Meta removed its Galactica AI, a large language model designed to generate academic papers, three days after releasing it on November 15. The decision followed sharp criticism from artificial intelligence experts and social media users who highlighted the system's tendency to produce fabricated content that appeared credible.
The tool, trained on scientific papers, was taken offline after critics demonstrated it could generate convincing but false information. AI expert Gary Marcus called the system's output "bullshit" in his Substack, writing, "How do I put this politely? It prevaricates. A lot." Marcus compared Galactica to OpenAI's GPT-3, noting both systems excel at producing grammatically sound but factually incorrect text.
Examples of Fabricated Output
David Chapman, another AI expert, shared an example where a user prompted Galactica to write a Wikipedia article about "bears in space." The model generated a false account of a Soviet space bear named "Bars" allegedly launched aboard Sputnik 2, mirroring the story of Laika, the Soviet space dog. Chapman described the output as "hilariously bad."
Meta's chief AI scientist Yann LeCun defended the project before and after its removal, arguing critics were taking their "shitposting" too far. Marcus characterized the rapid takedown as a public admission that the bot was released prematurely, contradicting LeCun's defense.
Marcus tweeted that large language models like GPT-3 and Galactica "are like bulls in a china shop, powerful but reckless" and "are likely to vastly increase the challenge of misinformation." The incident underscored a persistent issue in contemporary AI: systems can generate substantial amounts of on-topic text that makes little sense upon closer inspection.
Editorial Notes