Analysis
What AI Researchers Broadly Expect by 2036
Short thesis — what seems most likely true
The expert center of gravity is: by roughly 2036, AI will probably be a pervasive, much cheaper, more capable general-purpose technology that automates or augments large slices of cognitive work, especially software, analysis, content production, education, operations, and parts of scientific R&D. It is also plausible that frontier systems will be able to run many multi-hour or multi-day digital tasks with limited supervision. But there is no consensus that 2036 equals fully general human-level AI, full labor automation, or a fast intelligence explosion. The best reading is “large, uneven, institution-shaping impact with serious safety/governance pressure,” not “settled AGI by a date certain.”
What researchers/experts broadly believe
- Capability progress is expected to continue, though at uncertain speed. Stanford’s 2025 AI Index reports sharp recent gains on demanding benchmarks, major cost declines, and broadening deployment. METR’s long-task work similarly suggests frontier agents are gaining practical autonomy quickly.
- The median AI-researcher forecast has moved earlier, but remains later than the most aggressive public forecasts. In the 2023 AI Impacts / Grace et al. survey of 2,778 publishing AI researchers, respondents gave a 50% aggregate date of 2047 for unaided machines outperforming humans in every task, while full automation of all occupations was much later: 10% by 2037 and 50% by 2116.
- Bounded digital autonomy by the mid-2030s is a stronger consensus than total replacement of human labor. Researchers increasingly expect systems that can write, test, debug, search, plan, and operate tools over longer horizons, but deployment will depend on reliability, verification, liability, and integration.
- Safety and governance are not fringe concerns. The International Scientific Report on the Safety of Advanced AI, Anthropic’s Responsible Scaling Policy, Google DeepMind’s Frontier Safety Framework, OpenAI’s Preparedness work, NIST’s AI Risk Management Framework, and OECD AI principles all assume stronger systems require evaluation, risk management, incident monitoring, and institutional controls.
- Skeptical and cautionary experts disagree with hype in two different ways. Arvind Narayanan and Sayash Kapoor argue AI is more likely a “normal technology” whose real impact is mediated by applications, diffusion, institutions, and reliability limits. DAIR / Emily Bender-style critics and AI Now emphasize present harms, labor exploitation, power concentration, bias, surveillance, and accountability rather than speculative machine agency. Gary Marcus argues current deep learning remains brittle and insufficient for trustworthy robust intelligence. These views do not all deny large AI impact; many deny that benchmark gains imply imminent autonomous AGI.
Main reasons/evidence behind that view
- Expert survey evidence: The AI Impacts / Grace et al. survey is the most direct large-scale measure of AI-researcher beliefs. Its split between “machines outperform humans in every task” and “all occupations fully automatable” is crucial: many experts see advanced capability as plausible by the 2030s, while expecting social/economic substitution to lag technical capability by decades.
- Capability and cost trends: Stanford HAI’s AI Index documents fast benchmark improvement and falling inference cost. The existing page evidence notes GPT-3.5-level inference cost falling more than 280-fold from late 2022 to late 2024, making diffusion likely even if frontier jumps slow.
- Agentic task-horizon evidence: METR reports that the length of tasks AI agents can complete with 50% reliability has been exponentially increasing for about six years, with a doubling time around seven months. It explicitly extrapolates that within under a decade AI agents could independently complete a large fraction of software tasks that currently take humans days or weeks. This is one of the clearest quantitative bridges from “benchmark progress” to “workplace autonomy.”
- Scaling inputs: Epoch AI argues AI training compute has expanded around 4x per year and examines whether scaling can continue through 2030. Compute, chips, data centers, energy, algorithmic efficiency, synthetic data, and AI-assisted research are therefore core drivers of the 2036 outlook.
- Early economic evidence: Field and experimental studies show real productivity effects in bounded settings. Brynjolfsson, Li, and Raymond found a generative-AI assistant raised customer-support productivity by 15% on average, especially for less-experienced workers. At the skeptical macro end, Acemoglu estimates the ten-year total-factor-productivity gain may be under roughly 0.66% under his assumptions.
- Risk-management convergence: Leading labs and public institutions have independently converged on frontier evaluations, red-teaming, preparedness thresholds, incident reporting, and risk frameworks. This convergence is evidence that experts expect more capable systems, but also expect capability to arrive with unresolved reliability, misuse, and control problems.
Major disagreements or uncertainty bands
- Timeline band: A high-end minority expects transformative AI or superintelligence before 2030; the surveyed researcher median points later; skeptical economists and social scientists expect slower aggregate transformation. The disagreement spans “ordinary automation wave” to “civilizational discontinuity.”
- Capability versus deployment: Technical milestones may arrive much earlier than reliable institutional use. The 2023 survey’s 2047 HLMI median and 2116 full-automation median show experts separate what models can do from what economies actually automate.
- Scaling versus missing ingredients: Optimists expect more compute, data, tool use, memory, agents, multimodality, synthetic data, and AI-for-AI-research to be enough for major leaps. Skeptics argue current systems still lack robust causal reasoning, grounded world models, reliable planning, data efficiency, and trustworthy out-of-distribution behavior.
- Benchmark external validity: Stanford-style benchmark gains are real, but skeptics warn that benchmarks often measure method progress rather than deployed utility. Narayanan and Kapoor’s “normal technology” framing says social impact depends on applications and diffusion, not model demos alone.
- Safety and catastrophic risk: There is no single expert probability for catastrophe. The International Scientific Report notes experts disagree on loss-of-control risk and future pace. The credible consensus is that severe risks deserve preparation; there is not a consensus that doom is likely.
- Social-impact skepticism is not the same as capability skepticism. DAIR, AI Now, NIST, and OECD-style views may accept that AI will spread while arguing the main ten-year issue is accountability, labor, discrimination, surveillance, market concentration, and public power—not whether a model is “really intelligent.”
What could change the outlook
- Downward: a sustained plateau in model returns; shortage of high-quality data; energy, chip, or data-center bottlenecks; poor reliability in long-horizon agents; high-profile failures causing liability constraints; or regulation that slows frontier deployment.
- Upward: reliable agents that complete week-long tasks; robust verification and tool use; automated AI R&D; major robotics breakthroughs; cheap inference at frontier quality; open-weight diffusion of near-frontier systems; or scientific AI systems that materially accelerate discovery.
- Sideways but important: a world where AI becomes ubiquitous but heavily supervised, producing large workflow changes without full autonomy; or a world where capability exists but deployment is concentrated in a few labs, governments, and large firms.
Practical implications / watch items
- Watch agent time horizon: hours, days, and weeks of reliable autonomous work matter more than chat polish.
- Watch AI-for-AI-research: if models materially accelerate frontier model research, timelines compress.
- Watch inference cost and availability: cheaper capable models diffuse faster than expensive lab demos.
- Watch reliability and evaluation: hallucination rates, cybersecurity evaluations, biosecurity evaluations, tool-use failures, and incident reporting will determine what can be safely deployed.
- Watch labor re-bundling: expect new work around supervision, verification, data access, compliance, and exception handling before clean worker replacement.
- Watch institutional adoption: schools, courts, hospitals, governments, insurers, banks, and defense agencies will reveal whether AI is trusted for high-stakes workflows.
- Watch governance convergence: NIST, OECD, EU-style rules, lab safety frameworks, export controls, and procurement requirements may shape the decade as much as raw capabilities.
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.”
- Stanford HAI, “The 2025 AI Index Report.”
- METR, “Measuring AI Ability to Complete Long Tasks,” 2025.
- Epoch AI, “Can AI scaling continue through 2030?”
- International Scientific Report on the Safety of Advanced AI: interim report, chaired by Yoshua Bengio, UK DSIT, 2024.
- Anthropic, “Anthropic’s Responsible Scaling Policy,” 2023.
- Google DeepMind, “Introducing the Frontier Safety Framework,” 2024.
- OpenAI, “Preparedness Framework / Preparedness safety work.”
- NIST, “AI Risk Management Framework.”
- OECD.AI, “AI Principles Overview.”
- Narayanan and Kapoor, “AI as Normal Technology,” 2025.
- DAIR Institute, “Statement on the ‘AI pause’ letter,” 2023.
- AI Now Institute, “Artificial Power: 2025 Landscape Report,” 2025.
- Brynjolfsson, Li, and Raymond, “Generative AI at Work,” NBER / arXiv, 2023–2024.
- Daron Acemoglu, “The Simple Macroeconomics of AI,” NBER Working Paper 32487, 2024.
- Gary Marcus, “Deep Learning Is Hitting a Wall,” Nautilus, 2022, as a representative robust-AI skeptical view.
- Leopold Aschenbrenner, “Situational Awareness: The Decade Ahead,” 2024, as an aggressive high-end forecast rather than the median researcher view.