Meta announced on November 15 a new large language model called Galactica, trained on 48 million scientific papers, textbooks, and encyclopedias. The company promoted it as a tool that “can summarize academic papers, solve math problems, generate Wiki articles, write scientific code, annotate molecules and proteins, and more.” Researchers and the public were invited to try the demo.

Within hours, scientists began sharing examples of Galactica fabricating papers—sometimes attributing them to real authors—and generating authoritative-sounding nonsense, such as a wiki article on “the benefits of eating crushed glass” or a history of bears in space. The model also produced racist and biased content when prompted with certain topics, and refused to generate text on “racism” or “AIDS,” returning a content filter message.

Chief AI scientist Yann LeCun defended the model, tweeting “Type a text and Galactica will generate a paper with relevant references, formulas, and everything.” The episode mirrors Microsoft’s 2016 Tay chatbot debacle, where a public demo was shut down after users turned it into a racist bot. It highlights the gap between the promise of large language models and their current inability to reliably separate fact from fiction.