Microsoft's Own Director Called AI Scraping "the Largest Theft of Labor in Human History." That Sentence Does Not Decide Fair Use. It Decides the Damages.

Eleven words unsealed on 17 September 2026 have been circulating for four days now, and almost every outlet carrying them has filed them under the wrong question. Brent Hecht, Microsoft’s Director of Applied Science, wrote in a January 2023 internal memo that the mass collection of published work for AI training was “the largest theft of labor in human history,” and, in the same document, “an astonishing theft of unprecedented proportions.” The statements surfaced through newly unredacted briefs in the consolidated copyright litigation brought by The New York Times and other news publishers against OpenAI and Microsoft, and were reported by TechCrunch, the Washington Post, Tom’s Hardware and TechRadar the same day.

The near-universal framing is that a defendant has confessed. It has not, and the reason is structural rather than rhetorical. A fair-use ruling does not ask what an employee of the defendant thought about the ethics of the practice. There is no statutory slot for it. Where a document like this one actually does damage is somewhere else entirely — in the willfulness finding that sets the ceiling on statutory damages. The memo is close to irrelevant to whether OpenAI and Microsoft win. It is potentially decisive as to what losing costs.

The eleven words, and who said them

Start with the title, because several outlets have been loose with it. Hecht is Microsoft’s Director of Applied Science. He is not general counsel, not a corporate officer, and not a decision-maker on training-data policy. Headlines calling him an “exec” are doing work the org chart does not support.

Microsoft has answered on the record. Spokesman Alex Haurek said Hecht’s writings reflected “one employee’s individual perspective,” were not legal analysis, and did not represent the company’s views, and restated Microsoft’s position that training on published material is transformative fair use. That is a real response, not a no-comment, and it is also the strongest version of the defendants’ argument: a large company that debates a contested practice internally, in writing, and proceeds anyway looks exactly like a large company that debated a contested practice internally and concluded it was lawful.

A second Hecht document in the same tranche warned that the companies’ systems risked creating a “doom loop” — suppressing traffic to the publishers whose work trains the models, and so starving the content supply the models depend on. That one is an economic argument, not a legal one, and it is the more interesting of the two. It has received a fraction of the attention, because it is not eleven quotable words.

The four questions a fair-use ruling actually asks

Section 107 of the Copyright Act gives courts four factors. None of them is about intent, remorse, or internal dissent.

• Purpose and character of the use — including whether it is commercial, and how far it transforms the original into something with a different purpose. This is where the defendants live or die, and it turns on what a trained model is, not on who objected to building it.
• Nature of the copyrighted work — factual reporting sits closer to the thin end of protection than fiction does, which cuts modestly toward the defendants on news content.
• Amount and substantiality used — whole articles were ingested, which cuts toward the plaintiffs, though courts have accepted wholesale copying where the end use is genuinely transformative.
• Effect on the market for the original — the factor that has decided most recent AI copyright disputes, and the one the “doom loop” memo is arguably evidence about.

Read that list again with the viral quote in hand. An employee’s moral objection does not attach to any of the four. It is not a legal conclusion, it is not admissible as one, and a court weighing factor one does not get closer to an answer by learning that a director in applied science found the practice distasteful. The only factor it touches at all is the fourth, and only indirectly — the doom-loop memo, not the theft line, is the document a plaintiff would actually cite on market harm.

Where the memo does land: willfulness

Under 17 U.S.C. §504(c), a plaintiff who elects statutory damages instead of proving actual losses recovers between $750 and $30,000 per infringed work, at the court’s discretion. If the infringement is found to have been willful, the ceiling rises to $150,000 per work. Willfulness is not a question about whether infringement occurred. It is a question about what the defendant knew, or recklessly disregarded, while it occurred.

That is precisely the question a contemporaneous internal memo speaks to. This is why discovery in copyright cases is hunted for documents of exactly this shape — not to prove the act, which is usually undisputed, but to characterise the state of mind around it. A January 2023 memo from inside the defendant, using the word theft, is the archetype.

Now the arithmetic, and it needs its caveats stated in the same breath. The filings describe a Microsoft crawler project, Project Mango, that produced a dataset containing at least 160,903 unique works belonging to the news-publisher plaintiffs. Apply the statutory range to that pleaded figure and you get the commercial stake of the willfulness finding: roughly $121 million at the $750 floor, about $4.8 billion at the $30,000 standard ceiling, and about $24.1 billion if every work were awarded the $150,000 willful maximum.

Those numbers are a statutory range applied to a pleaded count. They are not a forecast of an award, and treating them as one would be the same error as treating the memo as a confession. Three things constrain them hard. Works must be separately registered with the Copyright Office to be separately compensable. Courts routinely group works — all parts of a compilation count as one work for statutory purposes, which is a live question for a newspaper archive. And the $150,000 figure is a maximum a court may award, not a number it must; the discretion runs across the whole range even after a willfulness finding.

What the memo changes is which end of that range is reachable. That is worth more than a headline suggests and considerably less than the headlines have implied.

The unglamorous allegations that may matter more

Two other items in the unsealed material are harder to quote and easier to prove.

Project Taxi

Between 2019 and 2022, Microsoft supplied OpenAI with a compilation of billions of webpages assembled for the Bing search index, under the internal name Project Taxi. The filings describe this as an undisclosed transaction, and say the transferred data included Times content. The significance is not the copying — it is that a search index built for one purpose was commercially transferred for another, between two named parties, on dates. That is a cleaner factual record than a memo, and it bears on factor one in a way an employee’s opinion does not.

The copyright-notice allegation

The filings also allege paywall circumvention and the deliberate stripping of copyright management information from training data. That second allegation is not part of the infringement claim at all. Removing or altering copyright management information is a separate statutory violation under DMCA §1202, with its own damages provision and its own elements, and it does not require the plaintiff to win the fair-use argument. A defendant can prevail on fair use and still face §1202 exposure. Anyone merging the two claims into one story about “the AI copyright case” is losing the part with an independent path to liability.

What the quote is actually good for

The litigation is now a consolidated multi-district proceeding covering dozens of suits over pirated training corpora, model training on protected work, and outputs resembling the originals. Nothing unsealed on 17 September resolves any of it. No court has ruled. No new legal development occurred — what happened is that a redaction came off.

If you saw the quote and concluded the case was over, the honest correction is narrower and duller than the headline. The memo does not make OpenAI and Microsoft more likely to lose. It makes losing more expensive, by putting the willfulness multiplier in play. That is the question it answers, and it is not the question anyone is sharing it to answer.