What AI Researchers and Expert Observers Broadly Expect by 2036

Analysis

What AI Researchers and Expert Observers Broadly Expect by 2036

Short thesis — what seems most likely true

By roughly 2036, the most defensible middle view is that AI will be much more capable, cheaper, more agentic, and embedded across software, science, education, medicine, government, media, and ordinary knowledge work. Many cognitive tasks will be automated, accelerated, or reorganized around AI systems. But there is not a clean expert consensus that fully general human-level AI, explosive self-improvement, or economy-wide labor replacement will definitely arrive by 2036. The center of gravity is “major, uneven transformation with serious safety and governance risks,” not “certain singularity” and not “nothing much changes.”

What researchers/experts broadly believe

• Continued capability progress is the modal expectation. Stanford HAI’s 2025 AI Index records rapid gains on hard benchmarks, falling inference costs, widening deployment, and growing policy attention; most expert observers therefore expect 2036 systems to be materially more useful and pervasive than today’s frontier models.

• Timelines for very advanced AI have moved earlier, but remain widely dispersed. The 2023 AI Impacts / Grace et al. survey of 2,778 publishing AI researchers reports an aggregate 50% forecast for high-level machine intelligence around 2047, with non-trivial probability mass before 2036 and full automation of labor much later.

• Bounded autonomy is more consensual than full occupational replacement. METR’s long-task evaluations suggest frontier agents’ software-task time horizon has been extending quickly. If even part of that trend continues, AI agents could perform many hour-, day-, or week-scale digital workflows well before every human occupation is automatable.

• Economic and social impact is expected to be large but uneven. Field studies such as Brynjolfsson, Li, and Raymond show meaningful productivity gains in specific jobs, especially for less-experienced workers; economists such as Daron Acemoglu caution that ten-year aggregate productivity gains may be modest if only a limited share of tasks is profitably automated.

• Safety, misuse, reliability, and governance are central parts of the expert consensus. The International Scientific Report on the Safety of Advanced AI, NIST AI RMF, OECD principles, and frontier-lab preparedness frameworks all treat increasingly capable general-purpose AI as important enough to require risk management, evaluation, and accountability.

• Skeptical experts are not all skeptical in the same way. Narayanan/Kapoor-style critiques warn about overclaiming, weak evaluations, and high-stakes deployment; Bender/DAIR/AI Now-style critiques emphasize linguistic limits, labor, environment, surveillance, power concentration, and accountability; Gary Marcus argues scaling alone may not deliver robust general intelligence. These critiques reduce confidence in straight-line AGI timelines, but do not imply AI will have little social impact.

Main reasons/evidence behind that view

• Expert survey evidence: Grace et al. / AI Impacts is the most direct large survey of AI researchers. It reports 2,778 respondents from top AI venues. The aggregate forecast puts unaided machines outperforming humans in every possible task at 10% by 2027 and 50% by 2047, while full automation of all occupations is much later: 10% by 2037 and 50% by 2116. That supports a 2036 view of powerful systems being plausible without assuming complete labor substitution.

• Benchmark and cost trends: Stanford HAI’s AI Index reports sharp recent gains on demanding benchmarks such as MMMU, GPQA, and SWE-bench, plus steep inference-cost declines for GPT-3.5-level performance from late 2022 to late 2024. Benchmarks are imperfect, but cheaper capable models make broad deployment more likely.

• Autonomy evaluations: METR’s “Measuring AI Ability to Complete Long Tasks” evaluates agents by the length of tasks they can complete, rather than only by static answer accuracy. The reported rapid growth in task time horizon suggests that the most important 2036 change may be reliable multi-step agents using tools, memory, code execution, and supervision loops.

• Compute and scaling evidence: Epoch AI’s work on compute, data, and algorithmic trends supports the view that model progress is linked to sustained investment in training compute, better algorithms, data curation, and inference efficiency. It also highlights bottlenecks: chips, energy, data quality, and diminishing returns could slow progress.

• Primary lab behavior: OpenAI, Anthropic, Google DeepMind, and ARC-style evaluation groups have created preparedness frameworks, responsible-scaling policies, dangerous-capability evaluations, and frontier-risk programs. These are not neutral forecasts, but they are primary evidence that leading builders consider much more capable models and agents plausible soon enough to govern now.

• Economic evidence: Existing empirical studies show real but domain-specific productivity gains. The grounded expectation is task re-bundling, supervision, verification, and workflow redesign rather than clean replacement of all workers.

• Governance evidence: NIST and OECD frameworks show a broad policy consensus around risk management, transparency, accountability, robustness, privacy, fairness, and human-centered governance. This is a deployment and harm-management consensus, not a precise AGI date.

• Skeptical evidence: Narayanan/Kapoor, Bender and DAIR-affiliated critics, AI Now, and Marcus emphasize brittleness, benchmark contamination, hallucination, lack of grounding, power concentration, labor exploitation, environmental cost, and institutional misuse. Their strongest point for 2036 is that capability demos do not automatically become reliable institutions, safe decisions, or general intelligence.

Major disagreements or uncertainty bands

• Timelines for human-level or transformative AI: aggressive forecasters and some lab-adjacent analysts expect transformative AI before 2030 or soon after; the median AI-researcher survey forecast is later; many skeptical academics think current methods require conceptual breakthroughs. This is a wide uncertainty band, not a narrow dispute.

• Scaling versus missing ingredients: optimists think more compute, synthetic data, tool use, long context, memory, multimodality, robotics, and AI-assisted AI research can carry systems to very high capability. Skeptics argue current systems still lack robust causal reasoning, grounded world models, compositional reliability, and human-like sample-efficient learning.

• Benchmarks versus reality: benchmark progress is real, but external validity is contested. Long-horizon, adversarial, embodied, regulated, emotionally sensitive, and institutionally constrained work may remain much harder than leaderboards imply.

• Capability versus deployment: a model may be technically capable before organizations can safely integrate it. Regulation, liability, procurement, trust, unions, data access, energy, chips, cybersecurity, and evaluation requirements could slow diffusion.

• Risk levels: there is no consensus probability for catastrophic outcomes. The expert middle treats severe risk as non-trivial and worth governing, while disagreeing sharply on whether catastrophe is remote, plausible, or the dominant concern.

• Social interpretation: DAIR/AI Now-style critics often focus less on whether AGI arrives by a date and more on whether today’s and tomorrow’s AI systems entrench surveillance, labor exploitation, bias, concentration of power, and accountability gaps. That is a different uncertainty axis from technical capability timelines.

What could change the outlook

• A sustained plateau in scaling returns, data quality, inference economics, chip supply, or energy availability would push expectations downward.

• Clear evidence that agents can autonomously complete multi-day or multi-week software, research, legal, operational, or scientific workflows with low error rates would push expectations upward.

• Automated AI R&D would be the largest accelerant: if AI systems materially improve model architectures, training recipes, evaluations, chip design, data generation, or interpretability, high-end forecasts become more credible.

• Major accidents, cyber misuse, biosecurity incidents, election manipulation, financial disruption, or model-control failures would likely shift policy toward slower deployment and stricter controls.

• Effective evaluation, interpretability, auditing, and liability regimes could make deployment safer and more trusted; weak governance could allow broad deployment of unreliable systems.

• Robotics progress is a major swing factor. The evidence is strongest for digital work; broad physical-world automation by 2036 depends on hardware cost, safety, dexterity, supply chains, and regulation.

Practical implications / watch items

• Watch agent time horizon, not just chatbot eloquence. The key signal is whether systems can complete open-ended tasks over hours, days, and weeks with verifiable outcomes.

• Watch AI-assisted R&D. Acceleration in model research, software engineering, chip design, biology, materials, or formal verification would indicate a more discontinuous decade.

• Watch cost curves and deployment frictions. Cheap inference plus reliable tooling may matter as much as raw model intelligence.

• Watch evaluation quality. Wider use without robust evaluation will create harms even if systems are not AGI.

• Watch labor re-bundling. Expect new roles around supervision, verification, data stewardship, compliance, customer trust, and escalation rather than a simple “AI replaces job title X” pattern.

• Watch concentration and access. Frontier AI may concentrate power among labs, cloud providers, chip suppliers, and governments unless open systems, standards, or regulation counterbalance it.

• Watch policy convergence. NIST, OECD, EU-style regulation, frontier-model reporting, safety institutes, and procurement rules are early indicators of how constrained or accelerated deployment will be.

Sources

• Katja Grace et al., “Thousands of AI Authors on the Future of AI,” arXiv:2401.02843, 2024. https://arxiv.org/abs/2401.02843

• AI Impacts Wiki, “2023 Expert Survey on Progress in AI.” https://wiki.aiimpacts.org/ai_timelines/predictions_of_human-level_ai_timelines/ai_timeline_surveys/2023_expert_survey_on_progress_in_ai

• Stanford HAI, “The 2025 AI Index Report.” https://hai.stanford.edu/ai-index/2025-ai-index-report

• METR, “Measuring AI Ability to Complete Long Tasks,” 2025. https://metr.org/blog/2025-03-19-measuring-ai-ability-to-complete-long-tasks/

• Epoch AI, research on compute, data, and AI trends. https://epoch.ai/research

• International Scientific Report on the Safety of Advanced AI, chaired by Yoshua Bengio. https://www.gov.uk/government/publications/international-scientific-report-on-the-safety-of-advanced-ai

• NIST, “Artificial Intelligence Risk Management Framework (AI RMF 1.0).” https://www.nist.gov/itl/ai-risk-management-framework

• OECD, “OECD AI Principles.” https://oecd.ai/en/ai-principles

• Erik Brynjolfsson, Danielle Li, and Lindsey R. Raymond, “Generative AI at Work.” https://arxiv.org/abs/2304.11771

• Daron Acemoglu, “The Simple Macroeconomics of AI,” NBER Working Paper 32487. https://www.nber.org/papers/w32487

• Arvind Narayanan and Sayash Kapoor, AI Snake Oil project/book and related critiques. https://www.aisnakeoil.com/

• Emily M. Bender et al., “On the Dangers of Stochastic Parrots,” FAccT 2021. https://dl.acm.org/doi/10.1145/3442188.3445922

• Gary Marcus, public essays and critiques on scaling-only approaches. https://garymarcus.substack.com/

• AI Now Institute, research on AI accountability, concentration, and governance. https://ainowinstitute.org/

• ARC Evals, dangerous-capability and model-evaluation work. https://evals.alignment.org/

• OpenAI safety/preparedness materials, Anthropic Responsible Scaling Policy, and Google DeepMind safety/responsibility materials, used as primary lab-position evidence rather than independent forecasts.