Why Expert Predictions on AI Keep Missing the Mark

The article opens with a viral paper titled "A Taxonomy of Omnicidal Futures Involving Artificial Intelligence," co-authored by Jacob Tsimerman, who is the 2026 Fields Medal recipient. Though published on arXiv (a technical preprint site), the author notes it reads more like a LessWrong post and questions how it passed moderation, arguing that name recognition allows certain figures to operate under different standards. The author then challenges whether a Fields Medal adds credibility to AI risk arguments, pivoting to the broader claim that credentials and fame are poor predictors of forecasting ability.

To support this thesis, the author draws on historical examples. Paul R. Ehrlich, a former Stanford professor, gained worldwide recognition for his 1968 book predicting population collapse—a prediction that proved wrong, yet the idea of a "population crisis" remained embedded in public and intellectual consciousness for decades. Decorated military experts in the early 2000s confidently posited a link between nonexistent weapons of mass destruction in Iraq and al-Qaeda, yet the Iraq War continued for another 20 years despite intelligence failures from the outset. Well-credentialed finance and economics experts were blindsided by the 2008 financial crisis. During COVID, health experts championed lockdowns and other measures that proved ineffective at stopping virus spread while causing societal harm.

The author observes that "AI risk" has followed a similar cultural trajectory, becoming a constant refrain in media for three years. Self-described AI "safetists" frame themselves as outsiders fighting a pro-AI hegemony, yet their position is actually mainstream—reflected in publications such as Time, The New York Times, and The Atlantic. USA Today published the headline "The AI reckoning is here, and America is completely unprepared" just two days before the article was written. By contrast, unhesitatingly optimistic articles about AI are much harder to find than skeptical or hostile ones, which blame AI for academic dishonesty, declining literacy, joblessness, and other ills.

The author then examines LessWrong, founded in February 2009, which has accumulated 45,933 posts and 906,338 comments from contributors including many leading and would-be leading figures in AI and technology. Adding the millions of hours of podcasts and articles produced by mainstream outlets over 15 years yields an enormous body of AI commentary. Yet the author claims that no one in that entire corpus has a consistently good track record on AI prediction. Specifically: between 2016 and 2020, no one predicted the rise of LLMs as they actually emerged. No one predicted that Nvidia stock would surge due to worldwide GPU demand for language models. After Nvidia had already surged in 2023–2024, no one in 2025 was saying there would be a global memory shortage, despite it now having its own Wikipedia page and driving huge rallies in stocks such as Micron and SanDisk.

The author recalls that between 2015 and 2020, the focus was on machine learning broadly. AlphaGo and TensorFlow were major developments, and professors took sabbaticals to consult on recommendation algorithms and self-driving cars. By 2020–2021, the biggest stated threat was deepfakes, not self-improving agentic intelligence or mass job loss. When LLMs did become prominent in 2022–2023, the main concern was AI devaluing artists or infringing copyrights, as it excelled at image and text generation. Math and coding were considered too advanced for automation. The Guardian published the headline "'It's the opposite of art': why illustrators are furious about AI" in January 2023, which seemed hugely significant at the time but "seems quaint today now that AI is disproving century-old conjectures." The author also references experts in the 1950s who underestimated computer capabilities before the transistor revolution, and those who overestimated in the opposite direction by predicting moon colonies. The author concludes that whether the focus is AI crisis today or overpopulation in the 1960s, this pattern underscores the general difficulty of forecasting and suggests we cannot expect experts to see the future first.