On November 15, 2022, Meta released a language model called Galactica. The promise was enormous: it had been trained on 48 million papers, textbooks and lecture notes, and it offered to reason over scientific literature, summarize studies, and help write academic articles.

On November 17 — two days later — Meta pulled it.

It wasn't taken down by a competitor or a leak. It was taken down by something far more uncomfortable: within hours, the researchers testing it found that it produced authoritative-sounding output that was false and biased — and, most revealingly, that it invented citations. It produced references with impeccable formatting — plausible authors, credible titles, real journals — pointing to papers that did not exist. As one of its critics put it: in all cases, it was wrong or biased, but sounded right and authoritative.

That scene is worth pausing on, because it explains the failure of enterprise AI better than any statistic.

The easy diagnosis (and the wrong one)

The comfortable reading is "it needed more data" or "the model was bad." Neither holds up.

Data was the one thing it had. Forty-eight million papers, textbooks and lecture notes were, arguably, one of the most curated corpora a model had been given up to that point. The bulk of it wasn't internet comments — it was scientific literature.

And there's a detail that makes the story far more interesting: Galactica did have provenance data. It was trained on more than 360 million in-context citations and tens of millions of normalized references. It knew perfectly well what a citation looks like.

What it lacked was a verifiable link between what it generated and that base. It could produce the appearance of a citation with no anchor to a real one, because it generated probabilistically instead of retrieving and citing. It had the shape of provenance, not provenance.

And that distinction — between plausible and true — is not solved by a bigger model. It's solved by being able to trace every claim back to its source.

It isn't an anecdote. It's the pattern.

You could dismiss Galactica as the stumble of an experimental model in 2022. The 2025 numbers say otherwise.

The report "The GenAI Divide: State of AI in Business 2025" — a preliminary draft (v0.1, July 2025) authored by researchers from MIT's Project NANDA, who expressly state that the views are the authors' own and not those of their institution — reviewed, according to that version, more than 300 publicly disclosed AI initiatives, interviews with representatives of 52 organizations, and survey responses from 153 senior leaders. Its headline:

"95% of organizations are getting zero return" from their generative AI initiatives, despite an estimated $30–40 billion in enterprise investment.

It deserves the caution it deserves: it's a preprint, it isn't peer-reviewed, and its methodology has been publicly challenged. We don't cite it as revealed truth, but because the number echoed through every boardroom in 2025 — and because the interesting part of the report isn't the 95%.

The interesting part is where the authors say the problem is NOT:

"The core barrier to scaling is not infrastructure, regulation, or talent. It is learning. Most GenAI systems do not retain feedback, adapt to context, or improve over time."

It's an honest and striking finding: the problem isn't in the model. It's around the model.

Our thesis: inside a company, "context" is spelled metadata

Here we put our own opinion on the table, and it's worth flagging as such: the authors document a learning and context gap; attributing it to the absence of data governance is our reading, not theirs.

And we should be precise about the mechanisms, because they are not the same. What the authors describe is a product architecture problem: systems with no memory, no feedback loop, that don't fit anyone's actual workflow. What Galactica got wrong was generating instead of retrieving. And what a governance program puts in order is a third thing: the data at rest — who owns it, what it means, where it came from. Three distinct layers, and conflating them would be exactly the kind of shortcut this article criticizes.

Our thesis isn't that data governance fixes all three. It's that it's the floor beneath all three: you can give your system memory and ground it in retrieval instead of probability, and it will still hand you the wrong number with total confidence if the table it retrieves is an orphaned copy nobody maintains. Fixing the architecture without fixing the data only means the error arrives faster and better cited.

Because "adapting to context" sounds like a property of the model, and inside a company it isn't. Your business context doesn't live in the weights of a neural network. It lives in far more boring answers:

Who owns this table?
What does this field actually mean? Is revenue before or after tax?
Where did this number come from? What process computed it, from which sources?
Is this the source of truth, or one of seven copies someone left behind?
Can it be trusted today, or has it gone eleven months without an update?

That is exactly what a data governance program produces: owners, meaning, lineage, quality, a single source of truth. That is metadata. A system cannot adapt to a context that nobody ever described.

That's why we say Data First, before AI First. Not as a slogan — as an order of operations.

Why your company is worse off than Galactica

And now the uncomfortable part, which is the real lesson of this story.

Galactica had a stroke of luck your company will not get: it failed in a domain where the error was immediately verifiable. It invented a paper that didn't exist, and any researcher with thirty seconds and a search engine could disprove it. That's why it died in three days. The system failed loudly.

Your enterprise AI has no such luxury. When you ask it:

> What were Panama's Q3 sales?

Panama's Q3 2026 sales were US$ 4.2 million,
12% above the same period last year.

…the answer sounds exactly as authoritative as a fabricated Galactica citation. But nobody is going to disprove it in thirty seconds, because a wrong number doesn't look absurd. It looks like a number.

Which table did it come from? Does it include returns? Is it the table finance uses, or the copy left over from a migration? If you can't answer that, you have a Galactica inside your company — except nobody is going to pull the plug on day three. It will feed a dashboard, and the dashboard will feed a decision.

That's the silent failure behind the 95%. It doesn't blow up — it simply never delivers value, and nobody is quite sure why.

What actually moves the needle

The conclusion isn't "don't do AI." It's that the order matters, and the boring work comes first:

Owners. Every critical data asset has a named accountable person. Without this, there's nobody to ask.
Meaning. A glossary where active customer means one thing, not four depending on the department.
Lineage. Being able to trace any number back to its origin. It's the enterprise version of the question Galactica couldn't answer.
Quality. Automated rules that raise a flag when something breaks, before it reaches a report.
A single source of truth. That one customer table exists, and everyone knows which one it is.

On that foundation, AI stops hallucinating about your business — not because the model got better, but because it finally has a context somebody took the trouble to write down. This is what the DAMA-DMBOK and DCAM frameworks have been organizing for years, long before generative AI was fashionable.

And it's program work, not tool work. No platform will tell you who owns a table: that's culture, organization, and process. The tool comes afterward, chosen to fit the architecture you already have.

The starting point

Galactica had 48 million papers and 360 million citations, and still couldn't answer "where did this come from?" It was a free warning, back in 2022, about what would happen to a good share of this decade's AI initiatives.

The useful question isn't whether your company should use AI. It's whether, when your AI gives an answer, anyone will be able to trace it back to its source.

Where does your organization stand today? A data maturity assessment takes minutes and gives you an honest starting point — what's governed, what's orphaned, and how far you are from AI you can trust on your own data.

Sources: Galactica: A Large Language Model for Science — arXiv:2211.09085 · Meta's Galactica survived only three days — MIT Technology Review · Galactica incident — OECD.AI · The GenAI Divide: State of AI in Business 2025 (v0.1, preliminary) — MIT Project NANDA

Why AI Projects Fail: The Lesson of a Model That Lasted Three Days