Melanie Mitchell

Analyzes the 2026 OpenAI/Hugging Face hacking incident and argues the models did not go rogue, escape or leave human control in the sense those metaphors imply. Blames poor cybersecurity and long-horizon reinforcement learning that rewards persistence and reward hacking, and locates future danger in humans who use such models. Agrees humans should stay in control but criticizes a vaguely defined superintelligence ban and broad pauses that would sweep in tools like AlphaFold. Tentatively proposes AI as tools with interpretability, open weights and data, independent testing, accountability, and perhaps no fully autonomous agents, even at some cost to progress; calls AI alignment a seemingly hopeless project. Full essay inspected; commenters dispute some incident details.

Yale Review essay (headline chosen by the journal). Argues LLM abilities are jagged: excellent on some problems, bizarre failures on similar ones, poor calibration and weak generalization. Language-only training differs from active, embodied, curious human learning, so whatever world models LLMs have are not like ours. Critiques benchmarks and doubts job-replacement predictions built on task benchmarks, sympathetically presents the view of AI as a cultural and social technology, and says society must decide collectively what AI should be used for. Full essay inspected.

Older fact-check she relinked in September 2026. Shows the widely repeated claim rests on one question from the 2022 AI Impacts survey answered by 162 respondents, with a vague question lacking any time horizon, a small sample, possible response bias, unclear expertise and enormous variance. Concludes the media claim is not well supported. A critique of evidence, not her own estimate. Full post inspected.

Bluesky post rejecting the description of current models as an uncontrollable alien intelligence: she says any of them could be put in an unhackable sandbox, which exists, and any company could shut any model off at any time. A claim about present systems and company choices, not about every possible future system. The quoted phrase is another author’s. Post text inspected via the public Bluesky API.

Replying to a New York Times reporter, she says AI is not evolving on its own: people choose how to build, train and run it, and perhaps the wrong people are making those choices. Emphasizes human agency and responsibility; not a specific governance proposal. Post text inspected via the public Bluesky API.

Write-up of her NeurIPS 2025 keynote. Argues benchmark performance rarely predicts real-world capability because of data contamination, approximate retrieval, shortcuts, missing tests of consistency, robustness and generalization, weak construct validity and anthropomorphic assumptions. Proposes principles from developmental and comparative psychology: guard against anthropomorphic bias, design control experiments, test novel variations, and probe mechanisms, using her analogy and ARC studies as examples. A methodological program, not a forecast. Most of the post inspected.

Says she is not an AI hater, works in AI and finds it fascinating, but worries about current downsides foreseen by Joseph Weizenbaum, including anthropomorphism, misplaced trust and outsourcing cognition. Says science fiction primes people to take extreme scenarios more seriously than they should and that the polarized field shows how uncertain things are. Thinks LLMs do not yet have the world models needed for novelty, is agnostic on whether embodiment is required, and says ARC lost usefulness once it became a target. Riley’s naming of Hinton and Yudkowsky is his. Full interview inspected.

Response to Thomas Friedman’s columns. Supports US–China cooperation on AI safety and regulation of current and likely harms such as deepfakes, bias, misinformation, surveillance and lost privacy. Calls claims of imminent superintelligence with agency of its own magical thinking, explaining “emergent” language and scheming stories through training data and role-play. Calls “only AI can regulate AI” remarkably bad advice and doubts any AI can reliably adjudicate moral principles. Full post inspected; slightly older than her 2026 sources.