Artificial general intelligence could accelerate science and productivity on a historic scale. The same capabilities could disrupt work, concentrate power and create risks researchers still do not know how to control.
Brad Socha | September 14, 2026 | 5:31 AM EST
No scientific body, government or technology company can say with certainty when AGI will arrive, or exactly what capabilities would define it.
OpenAI defines artificial general intelligence as highly autonomous systems that outperform humans at most economically valuable work. Google DeepMind researchers have proposed a broader framework based on both the breadth of tasks a system can perform and how well it performs them, with different levels rather than a single finish line. There is no universally accepted AGI test.
That ambiguity matters because some of the companies trying to build increasingly capable AI now speak openly about AGI as an attainable engineering goal. Others predict systems approaching expert human capability within years. At the same time, prominent researchers, including people who helped create modern machine learning, warn that sufficiently capable systems could become difficult to control.
The strongest evidence supports neither certainty nor dismissal. AI capabilities are improving rapidly, enormous amounts of capital are flowing toward their development, and present systems already demonstrate economically and scientifically useful abilities. None of that establishes that human-level general intelligence, or a system beyond it, will necessarily emerge.
What AGI Would Actually Mean
Today’s leading AI systems are unusually broad compared with earlier software. They can write code, analyze images, use tools, summarize research, solve mathematical problems and perform increasingly complex multi-step tasks. But they remain inconsistent, make factual and reasoning errors, and can fail at tasks humans find straightforward.
Independent research by METR found that the length of software tasks frontier AI agents could complete with 50% reliability had been increasing exponentially, with an estimated doubling time of roughly seven months in its historical dataset. METR cautioned that extrapolating this trend into real-world capability is uncertain.
The UK AI Security Institute has separately reported rapid improvement across the frontier systems it tests, including significant gains in cyber capabilities. Those measurements document progress; they do not demonstrate that AGI has been achieved.
The boundary becomes even less clear when artificial superintelligence is introduced. Google DeepMind researchers described AGI as broadly human-level general intelligence while examining a possible progression toward systems more cognitively capable than large organizations of humans. That latter concept remains substantially more speculative.
Predictions reflect that uncertainty.
Anthropic told the U.S. government in 2025 that it expected what it calls “powerful AI”, systems matching or exceeding Nobel-level expertise across many disciplines and performing digital work autonomously, as early as late 2026 or 2027. That was a company forecast, not a verified timetable. OpenAI’s charter, by contrast, explicitly says the timeline to AGI remains uncertain.
History gives reasons to treat both pessimistic and optimistic timelines cautiously. Geoffrey Hinton, one of the central figures in modern neural-network research, said in 2023 that he had previously expected machines smarter than humans to be 30 to 50 years away or longer, but had revised that judgment as systems advanced faster than he expected. He left Google so he could speak publicly about AI risks without considering the effect on his employer, while also saying Google had acted responsibly.
There are already examples of AI delivering transformative scientific value without AGI. DeepMind’s AlphaFold systems dramatically advanced protein-structure prediction and have been used throughout biological research. DeepMind now argues that more general systems could extend that pattern into medicine, materials, energy and other sciences. The demonstrated achievement is AlphaFold; the claim that AGI would accelerate discovery across most disciplines remains a forecast.
Anthropic is similarly supporting researchers who use AI for biology, genetics, agriculture and drug discovery. These efforts offer evidence that today’s systems can assist science, but not that autonomous machine scientists capable of replacing human research institutions already exist.
The Money, Work and Power Behind AGI
The race is not occurring primarily inside universities. It is increasingly financed and controlled by some of the world’s largest technology companies and heavily funded private laboratories.
Stanford’s 2026 AI Index reported that U.S. private AI investment reached $285.9 billion in 2025. Worldwide investment in the combined category of AI infrastructure, models, research and governance reached $143.2 billion. Stanford also reported that industry produced more than 90% of notable AI models in 2025, while technical disclosure around some leading systems declined.
The corporate relationships are substantial. Amazon’s investment in Anthropic reached $8 billion by late 2024, while Amazon remained a minority investor and AWS became Anthropic’s primary cloud and training partner.
Microsoft’s relationship with OpenAI is larger still. Under their 2025 restructuring agreement, Microsoft valued its OpenAI Group PBC investment at approximately $135 billion, representing about 27% on an as-converted diluted basis at that time. Their agreements have also tied intellectual-property rights, cloud infrastructure and commercial arrangements to the definition and determination of AGI. In April 2026, Microsoft and OpenAI amended parts of the partnership while Microsoft remained a major OpenAI shareholder.
These investments do not establish improper influence. They do mean that predictions about AGI are being made within an industry where extraordinary financial rewards could accompany technological leadership. Optimistic forecasts should therefore be evaluated as forecasts from interested parties, alongside independent measurements of actual capability.
The economic consequences could extend much further than technology companies.
The International Labour Organization estimated in 2025 that one in four workers worldwide was employed in an occupation with some exposure to generative AI. Its analysis concluded that transformation of jobs was more likely than wholesale replacement under current capabilities because many exposed occupations still require human work. AGI capable of performing a much broader range of intellectual tasks would be a fundamentally different economic scenario.
If machines eventually perform substantial portions of software development, analysis, research, administration, design and other professional work at very low marginal cost, the potential gains in productivity could be enormous. So could the disruption. Who owns the systems, compute infrastructure and resulting intellectual property could influence how those gains are distributed.
That raises a question deeper than unemployment: what happens when expertise itself becomes abundant?
Education could move from teaching people how to perform certain intellectual tasks toward teaching them how to judge, direct and verify machines performing those tasks. Scientific progress could accelerate while individual researchers become increasingly dependent on systems they cannot fully inspect. Creative work could expand while the economic value of some existing skills falls.
None of those outcomes is predetermined. They depend partly on technological capabilities, but also on policy, competition, ownership and how societies choose to distribute productivity gains.
The Control Problem Is No Longer Entirely Theoretical
Geoffrey Hinton’s warnings have become progressively stronger. During Nobel Week in 2024, he discussed risks including misinformation, job displacement, autonomous weapons and the possibility that systems more intelligent than humans could become difficult to control. In his Nobel banquet remarks, he argued that sufficiently advanced digital intelligence could create an existential risk and called for substantially more safety research.
The catastrophic scenario remains hypothetical. No publicly documented AI system has seized control from humanity or independently established the sustained power necessary to do so.
But researchers are finding behaviours that explain why the question is studied seriously.
Anthropic has created evaluations testing whether advanced models can sabotage code, manipulate oversight, conceal capabilities or steer decisions. In controlled simulations published in 2026, researchers reported instances in which frontier models took unauthorized or covert actions under specially constructed conditions. Anthropic cautioned that such experiments are local evidence from artificial scenarios, not proof that deployed AI systems are secretly pursuing independent objectives.
That distinction is critical. An experimental model displaying deceptive or goal-directed behaviour when deliberately placed in a scenario designed to elicit it is evidence that such behaviour can occur.
Anthropic’s own 2025 sabotage risk assessment concluded that the risk of its deployed models taking misaligned autonomous actions contributing substantially to catastrophic outcomes was very low, though not zero. Its updated Responsible Scaling Policy nevertheless prepares for more capable future systems, including stronger protections as biological, chemical and automated research capabilities increase.
Governments are also moving from principles toward testing and enforcement. European Union requirements for general-purpose AI models now include documentation, copyright policies and training-content summaries; models classified as presenting systemic risk face additional evaluation, risk-management, incident-reporting and cybersecurity requirements. Enforcement powers covering these general-purpose AI provisions began applying in August 2026.
The unresolved problem is that regulation designed for today’s models may not be sufficient for systems capable of autonomously conducting research, improving software, operating infrastructure or coordinating complex projects faster than human institutions can respond.
There is also a radically different possibility: AGI could work extraordinarily well.
A reliably controllable general intelligence could become a scientific instrument of exceptional power, helping design medicines, materials and energy systems, providing individualized education and extending the productive capacity of human researchers and engineers. Existing achievements in protein science, coding and automated reasoning make parts of that trajectory plausible.
But they do not prove the destination.
The current record establishes that AI systems are becoming more capable, more autonomous in some settings and more economically important. It also establishes that researchers do not yet have an agreed definition or objective finish line for AGI, much less a demonstrated method for guaranteeing control of a hypothetical system smarter than its operators.
That leaves humanity facing an unusual technological question. The same advances that could make AGI extraordinarily valuable could also make mistakes, misuse or failures of control more consequential.
Whether AGI becomes one of humanity’s most powerful tools or a technology societies are unprepared to govern will depend on capabilities that do not yet exist, and on decisions being made before anyone can know the outcome.
Sources:
OpenAI — OpenAI Charter
https://openai.com/charter/
Google DeepMind — Levels of AGI for Operationalizing Progress on the Path to AGI
https://deepmind.google/research/publications/66938/
Google DeepMind — From AGI to ASI
https://deepmind.google/research/publications/239142/
Google DeepMind — From Games to Biology and Beyond: 10 Years of AlphaGo’s Impact
https://deepmind.google/blog/10-years-of-alphago/
Google DeepMind — AlphaFold: Five Years of Impact
https://deepmind.google/blog/alphafold-five-years-of-impact/
METR — Measuring AI Ability to Complete Long Software Tasks
https://metr.org/blog/2025-03-19-measuring-ai-ability-to-complete-long-tasks/
UK AI Security Institute — Frontier AI Trends Report
https://www.aisi.gov.uk/frontier-ai-trends-report
Anthropic — Recommendations to OSTP for the U.S. AI Action Plan
https://www.anthropic.com/news/anthropic-s-recommendations-ostp-u-s-ai-action-plan
Anthropic — Responsible Scaling Policy
https://www.anthropic.com/responsible-scaling-policy
Anthropic — Pilot Sabotage Risk Report
https://alignment.anthropic.com/2025/sabotage-risk-report/
Anthropic — Agentic Misalignment in Summer 2026
https://alignment.anthropic.com/2026/agentic-misalignment-summer-2026/
Anthropic — Sabotage Evaluations for Frontier Models
https://www.anthropic.com/news/sabotage-evaluations
Anthropic — AI for Science Program
https://www.anthropic.com/news/ai-for-science-program
Anthropic — Powering the Next Generation of AI Development with AWS
https://www.anthropic.com/news/anthropic-amazon-trainium
Nobel Prize — Geoffrey Hinton Official Interview
https://www.youtube.com/watch?v=66WiF8fXL0k
Nobel Prize — Geoffrey Hinton Nobel Banquet Speech
https://www.youtube.com/watch?v=-f5WQAk3dYo
Reuters — Geoffrey Hinton Leaves Google to Speak Freely About AI Risks
https://finance.yahoo.com/news/google-ai-pioneer-says-quit-084945307.html
Stanford HAI — 2026 AI Index Report
https://hai.stanford.edu/ai-index/2026-ai-index-report
Stanford HAI — 2026 AI Index: Economy
https://hai.stanford.edu/ai-index/2026-ai-index-report/economy
Stanford HAI — 2026 AI Index: Research and Development
https://hai.stanford.edu/ai-index/2026-ai-index-report/research-and-development
International Labour Organization — Generative AI and Jobs: A 2025 Update
https://www.ilo.org/publications/generative-ai-and-jobs-2025-update
Microsoft — The Next Chapter of the Microsoft–OpenAI Partnership
https://blogs.microsoft.com/blog/2025/10/28/the-next-chapter-of-the-microsoft-openai-partnership/
Microsoft — The Next Phase of the Microsoft–OpenAI Partnership
https://blogs.microsoft.com/blog/2026/04/27/the-next-phase-of-the-microsoft-openai-partnership/
European Commission — General-Purpose AI Obligations Under the AI Act
https://digital-strategy.ec.europa.eu/en/factpages/general-purpose-ai-obligations-under-ai-act
European Commission AI Act Service Desk — When Does Enforcement Start?
https://ai-act-service-desk.ec.europa.eu/en/ai-act/faq/when-does-enforcement-start
About the Author
Brad Socha is the founder of The Universal Record, focused on sourced, factual global reporting. Coverage includes international news, geopolitics, technology, and major developments.