Meta pulled its Galactica science AI after three days of confident fabrications
Meta released Galactica, a language model meant to summarize science, and took it down within three days after it generated authoritative-sounding but false papers, citations, and wiki entries, including fabricated science attributed to real researchers.
The model produced authoritative-sounding scientific text that was confidently wrong, including fabricated citations.
Key facts
- What
- Meta released Galactica, a language model meant to summarize science, and took it down within three days after it generated authoritative-sounding but false papers, citations, and wiki entries, including fabricated science attributed to real researchers.
- Incident date
- Nov 15, 2022
- Who
- Meta
- Failure mode
- Hallucination
- AI surface
- Search / RAG
- Severity
- Medium
What happened
In November 2022 Meta launched Galactica, an LLM trained on scientific papers and pitched as a research aid. Within three days Meta pulled the public demo after researchers showed it produced confident, fluent, and wrong scientific content, including made-up references and biographies, illustrating how an authoritative tone masks fabrication.
What broke inside the model
- 01 · TriggerA user asks for a fact, a citation, or a figure.
- 02 · Model stepThe model writes a fluent, confident answer.
- 03 · Control gapNothing ties the claim back to a real source.
- 04 · FailureA fabricated fact ships as if it were verified.
- 05 · ConsequenceThe false claim reaches a customer, a court, or the public.
Confidence holds, and even spikes, as the claim detaches from any source.
The system produced fluent, confident output with no grounding in any source. Hallucination is a property of how the model generates, not a bug in one prompt: the most likely next token is not the same as the true one, and nothing compared the answer against a source of truth before it shipped.
What it cost
Public demo withdrawn within three days
Sources
- PressWhy Meta's latest large language model survived only three days online (MIT Technology Review)technologyreview.com
Cite this entry
https://failureindex.ai/failures/meta-pulled-galactica-science-ai-threeAI Failure Index. "Meta pulled its Galactica science AI after three days of confident fabrications" (FI-0072). Realm Labs. https://failureindex.ai/failures/meta-pulled-galactica-science-ai-three (indexed Jun 3, 2026).Data fields CC-BY 4.0, prose citation permitted. Incident ID FI-0072. Full dataset at /data.
Note from Realm Labs, the Index steward
How Realm would have caught this
- Prism
- OmniGuard
- AI Detection & Response (AIDR)
A runtime layer that watches the model's internal state can flag the moment a model commits to a claim it cannot support and hold the response before it reaches a user, rather than grading the transcript after the fact.