On November 15, 2022, Meta released Galactica, described in its paper as “a large language model for science” trained on scientific papers, reference material, and knowledge bases. The paper claimed the model could “store, combine and reason about scientific knowledge,” reported it beat GPT-3 on technical tasks such as writing LaTeX equations, and announced the team would “open source the model for the benefit of the scientific community.” A public demo let anyone type a prompt and get a generated paper-like response.
The problem surfaced almost immediately. Because Galactica produced fluent, confident, citation-laden text regardless of whether the underlying claims were true, users found it would generate plausible-looking but fabricated or biased scientific content. Meta’s Galactica account (Papers with Code) posted that it had “paused the demo for now,” writing: “Thank you everyone for trying the Galactica model demo. We appreciate the feedback we have received so far from the community, and have paused the demo for now. Our models are available for researchers who want to learn more about the work and reproduce results in the paper.” That statement was dated November 17, 2022 - three days after launch.
Meta’s chief AI scientist Yann LeCun, who had promoted the demo, responded to the takedown on the same day with a terse post: “Galactica demo is off line for now. It’s no longer possible to have some fun by casually misusing it. Happy?” His framing as misuse, versus critics’ framing as a fundamental reliability problem, became part of the story.
The one-line lesson: fluency is not accuracy, and a model that wraps wrong answers in the format of authoritative science is more dangerous than one that obviously fumbles.