What AI Researchers Broadly Expect by 2036

Analysis

What AI Researchers Broadly Expect by 2036

Short thesis

The center of expert opinion is not “AGI is definitely here in ten years” and not “AI stalls.” The most defensible consensus is: by 2036, AI systems are very likely to be much more capable, cheaper, more agentic, and widely embedded in software work, research, education, medicine, media, and bureaucracy. Many researchers expect AI to automate substantial slices of cognitive work and accelerate some scientific and engineering workflows. But expert opinion remains highly dispersed on whether systems will reach fully general human-level capability, cause explosive self-improvement, or produce macroeconomic transformation on the scale implied by the most aggressive forecasts.

What researchers and experts broadly believe

• Capability progress will probably continue rapidly. The 2025 Stanford AI Index reports sharp one-year gains on demanding benchmarks such as MMMU, GPQA, and SWE-bench, major improvement in video generation, and cases where language-model agents outperform humans in constrained programming tasks.

• The median AI-researcher forecast has moved earlier, but not all the way to “within a decade.” In the 2023 AI Impacts survey of 2,778 publishing AI researchers, the aggregate forecast put a 50% chance of high-level machine intelligence by 2047, while several concrete milestones had 50% dates before 2028. That implies many experts see a serious chance of very advanced AI by the mid-2030s, but the median does not put complete human-level machine intelligence inside exactly ten years.

• More bounded autonomy is a stronger consensus than full labor replacement. METR’s 2025 long-task study estimates frontier AI agents’ software-task “time horizon” has doubled about every seven months since 2019; extrapolated, this implies systems within five to ten years that can independently complete many software tasks now taking humans days, weeks, or possibly longer.

• Labor-market effects are expected to be real but uneven. Field evidence already shows productivity gains: Brynjolfsson, Li, and Raymond find a generative-AI assistant raised customer-support productivity by 15% on average, especially for less-experienced workers. But Daron Acemoglu argues the ten-year macroeconomic effect may be modest, estimating no more than about 0.66% TFP growth over ten years under his assumptions.

• Safety, misuse, and governance will remain central. The International Scientific Report on the Safety of Advanced AI, chaired by Yoshua Bengio, says general-purpose AI can advance public interest and science, but capabilities are advancing rapidly, experts disagree on the pace of future progress, and understanding of systems’ inner workings remains limited.

Main evidence behind that view

• Expert elicitation: the AI Impacts / Grace et al. 2023 survey is the largest direct source on AI researchers’ expectations. It reports 2,778 respondents from top AI venues; aggregate forecasts put unaided machines outperforming humans in every possible task at 10% by 2027 and 50% by 2047 in the arXiv abstract, while full automation of all occupations is much later: 10% by 2037 and 50% by 2116. This split is important: researchers distinguish “systems can beat humans at tasks” from “the economy has fully automated every occupation.”

• Capability trend data: Stanford HAI’s 2025 AI Index documents fast recent benchmark progress: +18.8 percentage points on MMMU, +48.9 on GPQA, and +67.3 on SWE-bench in one year; inference costs for GPT-3.5-level performance dropped more than 280-fold from late 2022 to late 2024. Lower cost and higher capability make broad deployment more likely even if frontier progress slows.

• Autonomy evidence: METR’s “Measuring AI Ability to Complete Long Tasks” directly studies task duration rather than benchmark answers. Current frontier models in the paper have around a 50-minute 50% time horizon on selected software tasks, with time horizon doubling about every seven months. If that trend generalizes, ten-year AI agents become operationally important even without philosophical AGI.

• Economic experiments: controlled and field studies find meaningful productivity gains in specific work. The strongest evidence is not for universal replacement; it is for task-level augmentation, quality compression between junior and senior workers, and workflow redesign.

• Risk synthesis: the international scientific report captures the expert middle: advanced AI could improve wellbeing, prosperity, and science, but also enables disinformation, fraud, cyber misuse, biased decisions, loss-of-control concerns, and governance challenges. It explicitly notes disagreement over slow, rapid, and extremely rapid future progress.

Major disagreements and uncertainty bands

• Timelines: Some forecasters and lab-adjacent analysts expect transformative or superhuman systems before 2030; AI Impacts’ median researcher forecast is later; economists such as Acemoglu expect slower aggregate productivity effects. The disagreement is not marginal—it spans “ordinary automation wave” to “industrial-revolution-scale discontinuity.”

• Scaling versus missing ingredients: Optimists believe more compute, data, synthetic data, tool use, memory, agents, and AI-assisted AI research are enough for major advances. Skeptics argue current systems still lack robust causal reasoning, grounded world models, reliability, agency, and the ability to learn from the world as efficiently as humans.

• Benchmarks versus reality: Benchmark scores have risen quickly, but researchers disagree about external validity. A system that solves many benchmark items or short software tasks may still fail on messy, long-horizon, adversarial, politically constrained, or embodied real-world tasks.

• Economic diffusion: Even if technical capability arrives, institutions, regulation, liability, integration costs, trust, unions, procurement cycles, and energy or chip constraints can slow adoption. The AI Impacts survey itself separates technical human-level capability from full automation of occupations by many decades.

• Catastrophic risk: There is no single consensus probability. In the 2023 AI Impacts survey, respondents gave non-trivial probability to extremely bad outcomes, but most expected net-positive long-run outcomes. This supports taking risk seriously without pretending researchers agree on doom.

What could change the outlook

• A sustained plateau in frontier model returns, data quality, compute supply, or energy availability would push ten-year expectations downward.

• A breakthrough in reliable agents, long-context memory, verifiable reasoning, robotics, or automated AI R&D would push expectations sharply upward.

• Regulation, export controls, liability law, model-evaluation regimes, and insurance requirements could slow deployment even if labs can build stronger systems.

• Open-weight frontier systems could accelerate diffusion and misuse; closed, expensive systems could concentrate benefits and bottlenecks.

• Major AI-caused accidents, cyber incidents, biosecurity events, or election manipulation would likely shift researcher and policy opinion toward slower deployment and stricter controls.

• Strong evidence that AI systems can autonomously perform multi-week research, engineering, or business processes would be the clearest sign that the high-end forecasts are becoming more likely.

Practical implications and watch items

• Watch agent time horizon, not just chat quality. The most important metric is whether systems can complete open-ended tasks over hours, days, and weeks with low supervision.

• Watch AI-assisted R&D. If AI begins materially accelerating model research, chip design, drug discovery, or formal verification, the outlook becomes more discontinuous.

• Watch cost curves. Falling inference cost may matter as much as raw intelligence because it determines whether AI becomes ambient infrastructure.

• Watch labor re-bundling. Expect redesign of jobs around AI supervision, verification, data access, compliance, and customer trust rather than clean one-for-one worker replacement.

• Watch safety evaluations. More capability without interpretability, reliable evaluations, and incident reporting would widen the gap between deployment and understanding.

• Watch policy convergence. The world is moving from voluntary principles toward testing, reporting, procurement rules, and liability; the strictness and international coordination of those regimes will shape the next decade.

Sources

• Grace et al., “Thousands of AI Authors on the Future of AI,” arXiv:2401.02843, 2024.

• AI Impacts Wiki, “2023 Expert Survey on Progress in AI,” 2024 update.

• Stanford HAI, “The 2025 AI Index Report,” 2025.

• METR, “Measuring AI Ability to Complete Long Tasks,” 2025.

• International Scientific Report on the Safety of Advanced AI: interim report, chaired by Yoshua Bengio, UK DSIT, 2024.

• Brynjolfsson, Li, and Raymond, “Generative AI at Work,” NBER / arXiv, 2023–2024.

• Daron Acemoglu, “The Simple Macroeconomics of AI,” NBER Working Paper 32487, 2024.

• Leopold Aschenbrenner, “Situational Awareness: The Decade Ahead,” 2024, included as an aggressive, non-consensus forecast rather than the median researcher view.