AI researcher Matthew Barnett posted on May 23 that a 2016 statement by Eliezer Yudkowsky — in which Yudkowsky said he would be “pretty shocked” if an AI could pass an unrestricted one-hour Turing test before the end of the world — is relevant to evaluating his track record on AI doom forecasts. The post drew over 77,000 views and 96 replies. NathanpmYoung replied that the statement offers “some evidence but not a lot,” while Jaime Sevilla noted that Yudkowsky has since acknowledged the prediction was poor, and pointed to other calls that held up, including assigning over 16% probability to an AI winning IMO gold by 2025.
Critics in the thread argued Barnett was cherry-picking a single data point rather than evaluating Yudkowsky’s full prediction record, with David Manheim calling the framing “motivated.” Others questioned whether current models could actually pass such a test given their uneven capability profiles. Barnett clarified he is not certain today’s AI would clear the bar, but said he expects the prediction will ultimately prove wrong. The exchange reflects ongoing friction in the AI safety community over how to rigorously assess forecasters whose arguments underpin high-stakes policy and research decisions.