Analysis
What AI Researchers Mostly Expect by 2036
Short thesis
The center of gravity among AI researchers is not “nothing changes” and not a guaranteed near-term intelligence explosion. The mainstream 10-year expectation is that AI becomes a pervasive, agentic general-purpose technology: much better at coding, scientific assistance, office work, tutoring, media generation, and routine digital operations; widely embedded in institutions; economically disruptive but unevenly diffused; and risky enough that safety, evaluation, and governance become central research and policy work. A meaningful minority expects human-level-or-better AI on most tasks within the decade, while another serious skeptical camp expects slower diffusion, persistent reliability bottlenecks, and less abrupt labor-market transformation.
Broad researcher/expert beliefs
- Frontier systems will keep improving in reasoning, multimodal understanding, tool use, coding, long-horizon task completion, and autonomy, though progress will be lumpy and bottlenecked by data, compute, evaluation, reliability, and regulation.
- By the mid-2030s, AI will likely perform or assist with a large share of knowledge-work tasks, especially software, writing, analysis, customer support, design, education, and parts of biomedical and scientific work.
- Researchers are split on whether “human-level machine intelligence” arrives by then. The largest expert survey gives a median date around 2047 for machines outperforming humans in every task, but also gives roughly a 10% chance by 2027; that implies a non-trivial but not majority belief in transformative systems before 2036.
- Most experts expect substantial social upside and substantial risk at the same time. In the 2023 AI Impacts/Grace et al. survey, 68.3% said good outcomes from superhuman AI are more likely than bad, yet many still assigned non-negligible probability to catastrophic outcomes.
- There is broad agreement that AI risk-reduction research should receive more priority, especially for misinformation, cyber/offensive use, concentration of power, inequality, loss of control, and evaluation of dangerous capabilities.
- Skeptical researchers do not usually argue that AI will be unimportant. Their view is that AI should be treated more like a powerful normal technology: transformative over decades through application-building and diffusion, not necessarily an abrupt AGI discontinuity.
Main reasons/evidence behind that view
- Expert surveys point to fast but uncertain capability timelines. “Thousands of AI Authors on the Future of AI” surveyed 2,778 researchers who had published in top AI venues. The aggregate forecast gave at least 50% probability to several concrete milestones by 2028, including autonomous construction of a payment-processing site, creating a song indistinguishable from a popular musician’s new song, and autonomously downloading and fine-tuning a large language model. The same survey estimated a 50% chance of unaided machines outperforming humans in every possible task by 2047, with a 10% chance by 2027, and full automation of all occupations much later: 10% by 2037 and 50% by 2116.
- Capability benchmarks are moving quickly. Stanford HAI’s 2025 AI Index reports sharp one-year gains on demanding benchmarks introduced in 2023: +18.8 points on MMMU, +48.9 on GPQA, and +67.3 on SWE-bench, plus major progress in video generation and some programming-agent settings.
- Autonomy is becoming measurable. METR’s 2025 “time horizon” work estimates that frontier AI models’ 50%-success task-completion horizon was about 50 minutes for Claude 3.7 Sonnet-style systems and had doubled roughly every seven months since 2019. METR cautions about external validity, but if the trend generalized, it would support expectations of agents completing day- or week-scale software tasks within a decade.
- Real-world adoption is already concentrated where researchers expect first-order impact. Anthropic’s Economic Index found Claude usage concentrated in software development and technical writing; about 36% of occupations saw AI use in at least a quarter of their associated tasks, about 4% across three-quarters of tasks, and use leaned somewhat more toward augmentation than automation, 57% versus 43%.
- Deployment evidence is no longer limited to chatbots. Stanford’s AI Index notes AI moving into daily life: FDA approvals of AI-enabled medical devices rose from six in 2015 to 223 in 2023, and robotaxi services such as Waymo and Baidu Apollo Go are operating at scale.
- Risk evidence is also growing. Stanford reports rising AI incidents and a gap between major developers recognizing responsible-AI risks and consistently performing standardized evaluations. The AI Impacts survey found that 38% to 51% of respondents gave at least a 10% chance of advanced AI leading to outcomes as bad as human extinction, depending on framing.
Major disagreements or uncertainty bands
- Timeline disagreement is large. “Median 2047” and “10% by 2027” can coexist because researchers’ probability distributions are wide and skewed. The useful interpretation is not that 2047 is certain, but that many researchers see a plausible decade-scale path while most do not treat it as the single most likely outcome.
- Capability does not equal deployment. Even if models can perform many tasks in demos, institutions may be slowed by verification costs, liability, data governance, security, workflows, union and professional norms, regulation, and customer trust.
- Benchmarks may overstate or understate real capability. Some benchmark gains reflect better test-taking, tool scaffolding, or contamination risks; conversely, benchmarks may miss economically useful capabilities such as fast drafting, code migration, decision support, and interface automation.
- Skeptical “normal technology” views argue that sudden economy-wide impact is implausible because application development and diffusion historically take decades for general-purpose technologies. On this view, AI can be transformative without requiring a near-term superintelligence story.
- Safety disagreement remains deep: researchers disagree on whether the largest risks are near-term misuse and concentration, long-term loss of control, labor displacement, or hype-driven misallocation of attention.
What could change the outlook
- Faster outlook: major algorithmic breakthroughs, sustained compute scaling, cheaper inference, better agent scaffolds, reliable long-context memory, self-correction, strong tool-use standards, automated AI research, and faster enterprise integration.
- Slower outlook: compute or energy bottlenecks, weaker returns to scaling, data limitations, model reliability plateaus, security failures, expensive regulation, public backlash, copyright/data restrictions, or lack of profitable deployments outside a few domains.
- Riskier outlook: strong autonomous cyber capabilities, persuasive misinformation at scale, biological design assistance, weak evaluation regimes, model power concentrated in a few actors, or competitive pressure that erodes safety margins.
- Safer outlook: robust third-party evaluations, incident reporting, liability clarity, audit trails, secure deployment practices, international coordination, and measurable progress on interpretability, control, and misuse prevention.
Practical implications / watch items
- Watch autonomous-task duration, not just benchmark scores: can agents reliably complete multi-hour, multi-day, and eventually multi-week tasks with low supervision?
- Watch coding and science workflows first; they are the clearest early indicators for broader knowledge-work automation.
- Watch diffusion metrics: enterprise deployment, regulated-sector adoption, cost per useful task, and whether AI use remains augmentation-heavy or shifts toward end-to-end automation.
- Watch labor-market data by task, not just occupation. The likely 10-year effect is task recomposition before full occupation replacement.
- Watch safety-evaluation maturity: dangerous-capability evals, third-party audits, incident databases, and whether labs publish comparable safety cases.
- Watch compute, chips, energy, and geopolitics: model progress depends on physical and political infrastructure as much as algorithms.
- The practical default should be: prepare for highly capable AI assistants and semi-autonomous agents becoming normal in most digital work within 10 years, while avoiding plans that require either guaranteed AGI or guaranteed stagnation.
Sources
- Grace et al., “Thousands of AI Authors on the Future of AI,” arXiv:2401.02843, 2024.
- Stanford HAI, “2025 AI Index Report,” especially benchmark gains, deployment, incidents, and governance findings.
- METR, “Measuring AI Ability to Complete Long Tasks,” 2025; arXiv:2503.14499.
- Anthropic, “The Anthropic Economic Index,” 2025, based on anonymized Claude.ai conversations.
- Kapoor and Narayanan / AI as Normal Technology, “AI as Normal Technology,” skeptical diffusion-focused view.
- AI Impacts survey materials and discussion around expert forecasts, risk concern, and prioritization of risk-reduction research.