Thousands of AI Authors on the Future of AI
Katja Grace, Harlan Stewart, Julia Fabienne Sandkühler, Stephen Thomas, Ben Weinstein-Raun, Jan Brauner, Richard C. Korzekwa
cs.CY, cs.AI, cs.LG
2024-01-05
A 2023 survey of 2,778 top-venue authors puts 50% HLMI at 2047, 13 years earlier than 2022; 38–51% assigned at least 10% to extinction-level outcomes.
Governments and labs are making bets on how fast AI moves and how bad the tail looks, while front-line researchers have mostly been sampled in small surveys. AI Impacts ran related questionnaires in 2016 and 2022. In fall 2023 they sent the same family of timeline questions to authors from six venues: NeurIPS, ICML, ICLR, AAAI, IJCAI, and JMLR. 2,778 people answered, about four times the 2022 sample, and the largest AI-researcher survey they know of.
This is a snapshot of expert opinion, not a prediction market. The authors say so: these people know AI research, not forecasting.
Invites went to researchers with a peer-reviewed paper in the prior year at those six venues. The response rate was about 15%. Most people saw only a subset of items. Timing questions used two frames: half the sample filled probabilities for fixed years, half filled years for 10/50/90% probabilities. Each person's three points were fit with a gamma curve, then averaged.
Two wordings of "machines beat humans at everything" were split across respondents. High-Level Machine Intelligence (HLMI) means unaided machines can do every task better and more cheaply than human workers, feasibility not adoption, assuming science is not badly disrupted. Full Automation of Labor (FAOL) means every occupation meets that bar. The FAOL group first judged truck driver, surgeon, retail salesperson, and AI researcher, then named an occupation they expect among the last to go.
Impact items covered 11 risk scenarios, long-run value of HLMI, extinction or permanent disempowerment, whether safety work should be prioritized more, and Stuart Russell's alignment problem: important, hard, worth doing now.
Of 39 concrete tasks, all but four had better than even odds of being "feasible" within ten years, meaning a top lab could implement them in under a year if it chose to. Aggregate forecasts put at least 50% by 2028 on building a payment site from scratch, writing a song indistinguishable from a hit artist's new release, and autonomously downloading and fine-tuning a large language model. The slow tail includes writing the governing equations of a virtual world (12 years), wiring a new house (17), writing or replicating a high-quality ML paper (19 / 12), proving a top-journal theorem (22), and a Millennium Prize-level problem (27). Among 32 tasks shared with 2022, 21 moved earlier, by 1.0 year on average.
Aggregate 50% HLMI lands in 2047, 13 years before the 2022 figure of 2060. The 10% mark is 2027. FAOL 50% lands in 2116, 48 years before 2164; 10% is 2037. The two wordings differ by more than sixty years, as they did in 2016 and 2022. Under the fixed-year frame, 50% HLMI is 34 years out; under the fixed-probability frame it is 17. Respondents whose undergraduate training was in Asia expected HLMI 11 years earlier than Europe, North America, and other regions combined.
| Quantity | 2023 aggregate | 2022 |
| HLMI 50% | 2047 | 2060 |
| HLMI 10% | 2027 | 2029 |
| FAOL 50% | 2116 | 2164 |
| ≥10% extinction-level | 38%–51% of people | similar |
68.3% thought good outcomes from superhuman AI more likely than bad. Among those net optimists, 48% still put at least 5% on extremely bad outcomes; among net pessimists, 59% put at least 5% on extremely good ones. Direct extinction/disempowerment questions had medians of 5% or 10% and means around 9%–19%, depending on wording. On a 30-year horizon, more than half marked "substantial" or "extreme" concern for misinformation (86%), mass opinion manipulation (79%), dangerous groups getting better tools (73%), authoritarian population control (73%), and worse inequality (71%). About 70% said AI safety research should be prioritized more than it is. Most called alignment important and harder than other AI problems, but not more valuable to work on today. There was no agreement on whether the next five years should go faster or slower.
This is the first large top-venue snapshot after ChatGPT and GPT-4. The main signal is that timelines moved forward. It is a poor calendar. Framing can swing 50% HLMI from 17 years to 34, and the occupation wording runs sixty years later than the task wording, so the same experts do not share a stable definition of "human-level." The more durable signal is the uncertainty itself: most people leave non-trivial mass on both excellent and catastrophic outcomes, and they agree safety work should get more attention. Read it as what researchers thought in fall 2023, not as evidence that 2047 is the year.
Forecasting is hard, and these experts were not selected for forecasting skill. A 15% response rate can bias the sample; the authors looked and did not find a shift that would rewrite the headlines. Expanding from two venues to six complicates the 2022 comparison; a NeurIPS/ICML-only slice looked similar to the full 2023 pool. The HLMI vs FAOL gap is unexplained. It may be framing, or some people hearing occupation automation as adoption. The fixed-year frame is systematically later, and the paper does not know which frame is better, so the two are averaged with equal weight. Many items went to subsets, with n ranging from a few hundred to 1,714. The fieldwork is fall 2023. Near-term task medians should not be treated as current forecasts.