Meta Galactica
Meta Galactica was taken offline on November 17, 2022, surviving just three days after its public demo generated fluent scientific falsehoods and offensive text.
“In our view, Meta proved that training an AI on forty-eight million academic papers only makes it invent nonsense with peer-reviewed confidence.”
Our takeMeta Galactica was a large language model trained on scientific literature that was taken offline after just three days when users demonstrated it produced authoritative-sounding falsehoods and offensive content.
- Hype · our take
- 4/5 promised → 1/5 delivered
- Lifespan
- 2 days
- Timing · our take
- right time,
wrong execution
- Dates
- announced · shipped · discontinued
- Where
- Lab Research · Internal
The Pitch
Meta unveiled Galactica as a large language model designed to assist scientists and researchers in exploring academic literature E4E4E4E3E4. Meta trained the system on forty-eight million scientific papers, lecture notes, textbooks, and websites to store, combine, and reason about scientific knowledge E4E4E4E3E4. According to Meta, the model was capable of summarizing academic literature, solving math problems, generating Wiki articles, writing scientific code, and annotating molecules and proteins E4E4E4E3E4. The tool was intended to accelerate writing scientific literature by helping researchers explore papers and ask technical questions E4E4. Users accessed Galactica through a public web exploration demo where they could input prompts to generate academic texts complete with citations and mathematical formulas E4E4. Meta offered the model as an open-source research demonstration available directly on the web at no cost to users E4E3E4E4.
What Happened
Meta AI announced and launched the public demo of Galactica on November 15, 2022, hoping for a major public relations win following expensive metaverse investments and company layoffs E4E3E4E4. Meta promoted the release the same day, claiming the demo could generate complete papers with relevant references and formulas E4E4. However, Galactica immediately died with a whimper after meeting three days of intense criticism from researchers and the public E4.
Within hours of the launch, adversarial users and researchers discovered that the model readily generated realistic nonsense and confident scientific fabrications E3. Testers found that Galactica produced invented papers with nonexistent citations, fabricated author biographies, and made-up experimental data E4. Users also discovered that Galactica could easily be prompted to generate pseudo-scientific text defending offensive topics such as homophobia, anti-Semitism, and eating glass, while queries about racism or AIDS were blocked E4. Academics described the tool as dangerous and inaccurate, noting that its authoritative scholarly syntax masked complete fabrications E4.
Facing severe ethical backlash and widespread criticism, Meta took down the public demo on November 17, 2022, just three days after its release E4E3E4E4E3E4E4. While the public web interface was paused, Meta kept the model code and weights available under an open-source research licence E4.
Why It FailedOur take
In our view, Galactica failed primarily because of hallucination liability and the demo reality gap E3. While Meta benchmarked the system on scientific evaluations, language models optimize for linguistic fluency rather than factual truth E4. Galactica generated scientific prose with polished syntax, plausible author names, and formatted citations that did not exist in reality E4. In high-stakes scientific research, an authoritative tone paired with fluent fabrication is worse than obvious errors because only subject-matter experts can spot the falsehoods E4. Meta exposed a raw research model to the public without adequate grounding, safety guardrails, or verification mechanisms, turning a technical demonstration into an instant trust crisis E4E4.
Lessons and What Would Have Worked
- In our view, founders building AI for high-stakes domains must decouple fluency from factuality: when an AI sounds authoritative while inventing citations, users will treat every error as dangerous misinformation.
- If a product relies on technical verification, do not expose an unguarded generation demo to the public without automated fact-checking, citation retrieval, or strict output constraints.
- Benchmark performance inside an academic evaluation does not translate to public safety; test adversarial prompts and hallucination risks before opening open-ended generation interfaces.
Timeline
What happened to Meta Galactica?
- What was Meta Galactica?
- Meta Galactica was a large language model trained on scientific literature that was taken offline after just three days when users demonstrated it produced authoritative-sounding falsehoods and offensive content.
- What happened to Meta Galactica?
- Meta AI announced and launched the public demo of Galactica on November 15, 2022, hoping for a major public relations win following expensive metaverse investments and company layoffs .
- Why did Meta Galactica fail?
- In our view, Galactica failed primarily because of hallucination liability and the demo reality gap .
Receipts
- E4Why Meta’s latest large language model only survived three days online | MIT Technology Review
Supporting quote
On November 15 Meta unveiled a new large language model called Galactica, designed to assist scientists. But instead of landing with the big bang Meta hoped for, Galactica has died with a whimper after three days of intense criticism. Yesterday the company took down the public demo that it had encouraged everyone to try out.
- E3New Meta AI demo writes racist and inaccurate scientific literature, gets pulled - Ars Technica
Supporting quote
On Tuesday, Meta AI unveiled [a demo](https://galactica.org/explore/) of Galactica, a large language model designed to “store, combine and reason about scientific knowledge.” While intended to accelerate writing scientific literature, adversarial users running tests found it could also [generate realistic nonsense](https://twitter.com/mrgreene1977/status/1593274906707230721?s=20&t=hfoIU_CeB6GCWkw75vHkvw). After several days of [ethical criticism](https://twitter.com/Grady_Booch/status/1593033061423550464), Meta took the demo offline, [reports](https://www.technologyreview.com/2022/11/18/1063487/meta-large-language-model-ai-only-survived-three-days-gpt-3-science/) MIT Technology Review.
- E4Why Meta's Galactica AI Model Didn't Last Long; A Misinformation Generator?
Supporting quote
Previously, Meta's Galactica, the company's new large language model (LLM), was launched.  This photograph taken on October 28, 2021 shows the META logo on a laptop screen in Moscow as Facebook chief Mark Zuckerberg announced the parent company's name is being changed to "Meta" to represent a future beyond just its troubled social network. Photo by KIRILL KUDRYAVTSEV/AFP via Getty Images The tech giant manufacturer has lots of plans for this artificial intelligence. Meta claimed that Galactica could help scientists and other researchers to store, combine, and reason scientific knowledge. "You can use it to explore the literature, ask scientific questions, write scientific code, and much more," said Meta through the official [Galactica website](https://galactica.org/mission/). However, the company had to shut it down a few days after the first demo. ## Why Meta's Galactica AI Model Didn't Last Long According to [CNET](https://www.cnet.com/science/meta-trained-an-ai-on-48-million-science-papers-it-was-shut-down-after-two-days/)'s latest report, Meta released Galactica's demo version on Tuesday, Nov. 15.  A tablet is seen in the center of a high tech art installation at Salt Galata on May 6, 2017 in Istanbul, Turkey. The "Archive Dreaming" installation by artist Refik Anadol uses artificial intelligence to visualize nearly 2 million historical Ottoman documents and photographs from the SALT Research Archive. Controlled by a single tablet in the center of a mirrored room the artist used machine learning algorithms to combine historical documents. Photo by Chris McGrath/Getty Images **Also Read :** [**DeviantArt Revealed a New Safeguard That Blocks Art-generating AI Systems**](https://www.techtimes.com/articles/283280/20221111/deviantart-revealed-new-safeguard-blocks-art-generating-ai-systems.htm) The tech firm said Galactica was trained on 48 million papers, lecture notes, textbooks, and websites. But, based on the comments made by experts, the training conducted by Meta for its new LLM wasn't enough. Carl Bergstrom, a biology professor at the University of Washington, used the new Galactica demo version. After using the new AI tech, Bergstrom described it negatively, saying that it is a "random bullsh\*\* generator." [Michael Black](https://twitter.com/Michael_J_Black/status/1593133722316189696?ref_src=twsrc%5Etfw%7Ctwcamp%5Etweetembed%7Ctwterm%5E1593133722316189696%7Ctwgr%5E120f3aa238f31b75160917e55dc8a64893eba862%7Ctwcon%5Es1_&ref_url=https%3A%2F%2Farstechnica.com%2Finformation-technology%2F2022%2F11%2Fafter-controversy-meta-pulls-demo-of-ai-model-that-writes-scientific-papers%2F), the Max Planck Institute for Intelligent Systems director, shared his experience. He said that the misinformation generated by Galactica is dangerous. Black added that although the LLM of Meta sounded right and authoritative, it was biased or wrong in all cases. ## Meta Defends Galactica Because of Galactica was generating false information instead of reliable details, Meta had to shut down the AI on Thursday, Nov. 17.
- E4Trust large language models at your own peril | MIT Technology Review
Supporting quote
When Meta launched Galactica, an open-source large language model designed to help scientists, the company—reeling from criticism of its expensive metaverse investments and its recent massive layoffs—was hoping for a big PR win. Instead, all it got was flak on Twitter and a spicy blog post from one of its most vocal critics, ending with its embarrassing decision to take the public demo of the model down after only three days. According to Meta, Galactica can “summarize academic papers, solve math problems, generate Wiki articles, write scientific code, annotate molecules and proteins, and more.” But soon after its launch, it was pretty easy for outsiders to prompt the model to provide “scientific research” on the benefits of homophobia, anti-Semitism, suicide, eating glass, being white, or being a man. Meanwhile, papers on AIDS or racism were blocked. Charming!
- E3What Meta learned from Galactica, the doomed model launched two weeks before ChatGPT | VentureBeat
Supporting quote
One year ago — and two weeks before OpenAI released ChatGPT — Meta released a research demo called [Galactica](https://arxiv.org/abs/2211.09085). An open source "large language model for science" that was trained on data including [48 million scientific papers](https://www.cnet.com/science/meta-trained-an-ai-on-48-million-science-papers-it-was-shut-down-after-two-days/), Meta touted Galactica's ability to "summarize academic literature, solve math problems, generate Wiki articles, write scientific code, annotate molecules and proteins, and more." Galactica survived publicly for only three days. On November 17, 2022, Meta took down the demo after an outcry over what was, back then, a word that had not yet made it into the mainstream: Hallucinations. [Many were appalled](https://venturebeat.com/ai/what-metas-galactica-missteps-mean-for-gpt-4-the-ai-beat/) by Galactica’s sometimes very unscientific output, which, like other LLMs, included information that sounded plausible but was factually wrong and in some cases also highly offensive.
- E4Meta pulls Galactica after three days • It Does What Now?
Supporting quote
and put a public demo online alongside claims that it could write papers, literature reviews and Wikipedia-style articles complete with citations and formulas. Meta’s chief scientist [Yann LeCun](https://itdoeswhatnow.com/tag/yann-lecun/) promoted it the same day, saying the demo could “generate a paper with relevant references, formulas, and everything.” Within hours, researchers and other users found the demo readily produced confident, fluent, scientific-sounding text that was false — invented papers with plausible author names and citations that did not exist, fabricated data, and content criticised as biased. The problem was not that Galactica sometimes said something questionable; it was that its errors were formatted exactly like real scholarship, complete with the syntax of citation and method, so that only a reader already expert in the specific claim could tell fabrication from fact. Critics argued this made it more dangerous than a text generator with no scientific pretensions, since its output was designed to look authoritative. Meta took the public demo offline after three days, saying in a statement that it had “paused the demo for now” in response to feedback. LeCun continued to defend the underlying research, tweeting sardonically three days after launch that the demo was “no longer possible to have some fun by casually misusing it.” The model’s weights and code remained available for research use under a Creative Commons licence. The episode became a widely cited case study in the gap between a model performing well on benchmarks — Galactica’s paper reported strong scores on several scientific and mathematical evaluations — and a model being safe to expose to the public, arriving two weeks before ChatGPT’s launch made the same tension a mainstream concern rather than a research-community one.
- E4Meta pulled its Galactica science AI after three days of confident fabrications · AI Failure Index
Supporting quote
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.
Sentiment
At launch
No launch coverage quote recorded.
At death
No death coverage quote recorded.
No sentiment quotes recorded.
Last updated · Phil · How we source entries
Changelog
- — Published Meta Galactica.
- — Published Meta Galactica.