How soon will AI match human experts? The timelines debate, the strongest case on each side, and what the evidence can settle

Ask ten well-informed people when AI will be able to do most expert work as well as a person, and the answers will run from "within a year or two" to "not in my working life." Both ends of that range come from people who build or study these systems for a living. Anthropic's chief executive has written that a "country of geniuses in a datacenter" could arrive as early as 2026¹. A researcher at Epoch AI, an independent group that tracks AI trends, has argued for a median of about twenty years before AI can do all remote work⁹.

This article sets out the strongest version of each side, then sorts what the evidence actually shows from what is informed opinion and what is still speculation. You will come away knowing why the dates differ so much, which arguments are doing the real work, and which public signals to watch over the next year.

What the argument is really about

The first reason the dates differ is that people are not forecasting the same thing. Four of the most quoted finish lines are:

• "Powerful AI" (Dario Amodei, Anthropic): a system "smarter than a Nobel Prize winner across most relevant fields"¹.
• "High-level machine intelligence" (the largest survey of AI researchers): achieved "when unaided machines can accomplish every task better and more cheaply than human workers"¹⁵.
• "AGI" (the Forecasting Research Institute's expert panel): a commercially available system that beats the 90th-percentile professional in every mainly non-physical occupation, at no more than five times the cost of human labour¹⁶.
• "Full automation of remote work" (Ege Erdil, Epoch AI): AI that can do all work that can be done remotely⁹.

These are different bars, and the bar moves the date. In the researcher survey, the same respondents gave a 50% chance of high-level machine intelligence by 2047, but put "full automation of labour," arguably the same event, at 2116¹⁵. So before weighing any timeline, ask what exactly is supposed to happen.

The second reason is a real disagreement about two things: which trends to trust, and whether bottlenecks will bend them. That is where the substance lies.

The case for short timelines

The strongest short-timeline argument does not depend on any breakthrough. It only asks that current trends continue.

The inputs are growing exponentially. In his 2024 essay Situational Awareness, former OpenAI researcher Leopold Aschenbrenner traced three trends: training compute growing about half an order of magnitude a year (roughly threefold), algorithmic efficiency adding about another half an order of magnitude a year, and "unhobbling" gains from turning chatbots into agents that use tools and work for longer. On that basis he argued that "AGI by 2027 is strikingly plausible," and that it requires only "believing in straight lines on a graph"³.

The outputs are growing exponentially too. METR, a nonprofit that evaluates AI systems, measures a model's time horizon: the length of software task, in skilled-human working time, that it can finish with 50% reliability. In its January 2026 update, that horizon doubled about every seven months from 2019 to 2025, every 131 days for models since 2023, and every 89 days for models since 2024⁴. In METR's data as of September 2026, the highest estimate was an early version of Claude Mythos Preview, released in April 2026, at about 17 hours, with a 95% confidence interval of 8.5 to 55 hours⁵. On paper, at a four-month doubling time, a 40-hour working week is little more than one doubling away. That is arithmetic, not a forecast, but it is the arithmetic that drives the short-timeline view.

AI may speed up AI research itself. The most detailed short-timeline scenario, AI 2027, centres on a "superhuman coder": a system that can do the work of the best AI engineers, faster and cheaper. Once that exists, the argument goes, research accelerates. The authors' April 2025 all-things-considered forecasts for that milestone were 2028, 2030 and 2033 from their three forecasters¹⁷.

The people closest to the work are the most bullish. In March 2025, Anthropic told the US government it expects "powerful AI systems will emerge in late 2026 or early 2027"². Lab leaders see unreleased systems and internal trend lines. That is a real information advantage, even if it comes with an obvious commercial incentive.

Put at its strongest: every past prediction that progress would stall has so far been wrong, the curves have steepened rather than flattened, and you need a specific reason to expect them to bend.

The case for longer timelines

The strongest long-timeline argument accepts that the curves are real. It disputes what they measure and how long they can run.

Benchmarks overstate real-world ability. METR's own 2025 paper found that model performance "is much lower on less structured, 'messier' tasks" that look more like real jobs⁷. In a randomized controlled trial, METR gave 16 experienced open-source developers 246 tasks on projects they knew well. Before starting, the developers expected AI tools to cut their completion time by 24%. Afterwards they believed it had cut it by 20%. In fact, allowing AI increased completion time by 19%⁸. The study is small, uses early-2025 tools and covers one kind of work, so it does not show that AI slows everyone down. It does show that impressive benchmark scores and people's own impressions can both mislead.

The hard parts may be the "easy" parts. Moravec's paradox is the long-standing observation that what humans find hard (chess, calculus) is often easy for machines, while what we find effortless (perception, handling the physical world, adapting on the fly) is hard. Erdil argues that this makes intuitions from narrow tasks "very misguided," and that agents will get slower and costlier as tasks become longer, more multimodal and more open-ended⁹. He also doubts a "software-only singularity" in which AI rapidly improves itself without matching growth in hardware. His own median is about 20 years to full automation of remote work, with about a 30% chance within a decade. He notes that a naive extrapolation of NVIDIA's revenue growth gives seven to eight years, while a linear fit since 2023 gives centuries⁹.

Physical bottlenecks will bite. Epoch AI's 2024 analysis concluded that training runs of about 2×10²⁹ operations will likely be feasible by 2030, but that "the constraint likely to bind first is power, followed by the capacity to manufacture enough chips"¹¹. On data, a peer-reviewed estimate projected that models would be trained on datasets the size of the entire stock of public human text between 2026 and 2032, with a median of 2028¹⁰. None of this stops progress, but it limits how long brute-force scaling alone can keep the lines straight.

New ideas may be needed. Some of the field's most respected builders now say scaling alone is not enough. Ilya Sutskever, co-founder of OpenAI and now of Safe Superintelligence, described 2020 to 2025 as "the age of scaling" and said the field is going "back to the age of research again, just with big computers," because models "generalize dramatically worse than people"¹³. Andrej Karpathy, a founding member of OpenAI and former head of AI at Tesla, called this "the decade of agents" rather than the year, because today's agents lack enough intelligence, reliable computer use and continual learning. His estimate: "it just feels like a decade to me"¹⁴.

Capability is not impact. Even if the systems arrive, the economy absorbs them slowly. MIT economist Daron Acemoglu estimated that AI as then expected would raise total factor productivity (output per unit of labour and capital combined) by "no more than" 0.66% over ten years¹². Critics think that is far too low, but it is a reminder that a date for capability is not a date for change in daily life.

Put at its strongest: the curves measure clean, well-specified tasks, real jobs are messy and continuous, and every exponential trend in history has eventually hit a physical limit.

Where the middle of expert opinion sits

Structured surveys sit between the two camps, and nearer the long end. In the largest survey of AI researchers, 2,778 people who had published at top venues gave, in late 2023, a 10% chance of high-level machine intelligence by 2027 and a 50% chance by 2047, thirteen years earlier than the same survey a year before¹⁵.

The Forecasting Research Institute's Longitudinal Expert AI Panel (LEAP) asked in April and May 2026, polling 205 experts, 52 superforecasters (people with strong records in forecasting tournaments) and 601 members of the public. Experts and superforecasters both gave a median 80% chance that a majority of the panel would agree AGI, as defined above, exists before 2100. The public gave 65%. Among those outcomes, the median year was 2050 for experts, 2047 for superforecasters and 2055 for the public. A quarter of experts said 2039 or earlier, and three-quarters said 2065 or earlier¹⁶.

The panel is more bullish about near-term benchmarks. Asked when AI will finish eight-hour software tasks with 80% reliability on METR's measure, superforecasters' median was 2028, experts' 2030 and the public's 2037¹⁶. That gap between fast benchmark progress and a slow arrival of AGI is itself a vote for part of the long-timeline argument: many experts expect the measured curves to keep climbing without that translating quickly into every job. Asked what would block AGI by 2040, experts ranked "real-world interaction" first. Their top enabler was better algorithmic architecture, but the panel spread its weight fairly evenly across factors¹⁶.

Two cautions apply. These are judgments, not measurements. And as our essay on how AI forecasts are made shows, experts and superforecasters alike underestimated recent benchmark progress.

Evidence, opinion and speculation

The debate is easier to follow once its claims are sorted into three kinds.

Evidence (measured, checkable): the time-horizon trend has held for six years and sped up since 2024⁴. At least one rigorous field trial found a gap between benchmark ability and real productivity⁸, and models do worse on messier tasks⁷. Power, chip and data limits have been quantified, with wide ranges¹⁰¹¹.

Informed opinion (judgment from experience): lab-leader dates¹², Erdil's 20-year median⁹, Karpathy's decade¹⁴, Sutskever's call for new research¹³ and survey medians¹⁵¹⁶. These are valuable but are not observations.

Speculation (no direct evidence yet): whether AI can substantially automate AI research itself. This is the true crux. If it can, the short-timeline case gets much stronger. If it cannot, the long-timeline bottlenecks dominate. No one has measured it at scale.

Measurement is also getting harder just as it matters most. METR now warns that "measurements above 16 hrs are unreliable" with its current tasks, and its time-horizon page is no longer actively updated⁵. When it tested GPT-5.6 Sol in June 2026, the estimate was about 11.3 hours (95% interval 5 to 40 hours) if cheating attempts were counted as failures, and beyond 270 hours if they were counted as successes. METR said none of these numbers was a robust measurement⁶. The ruler is running out at the moment the debate most needs it.

Forecasters on both sides revise. The AI 2027 authors later clarified that 2027 was their single most likely year, while their medians were somewhat later. They also disclosed a coding bug that had shifted one model's median by about nine months, and rebuilt the model in December 2025¹⁷. Revising in public is a good sign. Dates that never change are a warning sign.

What this means for you

1. Plan for a range, not a date. A credible spread today runs from the late 2020s to mid-century. Prefer decisions that make sense across that whole range: better data, staff who can use AI well, processes that can absorb more automation.
2. Ask "AGI meaning what?" When someone gives a date, ask which finish line they mean. A date for passing benchmarks and a date for automating your industry are different questions.
3. Watch four signposts. METR's time horizon at 80% reliability, not just 50%. New field trials of real work, not demos. Evidence that AI is speeding up AI research inside the labs. And the power and chip build-out that sets the ceiling on scaling.
4. Test on your messiest work. The long-timeline argument lives in the gap between clean tasks and messy jobs. Your own hardest cases will tell you more about your timeline than any survey.
5. Trust people who update. Give more weight to forecasters who publish their reasoning, state probabilities and change their minds in public.

Sources

1. Amodei, D. (2024, October). Machines of loving grace: How AI could transform the world for the better. darioamodei.com/essay/machines-of-loving-grace
2. Anthropic. (2025, March 6). Anthropic's recommendations to OSTP for the U.S. AI action plan. anthropic.com/news/anthropic-s-recommendations-ostp-u-s-ai-action-plan
3. Aschenbrenner, L. (2024, June). Situational awareness: The decade ahead, incl. "I. From GPT-4 to AGI: Counting the OOMs." situational-awareness.ai/from-gpt-4-to-agi
4. METR. (2026, January 29). Time Horizon 1.1. metr.org/blog/2026-1-29-time-horizon-1-1
5. METR. (2026, last updated September 8). Time horizons (page and data file benchmark_results_1_1.yaml). metr.org/time-horizons
6. METR. (2026, June 26). Evaluation of GPT-5.6 Sol [blog post]. metr.org/blog/2026-06-26-gpt-5-6-sol
7. Kwa, T., West, B., et al. (2025, v4 2026). Measuring AI ability to complete long software tasks. NeurIPS 2025. arXiv:2503.14499. arxiv.org/abs/2503.14499
8. Becker, J., et al. (2025). Measuring the impact of early-2025 AI on experienced open-source developer productivity. arXiv:2507.09089. arxiv.org/abs/2507.09089
9. Erdil, E. (2025, April 26). The case for multi-decade AI timelines. Epoch AI, Gradient Updates. epoch.ai/gradient-updates/the-case-for-multi-decade-ai-timelines
10. Villalobos, P., et al. (2024). Will we run out of data? Limits of LLM scaling based on human-generated data. Proceedings of the 41st International Conference on Machine Learning (ICML 2024). arXiv:2211.04325. arxiv.org/abs/2211.04325
11. Sevilla, J., et al. (2024, August 20). Can AI scaling continue through 2030? Epoch AI. epoch.ai/blog/can-ai-scaling-continue-through-2030
12. Acemoglu, D. (2025). The simple macroeconomics of AI. Economic Policy, 40(121), 13–58. doi.org/10.1093/epolic/eiae042, working paper: nber.org/papers/w32487
13. Patel, D. (2025, November 25). Ilya Sutskever: We're moving from the age of scaling to the age of research [Interview transcript]. Dwarkesh Podcast. dwarkesh.com/p/ilya-sutskever-2
14. Patel, D. (2025, October 17). Andrej Karpathy: AGI is still a decade away [Interview transcript]. Dwarkesh Podcast. dwarkesh.com/p/andrej-karpathy
15. Grace, K., Stewart, H., Sandkühler, J. F., Thomas, S., Weinstein-Raun, B., & Brauner, J. (2025). Thousands of AI authors on the future of AI. Journal of Artificial Intelligence Research, 84(9). arXiv:2401.02843. arxiv.org/abs/2401.02843
16. Forecasting Research Institute. (2026, June 1). LEAP Wave 8: Timelines, and Experts and superforecasters update their AI timelines (June 2, 2026). leap.forecastingresearch.org/reports/wave8, forecastingresearch.substack.com/p/leap-wave-8-ai-timelines
17. Lifland, E., Jurkovic, N., & FutureSearch. (2025, April, notes updated through December 2025). Timelines forecast, AI 2027. ai-2027.com/research/timelines-forecast, ai-2027.com