The Promise Failed Once. The Warnings Already Fail Today.

When you hire a consultant, a good one signs something before opening your books. It states that they can use what they learn to help you, not to help the next client who looks just like you. That is a boundary. Someone wrote it down, and someone can be held to it.

Two of the loudest voices in enterprise software just told you that you might not have that kind of agreement with your AI vendor.

On July 12, 2026, Satya Nadella published an essay called “The Reverse Information Paradox.” His argument: you pay for a frontier model twice, once with money, and again with something even more valuable. “Models learn from ‘exhaust,’” he wrote, “the prompts people write, the tools agents use, and especially the corrections people make when the model is wrong. Every correction is distilled into institutional know-how,” the kind, he said, “a competitor could never buy.” His fix: keep your data, build your own evaluation systems, run an orchestration layer that lets you switch models without losing what you taught one of them.

Three weeks later, on Palantir’s August 3 earnings call, CEO Alex Karp said the same thing in his own register, not Nadella’s. He called the frontier labs’ relationship to their own customers a set of “Marxist overtones,” said they intend “to capture the means of production of their purported partners,” and put it as plainly as this: “You are paying for the right for them to migrate your IP, your know-how, your expertise to their model, so that they can build a competitive business that doesn’t require your business or people.” Asked why a lab would do that, he answered himself: “it’s actually being done for what they believe are moral reasons.” Then: “They are superior to you. They deserve to colonize your enterprise.” He said this after his company posted $1.9 billion in quarterly revenue, up 93 percent, and $1.1 billion in profit.

Read past the hyperbole and both men are saying something concrete: your prompts and corrections become the vendor’s training data, and the vendor turns that into products that compete with you.

So, because it is part of what I do for a living, I went and read the current public policies.

OpenAI’s own page states it plainly: “By default, we do not use data from ChatGPT Enterprise, ChatGPT Business, ChatGPT Edu, ChatGPT for Healthcare, ChatGPT for Teachers, or our API platform, including inputs or outputs, for training or improving our models.” Opt in if you want to contribute, through an explicit dashboard setting. Otherwise, nothing.

Anthropic’s page says the same thing in its own words: “Anthropic does not train generative models using code or prompts sent to Claude Code under commercial terms, unless the customer has chosen to provide their data to us for model improvement,” through a named, opt-in Development Partner Program.

Neither company hides this. Both publish it on a page anyone can read, and none of the coverage I found of Nadella’s essay or Karp’s remarks checks the “your exhaust becomes their training data” claim against the page that says the opposite, for the exact customers being warned.

The easy version of the skeptical read stops right here: sure, but of course a company pays attention to what its customers are doing in aggregate, that isn’t a scandal, that’s just how software gets built. It’s a fair instinct, and it would end the piece, except this exact promise has already been tested once, in public, at the company built almost entirely on watching what its customers do.

On July 16, 2019, Amazon’s associate general counsel, Nate Sutton, told the House Judiciary Committee’s antitrust subcommittee, under oath, that Amazon does not use individual sellers’ data to inform its own competing products: “we do not use their individual data when we’re making decisions to launch private brands.” Four months later, on November 19, 2019, in a 69-page written response to the same subcommittee, Amazon acknowledged it does use “data that is aggregated across all third-party sellers and Amazon’s first-party sales and is therefore not specific to an individual seller” to make exactly those decisions. That is the same seam Nadella and Karp are gesturing at: not your specific data, only the pattern across everyone who looks like you.

It did not stop at the seam. On April 23, 2020, a Wall Street Journal investigation reported that Amazon employees had gone further than even the aggregate promise, using an internal workaround they called “going over the fence” to pull individual sellers’ sales reports directly, across categories that reportedly included batteries and Amazon’s own Kindle business. On July 29, 2020, Representative Pramila Jayapal read Sutton’s 2019 denial back to Jeff Bezos and asked him to make the same guarantee himself. He would not: “But I can’t guarantee you that policy has never been violated.”

That is the full arc of one company’s version of the exact promise OpenAI and Anthropic are making today: a sworn denial that individual data gets used, an admission that aggregate data is fair game, an investigation that found individual data got used anyway, and a chief executive who could not personally guarantee his own company’s policy had held. Nobody had to prove malice at any step for the pattern to run its full course. The aggregate carve-out was real, it got used exactly the way an economically rational company would use it, and the narrower promise sitting on top of it did not fully hold, once someone looked closely enough.

I am not claiming OpenAI or Anthropic has done this or will. Neither company is under subpoena the way Amazon was, and I found no evidence, from either company or from any reporting on either company, that this has happened with ChatGPT Enterprise or Claude Code data. What I can say is that both vendors are asking enterprises to trust the identical distinction, your specific prompts are off-limits, the pattern across everyone who resembles you is simply not addressed, that a company with the most public, most litigated version of the same promise did not fully keep once the incentive to break it grew large enough. Nobody has to be lying today for that to be the live risk. The seam has already been proven to work as a seam, in public, under subpoena. Nobody has said it is being used that way here. Nobody has said it is not.

It is worth adding that even the narrower reading only holds for the paying, contracted tier. Consumer ChatGPT and Claude, the free and personal-subscription products, do train on your conversations by default unless you turn it off yourself. Neither Nadella’s essay nor Karp’s remarks specify which tier they mean. A warning that doesn’t say which product it is about gets to sound like it is about whichever one you use for work.

Nobody in this story is a neutral party, and that is worth noting rather than brushing past. Microsoft invested billions in OpenAI and hosts ChatGPT on Azure, and Copilot itself is built to reach into a company’s email, files, and chat, which makes Nadella’s warning apply to products his own company sells, not only his rivals’. Palantir’s entire pitch is that you should not trust a frontier lab with your operational core at all, which makes Karp’s alarm and Palantir’s product the same pitch. An interested party can still be right.

I am not exempting myself from this pattern either. Fellowship Intelligence uses the same tools everyone is being warned about, and this is not new territory for me. On June 27, 2026, under this newsletter’s previous name, I wrote a piece called “What Your AI Contract Doesn’t Cover,” which found that OpenAI’s, Anthropic’s, and Google Cloud’s own enterprise terms protect your prompts and outputs and never define or protect usage metadata, API telemetry, or the pattern across everyone who looks like you. I named the gap, told readers to ask their vendors about it, and moved on. I never went looking for a real case where that gap had actually been used against someone. Amazon’s history is the case in practice: the aggregate carve-out is not a risk you point out once and leave sitting there, it is a mechanism with a documented, admitted, litigated record of getting used exactly the way it’s incentivized.

The contract says they do not train on your prompts. Nobody has written down what they are allowed to notice or what they are allowed to do with it. We have already watched, once, in public, under oath, what a company does with that silence when the number gets big enough to matter.

SOURCES

Satya Nadella, “The Reverse Information Paradox,” July 12, 2026. https://snscratchpad.com/posts/reverse-information-paradox/

TechCrunch, After killer quarter, Palantir CEO Alex Karp calls AI industry ‘Marxist,’ August 3, 2026. https://techcrunch.com/2026/08/03/after-killer-quarter-palantir-ceo-alex-karp-calls-ai-industry-marxist/

Benzinga, Alex Karp Rips Frontier AI Labs for ‘Marxist Overtones’ After Palantir’s Blowout Earnings. https://www.benzinga.com/markets/tech/26/08/60896658/alex-karp-rips-frontier-ai-labs-for-marxist-overtones-after-palantirs-blowout-earnings-says-they-want-to-colonize-your-enterprise

AOL, Alex Karp Rips Frontier AI Labs for ‘Marxist Overtones.’ https://www.aol.com/articles/alex-karp-rips-frontier-ai-103122000.html

OpenAI, Business data privacy, security, and compliance. https://openai.com/business-data/

Anthropic, Claude Code data usage. https://code.claude.com/docs/en/data-usage

Anthropic, About the Development Partner Program. https://support.claude.com/en/articles/11174108-about-the-development-partner-program

CNBC, Amazon tells House it doesn’t favor own brands in antitrust hearing, July 16, 2019. https://www.cnbc.com/2019/07/16/amazon-tells-house-it-doesnt-favor-own-brands-in-antitrust-hearing.html

CNBC, Amazon admits to Congress it uses ‘aggregated’ data from third-party sellers, November 19, 2019. https://www.cnbc.com/2019/11/19/amazon-uses-aggregated-data-from-sellers-to-build-its-own-products.html

CNBC, WSJ investigation: Amazon uses data from third-party sellers to develop its own products, April 23, 2020. https://www.cnbc.com/2020/04/23/wsj-amazon-uses-data-from-third-party-sellers-to-develop-its-own-products.html

Washington Post, Bezos testifies he can’t guarantee Amazon’s seller-data policy hasn’t been violated, July 29, 2020. https://www.washingtonpost.com/technology/2020/07/29/bezos-testimony-data-antitrust/