Microsoft spent its Build 2026 keynote on June 2 telling enterprise buyers exactly what they wanted to hear: its new MAI models were trained “exclusively on enterprise-grade, clean and commercially licensed data.” It was the cleanest possible pitch — no copyright landmines, no lawsuits waiting to detonate, no awkward questions about where the training data came from. Three days later, Microsoft’s own research paper torched the entire claim.
The MAI-Thinking-1 preprint — published by Microsoft’s own AI team — describes a training pipeline that includes Common Crawl, the giant open repository of scraped web pages that comes with precisely zero licensing guarantees. After filtering and deduplication, the Common Crawl slice of the corpus weighed in at 24.2 billion pages. That is not a rounding error. That is the foundation of the model.
A developer in the room caught it in real time
The contradiction was first surfaced by developer Simon Willison, who was actually present at Build when the models were announced and then did the thing Microsoft apparently hoped no one would do — he read the paper. The technology outlet The Decoder ran its own investigation and confirmed it. Willison, to his credit, later updated his post with the blunt verdict: the MAI models “have the same licensing problems as all of the other major large language models.”
That sentence is the whole story. Microsoft did not build a magically clean model. It built the same kind of model everyone else builds — trained on the open, scraped, legally murky internet — and then wrapped it in marketing language designed to make risk-averse enterprise CIOs feel safe.
The trick is in the two claims Microsoft keeps conflating
Here’s the translation, because Microsoft is leaning hard on the gap between two very different statements. The genuinely defensible claim is that MAI-Thinking-1 was trained “without OpenAI data” — no distillation from GPT models, no quietly siphoning a rival’s outputs. That part appears to be true, and it’s the headline Microsoft wanted, because it sells the “we’re independent from OpenAI now” narrative the company has been pushing since the $13 billion partnership turned into a rivalry.
But “we didn’t train on OpenAI” is a completely different promise from “we trained exclusively on clean, licensed data.” Microsoft let the second claim ride on stage in front of enterprise buyers, even though its own paper says otherwise. One statement is about a competitor. The other is about copyright liability — and only one of them holds up.
Follow the money: clean data was the product
This matters because the “clean data” line wasn’t a throwaway. It was the differentiator. Microsoft is trying to sell MAI as a lower-cost, lower-risk alternative to OpenAI and Anthropic for exactly the customers who lie awake at night worrying about getting sued — banks, hospitals, law firms, governments. For that audience, “commercially licensed training data” isn’t a feature. It’s the entire reason to switch.
Strip that away and what’s left? A capable mid-sized reasoning model — 35 billion active parameters, a sparse mixture-of-experts design, a 256,000-token context window, priced to undercut Claude and GPT. Genuinely competitive on the benchmarks, reportedly preferred over Claude Sonnet 4.6 in some blind tests. But “competitive model, priced aggressively” is a far weaker sales story than “the only AI you can deploy without a copyright lawyer on retainer.” Microsoft reached for the stronger story and the paper couldn’t back it up.
Who gets hurt
The second-order effect lands on the enterprises that believed the pitch. If a regulated company chose MAI specifically because it was told the training data was clean and licensed, and that turns out to be marketing rather than fact, the indemnification math changes overnight. The copyright exposure that haunts every other frontier model — the New York Times-style lawsuits, the unresolved fair-use questions — applies to MAI too. The buyers who switched to reduce legal risk may have simply moved it from one vendor’s balance sheet to another’s, while believing they’d eliminated it.
There’s also the trust cost, which is harder to quantify but more dangerous for Microsoft. The company just spent a year positioning its AI division as the grown-up in the room — the one building responsibly, “set free” from OpenAI to pursue serious work. Getting caught overstating your data provenance, in your own published research, three days after the keynote, is the kind of thing enterprise procurement teams remember at renewal time.
The verdict
To be clear about what this is and isn’t: Microsoft didn’t do anything technically unusual. Training on Common Crawl is the industry norm, and there’s nothing scandalous about a model using the open web. The scandal is the gap between the keynote and the paper. Microsoft had a perfectly honest story — “we built a strong, cheap model independent of OpenAI” — and chose to dress it up with a “clean, licensed data” claim it knew, or should have known, its own documentation contradicted.
In an industry where every lab is fighting copyright suits and every enterprise buyer is terrified of inheriting that liability, “trust us, our data is clean” is the most valuable sentence you can say — and the most expensive one to get caught saying falsely. Microsoft said it on the biggest stage it has. The receipts were in its own PDF the whole time.