Three stories landed this week that, read together, constitute something close to a crisis memo for the AI industry. None of them are about safety in the sci-fi sense. All of them are about a more boring, more devastating problem: AI systems that look like they're thinking but aren't.

The Scientific Reasoning Gap

A 2026 paper by Ríos-García et al. in arXiv CS.AI found that LLM-based systems increasingly deployed to conduct autonomous scientific research produce results without reasoning scientifically. The distinction matters. A system can output a correct answer through pattern matching while entirely missing the inferential chain that makes the answer meaningful or reproducible. Science isn't a lookup table. Neither is understanding distributions of model outputs, which a separate paper by Reif, Yang, et al. argues we catastrophically underestimate by evaluating only single LLM responses rather than the full probabilistic spread. Every benchmark is a lie of averages.

Automation Backlash and the Token Economy

This is the undercurrent beneath Nilay Patel's Decoder argument that people do not actually want automation: the backlash isn't Luddism, it's epistemic. People sense, correctly, that AI outputs lack the connective tissue of actual reasoning. And The Verge's report on AI monetization pressure adds a grim coda: as OpenAI and Anthropic squeeze users for token revenue, the pressure to deploy unreasoning systems at scale intensifies. Meanwhile, The New Yorker's piece on AI in education notes cognitive scientists warning that offloading reasoning to non-reasoning systems may be permanently impairing the very skill AI claims to augment. The feedback loop here is not virtuous.