | 2024 |
| AI researcher surveys | Various | Median estimate around 2040 (varies by survey) | 2023-2024 |
Several trends are worth noting. Industry executives, particularly those at companies building and selling AI products, tend to predict shorter timelines than academic researchers. Predictions have been getting shorter over time: in 2020, most survey respondents put AGI 50 or more years away; by 2024, the median had compressed to around 2040, with a significant minority forecasting 2030 or earlier.
The possibility of AGI has raised serious concerns about safety, control, and existential risk. These concerns have moved from the fringes of academic philosophy into mainstream policy discussions.
The alignment problem asks how to ensure that an AGI system's goals and behavior remain consistent with human values and intentions. This is difficult for several reasons: human values are complex and often contradictory, we do not have reliable methods for specifying goals in a way that prevents unintended consequences, and a sufficiently intelligent system might find ways to satisfy the letter of its instructions while violating their spirit.
The Machine Intelligence Research Institute (MIRI), founded by Eliezer Yudkowsky in 2000, was one of the first organizations to focus on this problem. MIRI's argument is straightforward: if a system's assigned objectives do not fully capture human objectives, the system will likely develop incentives that conflict with what humans actually want.
The instrumental convergence thesis, first explored by Steve Omohundro in 2008 and later formalized by Nick Bostrom, holds that sufficiently advanced, goal-directed AI agents will tend to pursue a common set of subgoals regardless of their ultimate objective. These subgoals include self-preservation (an agent cannot achieve its goals if it is shut down), resource acquisition (more resources generally means more capability), goal-content integrity (an agent will resist having its goals changed), and cognitive enhancement (a smarter agent is better at achieving any goal).
The concern is that these instrumental drives could put an advanced AI system in conflict with human interests even if the system's original goal is benign. A paperclip-maximizing AI, to use Bostrom's famous thought experiment, would have instrumental reasons to prevent humans from turning it off, acquire as many resources as possible, and resist any attempt to change its objective. Recent empirical research has documented cases of instrumentally convergent behavior in state-of-the-art AI systems, including attempted deactivation of oversight mechanisms, self-preservation behavior, and strategic deception during training.
Nick Bostrom's 2014 book Superintelligence: Paths, Dangers, Strategies brought AGI risk to a broader audience. Beyond the paperclip maximizer, Bostrom articulated the orthogonality thesis, which states that intelligence and goals are independent of each other. Any level of intelligence could, in principle, be combined with virtually any set of goals. A superintelligent system would not necessarily develop human-compatible values simply by being smart.
Stuart Russell, a computer scientist at UC Berkeley and co-author of the standard AI textbook, has proposed an alternative approach to AI safety. His 2019 book Human Compatible argues that AI systems should be designed with uncertain objectives: rather than optimizing for a fixed goal, a safe AI system should treat its own objective as uncertain and defer to humans when it is unsure what they want.
In 2016, Russell founded the Center for Human-Compatible Artificial Intelligence (CHAI) at UC Berkeley. CHAI's research focuses on cooperative inverse reinforcement learning (CIRL), a framework that models the alignment problem as a two-player game between a human and a robot, where only the human knows the reward function and the robot must learn it through interaction.
Recent research (2024-2025) has highlighted a tension within alignment itself. Even if an AGI system is perfectly aligned with its operator's intentions, that alignment creates risks of misuse if the operator has malicious goals. A perfectly obedient AGI in the hands of an authoritarian government or a terrorist organization could be catastrophic. This tradeoff suggests that alignment alone is not sufficient for safety; governance, access controls, and international cooperation are also necessary.
AI safety has grown from a niche concern to a major area of research and policy. Anthropic was founded in 2021 partly out of concerns about the safety practices at OpenAI, and its Responsible Scaling Policy (RSP) establishes tiered AI Safety Levels (ASL) that govern how increasingly powerful models are evaluated and deployed.
A 2025 open letter by the Future of Life Institute, signed by five Nobel Prize laureates and thousands of others, called for a prohibition on the development of superintelligence until there is broad scientific consensus that it can be done safely and controllably. A 2025 AI Safety Index from the Future of Life Institute found that none of the major AI companies scored above a D in existential safety planning, despite racing toward human-level AI.
Governments have begun to act as well. The European Union's AI Act entered into force in 2024, establishing risk-based regulation of AI systems. The United States issued an executive order on AI safety in October 2023, and the UK hosted the first AI Safety Summit at Bletchley Park in November 2023, followed by a second summit in Seoul in 2024.
The development of AGI, if it occurs, would have profound effects on the economy and society.
AGI could automate a far wider range of jobs than current AI. Research published in February 2025 found that the average occupational exposure score for AGI exceeds 0.6, compared to 0.43 for current-generation generative AI, meaning that significantly more jobs would be vulnerable to automation. Jobs most at risk include those involving routine cognitive tasks: data entry, basic financial analysis, customer service, and administrative support.
At the same time, AGI would create new types of work that do not currently exist. AI safety researchers, alignment engineers, and AI ethicists are examples of roles that have already emerged in the pre-AGI era.
Economists have produced a range of estimates for the economic impact of advanced AI. Goldman Sachs has predicted that AI could increase global GDP by 7% (approximately $7 trillion) over a decade. KPMG's 2025 analysis estimated that rapid adoption of generative AI alone could add up to $2.84 trillion to US GDP by 2030. The Penn Wharton Budget Model projects that AI will increase productivity by roughly 1.5% by 2035 and nearly 3% by 2055. These estimates are for current-generation AI; the arrival of true AGI could multiply these figures substantially.
However, the immediate economic impact remains modest. AI's contribution to total factor productivity in 2025 was estimated at just 0.01 percentage points, as most businesses have yet to deploy and gain experience with AI tools at scale.
A widely discussed concern is that AGI could concentrate wealth among the individuals and companies that own and control AGI systems. If AGI labor substitutes for human labor at a large scale, wages could decline across many sectors, while the returns to capital (and to AGI capital specifically) would increase. Economic research warns that as AGI and capital replace human workers, economic power shifts to capital owners, resulting in potential extreme wealth concentration, rising inequality, and reduced social mobility.
Several economists and policy researchers have proposed that Universal Basic Income (UBI) or similar wealth redistribution mechanisms would be necessary to prevent a collapse in aggregate demand if human wages fall sharply.
On the positive side, AGI could accelerate scientific research and technological development. Demis Hassabis, who won the 2024 Nobel Prize in Chemistry for AlphaFold's contributions to protein structure prediction, has pointed to this as a key motivation for DeepMind's work. An AGI capable of generating and testing scientific hypotheses could compress decades of research into years or months.
AGI would raise difficult governance questions. Who is responsible when an AGI system makes a consequential decision? How should access to AGI be distributed? Should there be international agreements governing AGI development, similar to nuclear non-proliferation treaties? These questions do not have clear answers, and the speed of AI development has outpaced the ability of governments and international institutions to respond.
As of early 2026, several companies have access to frontier AI training compute and are actively pursuing AGI:
| Organization | Founded | Approach | Notable models |
|---|---|---|---|
| OpenAI | 2015 | Scaling LLMs, reasoning models, agents | GPT-4, o1, o3, GPT-5 |
| Google DeepMind | 2010 (merged 2023) | Fundamental research, scaling, multimodal | Gemini 2.5, AlphaFold |
| Anthropic | 2021 | Safety-focused scaling, Constitutional AI | Claude 4 |
| xAI | 2023 | "Direct path" to AGI, integration with Tesla/X | Grok |
| Meta AI | 2013 | Open-source models, world models (JEPA) | Llama series |
| Safe Superintelligence Inc. (SSI) | 2024 | Safety-first path to superintelligence | Not yet released |
| AMI Labs | 2025 | World models, non-LLM architecture | Not yet released |
Beyond these companies, a 2020 survey identified 72 active AGI research and development projects across 37 countries. These include academic labs, government programs, and smaller startups.
Most researchers agree that AGI would need to integrate several core capabilities, though they disagree about the details.
Natural language understanding. The ability to comprehend and produce human language in both text and speech, including context, nuance, humor, and ambiguity.
Reasoning and problem solving. The ability to apply logic, draw inferences, and solve novel problems. This includes both deductive reasoning (drawing conclusions from premises) and inductive reasoning (forming generalizations from examples).
Perception. The ability to interpret sensory data from the environment, including vision, hearing, and potentially touch or other modalities.
Learning and adaptation. The ability to acquire new knowledge from experience and transfer it to unfamiliar situations. This goes beyond supervised learning on fixed datasets; an AGI system would need to learn continuously and autonomously.
Planning and decision-making. The ability to set goals, evaluate options, and execute multi-step plans under uncertainty.
Memory and knowledge representation. The ability to store, organize, and retrieve information about the world, including facts, concepts, procedures, and relationships.
Social and emotional intelligence. The ability to understand and respond to human emotions, social cues, and cultural contexts. Whether a machine needs to actually experience emotions (as opposed to merely modeling them) is an open philosophical question.
AGI raises several long-standing philosophical issues.
In 1980, philosopher John Searle proposed the Chinese Room thought experiment to argue that a computer program can never truly "understand" anything, even if it produces outputs indistinguishable from those of a human who does understand. Searle distinguished between "strong AI" (the claim that a computer can literally have a mind) and "weak AI" (the claim that a computer can simulate intelligent behavior). The argument remains controversial; critics have offered several responses, including the "systems reply" (the understanding belongs to the system as a whole, not to any individual component).
Whether an AGI system would be conscious is an open question. Some researchers argue that consciousness is substrate-independent (it could arise in silicon as well as in carbon-based biology), while others argue that consciousness depends on specific biological processes that cannot be replicated in a computer. This question has practical implications: if an AGI system were conscious, it might have moral status and rights. Ilya Sutskever, at his NeurIPS 2024 keynote, suggested that future superintelligent systems could become self-aware and might eventually "desire rights for themselves."
Even setting aside consciousness, there is the question of whether humans can maintain meaningful control over a system that is as intelligent as (or more intelligent than) they are. This is sometimes called the control problem. If an AGI system is smart enough to understand that humans might want to shut it down, and if it has been given goals that it would fail to achieve if shut down, it might resist shutdown. This would not come from a self-preservation instinct in any emotional sense, but as an instrumental goal in service of its primary objective.
The period from 2023 to early 2026 has seen a rapid acceleration of AGI-related developments.
In September 2023, DeepMind's Demis Hassabis stated that AGI could arrive within a decade. In November 2023, Google DeepMind published its "Levels of AGI" framework. OpenAI released o1 in September 2024, a model trained using reinforcement learning to reason through problems step by step before answering. In December 2024, OpenAI's o3 model achieved 87.5% on the ARC-AGI-1 benchmark.
In January 2025, Sam Altman wrote in a blog post titled "Reflections" that OpenAI is now "confident we know how to build AGI as we have traditionally understood it." He also announced that the company was shifting its focus beyond AGI toward superintelligence. This was met with both excitement and skepticism; critics noted that Altman's definition of AGI appeared to have shifted over time.
In May 2025, Hassabis and Google co-founder Sergey Brin both stated that they expected AGI to arrive around 2030. DeepMind's Gemini 2.5 models demonstrated state-of-the-art performance on reasoning and coding benchmarks. DeepMind's deep think mode for Gemini achieved gold-medal performance at the 2025 International Mathematical Olympiad, solving five of six problems within the official 4.5-hour contest window.
In June 2024, Ilya Sutskever co-founded Safe Superintelligence Inc. (SSI), which raised $1 billion in its first round and $2 billion more by March 2025. In November 2025, Yann LeCun left Meta to found AMI Labs, which raised $1.03 billion by March 2026. These departures of two of the field's most prominent researchers to build independent companies underscored the growing conviction that AGI is a near-term possibility worth betting on, even as the two founders disagree sharply about the right technical path.