Analysis
What Most AI Researchers Expect by 2036
Short thesis
The median expert view is not that 2036 will look like science fiction everywhere, but that AI will be a much more capable, cheaper, more agentic general-purpose technology embedded across software work, science, education, medicine, media, and government. A plausible central forecast is: AI systems will handle many multi-hour to multi-day cognitive tasks, accelerate R&D, reshape entry-level and routine knowledge work, and force much stronger governance. Researchers disagree sharply on whether this crosses into broadly human-level or superhuman AI by then; surveys put meaningful probability on it, but not certainty.
Broad researcher/expert beliefs
- Continued capability growth is the majority expectation. The largest recent survey of AI authors found aggregate forecasts giving at least a 50% chance of several substantial milestones by 2028, and a 10% chance of machines outperforming humans in every task by 2027 if science continues undisrupted; the median for that stronger “every task” threshold was 2047, not 2036.
- By the mid-2030s, many researchers expect AI agents to be useful on longer, messier workflows, especially software, analysis, design, tutoring, and operations. This does not require full AGI; it follows from cheaper inference, better tool use, larger context, multimodality, and improved reliability.
- Scientific and medical acceleration is a widely shared optimistic theme. Lab leaders such as Dario Amodei argue that powerful AI could compress years of biology, coding, and engineering work, though this is a contingent forecast rather than a settled fact.
- Risk concerns are mainstream, not fringe. In the Grace et al. survey, more than half of respondents said substantial or extreme concern is warranted for scenarios including misinformation, authoritarian control, and inequality; large minorities assigned at least 10% probability to extremely bad outcomes from advanced AI.
- Skeptical researchers do not usually deny large impacts; they doubt smooth extrapolation from current LLMs to robust reasoning, autonomy, causal understanding, and human-level generality. Their 10-year picture is more “powerful but brittle infrastructure” than “independent artificial scientists running the world.”
Evidence
1. Expert timelines: Grace et al. surveyed 2,778 researchers publishing in top AI venues. The headline result is a fast-moving but broad uncertainty band: some near-term milestones by 2028, 10% probability of human-outperforming systems across all tasks by 2027, 50% by 2047, and 10% probability of all occupations being fully automatable by 2037. This puts 2036 inside the serious-risk / serious-transformation window, but short of the median full-automation date.
2. Measured technical progress: Stanford’s 2025 AI Index reports sharp one-year gains on difficult benchmarks such as MMMU, GPQA, and SWE-bench, growing use in medicine and autonomous vehicles, falling inference costs, and wider deployment. That supports the belief that diffusion and capability improvement will both matter over the next decade.
3. Agent time-horizon evidence: METR’s 2025 work measures model capability by the length of tasks agents can complete. It reports roughly exponential improvement over six years, with a doubling time around seven months, and says extrapolation implies that within a decade agents may independently complete a large fraction of software tasks that currently take humans days or weeks. The extrapolation may break, but it is one reason many researchers expect economically meaningful agents before 2036.
4. Scaling constraints are real but not obviously binding by 2030: Epoch AI’s analysis argues that continuing rapid training-scale growth through 2030 is technically feasible under some assumptions, while flagging power, chip supply, data, and latency constraints. This supports “continued progress likely,” not “progress guaranteed.”
5. Frontier-lab expectations are aggressive: Anthropic’s Dario Amodei describes “powerful AI” as systems smarter than Nobel-level experts across many fields, with virtual-work interfaces, potentially arriving soon enough that the 5-10 years after arrival dominate social outcomes. AI 2027, a scenario project by former OpenAI/GovAI-affiliated forecasters and others, argues that superhuman AI this decade is plausible and would be more disruptive than the Industrial Revolution. These are influential forecasts, but they are not consensus surveys.
6. Benchmark skepticism remains important: François Chollet’s ARC-AGI work emphasizes that high benchmark scores can coexist with weak generalization and inefficient reasoning. This is a useful corrective to simple benchmark extrapolation: by 2036, systems may be extremely useful yet still fail at robust abstraction, causality, and out-of-distribution tasks.
7. Public-interest risk research: CAIS and related AI safety researchers emphasize loss of control, cyber/biological misuse, autonomous weapons, concentrated power, and organizational race dynamics. Even researchers optimistic about benefits increasingly treat these as governance problems that must be managed before systems become more autonomous.
Disagreements/uncertainty
- Timelines: Some experts expect transformative or superhuman AI before 2030; others expect no robust AGI by 2036. The survey median for machines beating humans at every task is after 2036, but the 10% and 25% bands are much earlier.
- Definitions: “AGI,” “human-level,” “transformative AI,” and “fully automatable occupations” are not interchangeable. Researchers can agree on major impact while disagreeing on whether the system deserves the AGI label.
- Reliability: Current systems still hallucinate, fail silently, overfit benchmarks, and struggle with long-horizon execution. The key open question is whether agent scaffolding, verification, synthetic data, and new architectures solve this enough for high-stakes autonomy.
- Economics: Capability does not equal adoption. Regulation, liability, integration costs, unions, organizational inertia, and trust may slow workplace substitution even if technical ability advances.
- Risk distribution: There is no settled expert consensus on whether the dominant harms are mundane concentration/misinformation/job disruption, catastrophic misuse, or loss of control. Different communities weight these very differently.
What could change outlook
- Slowdown factors: power shortages, chip export controls, data bottlenecks, training instability, model collapse, regulation, major safety incidents, or diminishing returns from scaling.
- Acceleration factors: algorithmic breakthroughs, successful autonomous AI R&D loops, much cheaper inference, better memory/tool-use, robotics progress, synthetic data that transfers well, or state-backed compute buildouts.
- Governance shocks: an AI-caused cyber/biological incident, election-scale manipulation, autonomous weapons crisis, or international treaty could sharply change deployment speed.
- Evaluation breakthroughs: better tests of long-horizon autonomy, deception, situational awareness, and real-world task completion would narrow today’s uncertainty more than another benchmark leaderboard.
Practical implications/watch items
- Watch agent time horizons: can systems reliably complete tasks that take humans a day, a week, then a month?
- Watch AI R&D automation: the biggest inflection would be models accelerating model research itself, not just writing routine code.
- Watch inference costs and open-weight frontier gaps: falling costs and capable open systems make diffusion faster and governance harder.
- Watch labor-market leading indicators: junior software roles, customer support, legal/document review, marketing/content, tutoring, and data analysis should show the first broad displacement or augmentation effects.
- Watch safety evaluations becoming mandatory: if frontier labs and governments converge on pre-deployment evals for cyber, bio, autonomy, and deception, that signals researchers believe the risk window is near.
- Watch physical-world bottlenecks: robotics, lab automation, manufacturing, energy, and permitting determine whether AI remains mostly cognitive infrastructure or spills rapidly into atoms.
Sources
- Grace, Stewart, Sandkühler, Thomas, Weinstein-Raun, Brauner, and Korzekwa, “Thousands of AI Authors on the Future of AI,” arXiv, 2024.
- Stanford HAI, “2025 AI Index Report,” 2025.
- METR, “Measuring AI Ability to Complete Long Tasks,” 2025.
- Epoch AI, “Can AI scaling continue through 2030?,” 2024/2025.
- Max Roser / Our World in Data, “AI timelines: What do experts in artificial intelligence expect for the future?,” 2023.
- Dario Amodei, “Machines of Loving Grace,” 2024.
- Daniel Kokotajlo, Scott Alexander, Thomas Larsen, Eli Lifland, and Romeo Dean, “AI 2027,” 2025.
- François Chollet / ARC Prize, “OpenAI o3 Breakthrough High Score on ARC-AGI-Pub,” 2024.
- Center for AI Safety, “AI Risks that Could Lead to Catastrophe,” accessed 2026.