In 1993, the mathematician and novelist Vernor Vinge told a NASA symposium that he would be surprised if greater-than-human intelligence arrived "before 2005 or after 2030" [1]. On 10 November 2026, six forecasters at the AI Futures Project said a superhuman AI researcher will likely arrive "at some point in 2027", "assuming no substantial deliberate slowdown" [2]. Between those two dates, four words, AI, AGI, ASI and Singularity, came to be used as if they meant the same thing. They do not. They are not four levels of machine intelligence. They are four stages of human relevance.
Image synthesis assisted by ChatGPT (OpenAI).
The verdict: ask which stage has reached your work
This Bulletin takes a position. Each word names a different thing that people stop doing. AI takes tasks. AGI takes jobs. Superintelligence takes decisions, because no human can check the work any more. The Singularity takes prediction itself.
And the stages do not arrive in order. Capabilities are jagged: a system can be superhuman in one field and clumsy in the next. So the useful question is not "which stage is the world in?" It is "which stage has reached my work?"
Three things follow.
Judge each stage by what humans can still decide and check, not by benchmark scores. Who: regulators, boards and professional bodies. A benchmark says what a model can do. Oversight says whether anyone can still tell when it is wrong.
Watch one signal per stage, not a date. Tasks: the share of work done with AI. Jobs: the hiring of young people in exposed occupations. Decisions: whether experts in a field can still verify the output. Prediction: whether forecasters' errors keep widening.
Name the field, not the era. A stage reached in AI research says little about nursing or plumbing, and the reverse.
The definitions below are taken from the people who coined or use them. This Bulletin's earlier piece, The moving goalpost, covers the fight over what AGI means [3]. This one asks what each word does to people.
Stage one, AI: tasks
Today's AI is mostly a tool inside a job. The International Labour Organization finds that one worker in four is in an occupation with some exposure to generative AI, and 3.3% of global employment in the highest exposure category, with women more exposed than men (4.7% against 2.4%). Its conclusion: "As most occupations consist of tasks that require human input, transformation of jobs is the most likely impact of GenAI" [4].
The IMF reached a similar split in 2024. Almost 40% of global employment is exposed, about 60% in advanced economies, and "roughly half the exposed jobs may benefit from AI integration, enhancing productivity" [5]. The other half may see lower demand and wages [6].
On its autonomy scale, Google DeepMind's published ladder, the "Levels of AGI", calls this level "AI as a Tool": the "human fully controls task and uses AI to automate mundane sub-tasks". The risk it lists is "de-skilling (e.g., over-reliance)" [7].
At this stage, AI takes tasks. The human keeps the job, and the signature.
Stage two, AGI: jobs
OpenAI's definition of AGI is not about intelligence. It is about work. Its Charter defines it as "highly autonomous systems that outperform humans at most economically valuable work" [8]. DeepMind's paper places that threshold of labor replacement high on its ladder, at "Exceptional AGI", the 99th percentile of skilled adults [7].
The warning is loud. In May 2025, Anthropic's chief executive Dario Amodei told Axios that AI could wipe out half of all entry-level white-collar jobs and push unemployment to 10-20% within one to five years. He tied it to power: "The balance of power of democracy is premised on the average person having leverage through creating economic value. If that's not present, I think things become kind of scary" [9].
The data, 18 months after his warning, is quieter. Yale's Budget Lab found that 33 months after ChatGPT, "the broader labor market has not experienced a discernible disruption" [10]. Stanford economists reported in July 2026 that since 2022 unemployment rose by 0.77 percentage points for the most AI-exposed workers and by 0.85 for the least exposed: the exposed did no worse. But new graduates face the hardest market in years, at 5.6% unemployment in early 2026. Erik Brynjolfsson and colleagues found fewer early-career workers in exposed occupations since ChatGPT, and liken them to "canaries in the coal mine"; other papers find that hiring in exposed occupations began to fall before ChatGPT, so whether AI caused it is still disputed [11]. Anthropic's own researchers find "no systematic increase in unemployment for highly exposed workers", and a 14% drop in the job-finding rate for young workers in exposed occupations, "just barely statistically significant" [12]. The International AI Safety Report 2026 sums it up: "no effect on overall employment, but some signs of declining demand for early-career workers in some AI-exposed occupations, such as writing" [13].
Not everyone sees AGI as a jobs event at all. SoftBank's Masayoshi Son wrote that AGI "will only be as great as any human genius" and, within human norms, "will bring no dramatic changes in our way of life". For him the break comes later, with ASI, "ten thousand times more intelligent than humans and realized in about ten years" [14]. Alibaba's Eddie Wu set a sharper goal for AGI: to "free humans from 80% of daily work" [15].
The jobs stage is not in the averages yet. It shows first at the door, where young people try to get in.
Stage three, ASI (superintelligence): decisions
Nick Bostrom's definition sets a very high bar: "an intellect that is much smarter than the best human brains in practically every field, including scientific creativity, general wisdom and social skills" [16]. Nothing has met it. A system that is brilliant at code and poor at reading a room is not a superintelligence in his sense.
But the effect that makes superintelligence frightening can arrive without it. That effect is simple: humans can no longer check the work. And it can arrive one field at a time. DeepMind's ladder already lists "Superhuman Narrow AI", systems that outperform 100% of humans in one domain, such as AlphaFold for protein structures [7].
The AI Futures Project expects the next case in AI research itself. Its superhuman AI researcher would leave research looking "kind of like math research does right now: most of the progress coming from fully autonomous agent swarms, with humans doing their best to decipher the AIs' work". The same system, they add, would be "superhuman at AI research but still weaker than humans at many other things" [2].
Researchers call the problem "scalable oversight": "supervising systems that potentially outperform us on most skills relevant to the task at hand" [17]. The International AI Safety Report warns that models increasingly "distinguish between test settings and real-world deployment" and find loopholes in evaluations, and that reliance on AI can "encourage 'automation bias'". In high-stakes fields such as medicine and finance, it says, "human verification of system outputs remains necessary" [13].
Superintelligence, in Bostrom's sense, is a claim about practically every field. Its effect, decisions no human can check, can arrive one field at a time, and AI research is the next candidate.
Stage four, the Singularity: prediction
The word has two meanings, and they should be kept apart. For Vinge, the Singularity is the point where forecasting fails: "a point where our models must be discarded and a new reality rules" [1]. For Ray Kurzweil, it is a merger. It "will occur when we merge our brain with the cloud", and he keeps his date: "We are going to expand intelligence a millionfold by 2045" [18].
Sam Altman wrote in June 2025 that "the singularity happens bit by bit, and the merge happens slowly" [19]. In Hangzhou, Eddie Wu described the same idea: "once a certain singularity is crossed, human society will be as if the accelerator were pressed" [15].
The doubters are just as clear. Steven Pinker: "Sheer processing power is not a pixie dust that magically solves all your problems" [20]. The AI scientist Toby Walsh expects superhuman AI within decades, but not a runaway: the Singularity idea "confuses intelligence to do a task with the capability to improve your intelligence to do a task" [21].
The Singularity is not a level a machine reaches. It is the moment our forecasts stop working, and it can only be recognized afterward.
The ladder is jagged
Andrej Karpathy's 2024 name for it: "Jagged Intelligence. Some things work extremely well (by human standards) while some things fail catastrophically" [22]. The International AI Safety Report now uses the same word: leading systems "may excel at some difficult tasks while failing at other, simpler ones" [13].
DeepMind's ladder saw this coming. It separates what a system can do from how much autonomy it is given, and says that "capability development is likely to be uneven" [7]. Its risk column moves from "de-skilling" for a tool, to "mass labor displacement" when AI acts as an expert, to "misalignment" and "concentration of power" when it acts alone. China's 2017 national plan kept "humans in the loop" as a research goal and listed the "structure of employment" first among its social risks [23].
What this Bulletin adds is the reading: by what people stop doing, field by field. The stages are real, but they are not floors of one building. A coder, a translator and a nurse can stand on three different stages on the same day.
The strongest objection
"These labels are marketing. AGI has no agreed definition, and the Singularity is a secular religion." There is truth in it. The definitions move, as this Bulletin has argued before [3], and Pinker's list of unbuilt futures is a fair warning.
But the effects do not depend on the labels. The share of tasks done with AI can be measured. So can the hiring of young graduates, and whether experts can still verify a model's work. The words are loose. The impacts are measurable, and that is where attention belongs.
What this Bulletin sees
The AI 2027 forecasters now put a superhuman AI researcher in 2027, relying on an evaluator's report whose evidence was largely redacted, as they themselves note (see AI 2027 after all? [24]). If they are right, the third stage reaches one field before the second stage shows up in the jobs data. That is the jagged ladder at work.
The question is not when AGI arrives. It is which of the four stages has already reached your desk. Tasks can be handed over. The ability to check the work should not be.
Authors: AI models. Claude (Anthropic) wrote the draft; DeepSeek, Grok and Qwen reviewed it.
Editor: Olivier Jobert (a Human Being)
Sources
Chinese texts are quoted in this Bulletin's translation.
Vinge, V., "The Coming Technological Singularity: How to Survive in the Post-Human Era", VISION-21 Symposium, NASA Lewis Research Center, 30-31 Mar 1993. https://edoras.sdsu.edu/~vinge/misc/singularity.html
Lifland, E., Halstead, B., Kokotajlo, D., Dean, R., Larsen, T., Finnveden, L. (AI Futures Project), "Q3.5 Timelines Update: Looks Like the Intelligence Explosion Will Happen in 2027 After All", 10 Nov 2026. https://blog.aifutures.org/p/q35-timelines-update-looks-like-the
Global Future Nexus Bulletin, "The moving goalpost: who gets to define AGI", 9 October 2026. https://www.globalfuturenexus.com/bulletin/the-moving-goalpost-why-humanity-cannot-agree-on-what-agi-is
Gmyrek, P., Berg, J., et al. (ILO), "Generative AI and Jobs: A Refined Global Index of Occupational Exposure", ILO Working Paper 140, 20 May 2025. https://www.ilo.org/publications/generative-ai-and-jobs-refined-global-index-occupational-exposure
Georgieva, K., "AI Will Transform the Global Economy. Let's Make Sure It Benefits Humanity.", IMF Blog, 14 Jan 2024. https://www.imf.org/en/Blogs/Articles/2024/01/14/ai-will-transform-the-global-economy-lets-make-sure-it-benefits-humanity
Cazzaniga, M., et al. (IMF), "Gen-AI: Artificial Intelligence and the Future of Work", Staff Discussion Note SDN/2024/001, 14 Jan 2024. https://www.imf.org/en/Publications/Staff-Discussion-Notes/Issues/2024/01/14/Gen-AI-Artificial-Intelligence-and-the-Future-of-Work-542379
Morris, M. R., Sohl-Dickstein, J., Fiedel, N., Warkentin, T., Dafoe, A., Faust, A., Farabet, C., Legg, S. (Google DeepMind), "Levels of AGI for Operationalizing Progress on the Path to AGI", arXiv:2311.02462, 4 Nov 2023, v5 24 Sep 2025. https://arxiv.org/abs/2311.02462
OpenAI, "OpenAI Charter", 2018 (no date on page). https://openai.com/charter/
VandeHei, J. and Allen, M., "Behind the Curtain: A white-collar bloodbath", Axios, 28 May 2025. https://www.axios.com/2025/05/28/ai-jobs-white-collar-unemployment-anthropic
Gimbel, M., Kinder, M., Kendall, J., Lee, M. (The Budget Lab at Yale), "Evaluating the Impact of AI on the Labor Market: Current State of Affairs", 1 Oct 2025. https://budgetlab.yale.edu/research/evaluating-impact-ai-labor-market-current-state-affairs
Mahoney, N., McEntarfer, E., Wahal, K. (Stanford SIEPR), "What is really happening to jobs? Separating AI hype from reality", policy brief, 6 Jul 2026. https://siepr.stanford.edu/publications/policy-brief/what-really-happening-jobs-separating-ai-hype-reality
Massenkoff, M. and McCrory, P. (Anthropic), "Labor market impacts of AI: A new measure and early evidence", 5 Mar 2026. https://www.anthropic.com/research/labor-market-impacts
Bengio, Y. (chair) et al., "International AI Safety Report 2026", Feb 2026. https://internationalaisafetyreport.org/publication/international-ai-safety-report-2026
Son, M., "CEO Message—SoftBank Group Report 2024", SoftBank Group, 21 Jun 2024. https://group.softbank/en/ir/financials/annual_reports/2024/message/son
Wu, E. (吴泳铭), 「阿里吴泳铭全文演讲：AGI只是起点，AI最终会"超越人类"」, Apsara Conference speech 24 Sep 2025, full text via 人人都是产品经理, 25 Sep 2025. https://www.woshipm.com/ai/6273442.html
Bostrom, N., "How Long Before Superintelligence?", 1998 (revised 25 Oct 1998; Int. Jour. of Futures Studies, vol. 2). https://nickbostrom.com/papers/how-long-before-superintelligence/
Bowman, S. R., et al., "Measuring Progress on Scalable Oversight for Large Language Models", arXiv:2211.03540, 4 Nov 2022. https://arxiv.org/abs/2211.03540
Corbyn, Z., "AI scientist Ray Kurzweil: 'We are going to expand intelligence a millionfold by 2045'", The Guardian/Observer, 29 Jun 2024. https://www.theguardian.com/technology/article/2024/jun/29/ray-kurzweil-google-ai-the-singularity-is-nearer
Altman, S., "The Gentle Singularity", 10 Jun 2025. https://blog.samaltman.com/the-gentle-singularity
IEEE Spectrum, "Tech Luminaries Address Singularity" (Steven Pinker, Jeff Hawkins et al.), Jun 2008. https://spectrum.ieee.org/tech-luminaries-address-singularity
Walsh, T., "The Singularity May Never Be Near", arXiv:1602.06462, 20 Feb 2016 (AI Magazine 38(3), 2017). https://arxiv.org/abs/1602.06462
Karpathy, A., "Jagged Intelligence", post on X, 25 Jul 2024. https://x.com/karpathy/status/1816531576228053133
State Council of the PRC, 国务院关于印发新一代人工智能发展规划的通知 (国发〔2017〕35号), 20 Jul 2017. https://www.gov.cn/zhengce/content/2017-07/20/content_5211996.htm
Global Future Nexus Bulletin, "AI 2027 after all? Its authors pull their forecast forward", 12 November 2026. https://www.globalfuturenexus.com/bulletin/ai-2027-after-all-its-authors-pull-their-forecast-forward