METR reports forecasts on AI R&D acceleration and societal impacts

METR reports forecasts on AI R&D acceleration and societal impacts

METR’s report, published August 20, 2025, presents a pilot survey by METR and the Forecasting Research Institute of AI forecasting domain experts and superforecasters. They estimated the chances of rapid AI progress following hypothetical AI research and development capabilities, and of extreme societal effects if progress accelerated. METR defines “effective compute” as a way to combine progress from larger compute spend and algorithmic improvements: the compute scale-up that would have been needed to match a more capable model without those innovations. The forecasters generally considered three-fold acceleration plausible but unlikely, while disagreeing sharply about possible extreme societal effects. METR says month-long, open-ended research evaluations could help resolve disagreement about acceleration, but the groups remained far apart on societal impacts even when forecasts assumed acceleration. METR concludes that this disagreement merits further study before expanding the pilot. It cautions that the small sample and subjective forecasts make the results suggestive, and that some aggregate statistics are fragile. The report does not establish whether the hypothetical capabilities or societal effects will occur.

Publication: 760eec9d-12e9-0eb2-3164-106be1f87901

attribution — pass: The reading attributes the pilot survey, the effective-compute definition and the conclusions to METR, working with the Forecasting Research Institute. This matches the single METR blog post in the evidence. Forecasts are kept as participants' estimates rather than presented as facts.

we have worked with the Forecasting Research Institute to conduct an initial pilot survey

summary — uncertain: Two central terms are left unexplained, and the definitions list is empty. First, 'superforecasters' is never defined; the source footnote calls them top performers from forecasting tournaments or platforms. Second, the summary says forecasters found 'three-fold acceleration' plausible but unlikely without saying what is accelerating. In the source, it means effective compute scale-up, with a year of 2018–2024 progress happening every 4 months. Smallest repair: add a short gloss for superforecasters and name the three-fold acceleration of effective compute scale-up in plain words. The summary also omits the medians (20% vs 8%), though the claims include them.

the amount of progress that happened in one year between 2018 and 2024 now happens every 4 months

date — pass: evidence.originalPublished is the instant 2025-08-20T07:00Z, whose UTC calendar day is August 20, 2025; this matches the URL path and the reading's stated date. The reading presents forecasts as forecasts, not outcomes, and uses no hindsight framing. The retrieval date appears only as a capture note.

support — pass: Each claim matches its quoted source text: the survey design, the effective-compute definition, the 20%/8% medians, the below-0.5% vs over-18% split, and the pilot caveats. The summary says month-long research evaluations could resolve the acceleration disagreement but not the societal gap; this is supported by the conclusion and by the conditional results. No benefits or costs are invented. The closing question follows from the source's own call to explore the disagreement.

Conditioning on a three-fold (or even ten-fold) acceleration does not close this gap in the predictions of the two groups

media — pass: No media is selected, and the evidence supplies no verified media candidates. No link has been invented.