What Should We Learn from Past AI Forecasts? (2016) (www.openphilanthropy.org)

In 2016 Luke Muehlhauser (Open Philanthropy Project) published a short study surveying historical AI forecasts to improve thinking about HLMI (human-level machine intelligence) timelines and related risks. He read histories and primary sources and found that exaggerated optimism was concentrated in a peak era (roughly 1956–1973), when many founders (e.g., Simon, Minsky, I.J. Good) made bold, multi-decade predictions — including explicit “intelligence explosion” scenarios and concrete dates (Good’s 1978 estimate, later revised). A second wave of hype in the early 1980s focused mainly on commercially useful narrow expert systems rather than HLMI. Muehlhauser also notes that the collected forecasts are not diverse — about 70% come from three overlapping groups — and that early criticism (notably Dreyfus) tempered later expert expectations. The study’s key implication for the AI/ML community is methodological: treat historical forecasts as informative about bias and provenance, not as direct evidence for current timelines. Forecasts tend to cluster by community and era, and early optimistic claims often conflated narrow benchmarks (e.g., chess) with general intelligence. Practically, forecasters should down-weight non-representative, hype-driven signals, explicitly model uncertainty and selection bias, and distinguish progress in narrow capabilities from genuine paths to HLMI when estimating timelines and risks.