Product1 publisher2 min readPublished
Internal OpenAI and Microsoft memos about substitution now anchor the Times' liability motion
A new filing in the New York Times case against OpenAI and Microsoft builds its market-harm argument out of the two companies' own internal documents, obtained in discovery. Anyone shipping an answer feature on these models sits downstream of it.
The Product Desk · Product desk
What happened
- A new filing in the New York Times case, reported by Thurrott.com, opens by quoting a Microsoft Director of Applied Science on "an astonishing theft of unprecedented proportions".
- OpenAI's head of ChatGPT wrote internally that the models were an existential threat to news publishers because they are "largely substitutive, period," and would get more so as they improved.
- An internal Microsoft document says the company's AI content strategy started a "doom loop" that will hurt both the performance of its models and the entire web.
- The Times is asking for summary judgment of liability across four categories of copying, one of which is the "horse trading" by which the two companies shared content with each other.
Compiled by The Product DeskSomething wrong?How this is made
Why it matters
- exposure A customer cannot audit its way clear of this one. The market-harm evidence sits in Microsoft and OpenAI files surfaced in discovery, so the risk reaches anyone shipping on the models without ever touching their own logs.
- decision Any roadmap item whose success measure is that the user never leaves your surface now has to be defended against the same substitution theory a judge is being asked to rule on.
- constraint Publishers negotiating content licences have internal admissions to quote back. That raises the floor on what a cheap corpus deal can look like.
- contradiction The companies argue publicly that their models transform the source into something that does not compete with it, while the documents quoted in the motion describe outputs that replace it, and the selection of quotes was made by the plaintiff.
A reader opens a chatbot, asks what happened overnight, reads four sentences and closes the tab. Microsoft CEO Satya Nadella is quoted in the filing describing that pattern, saying the chatbots "substituted" original sources by "giving you the information right there on ... the AI platform versus needing to go to the underlying source" [7]. The Times put the line in its motion because, it argues, substitution of that kind "eviscerates" the fair use defense [5].
Teams that ship answer surfaces tend to tell themselves the user comes back to the source for depth. The filing documents traffic drops at the New York Times, ZDNet and other content makers as the AI supplied what people wanted and they stopped clicking through [10]. OpenAI cofounder and president Greg Brockman wrote internally, on separate occasions, that the models were "excellent at news" and "very good at any news task" [6].
The same internal Microsoft document that describes a doom loop also names the dependency underneath it: "It is highly unusual that an end-product threatens the economic foundations of its essential suppliers, but that is the situation we have created for our LLM business with respect to its 'content supply chain'" [9]. That sentence is about Microsoft's own inputs, and it is the part an operator should read twice, because the suppliers in question are the same web your retrieval layer points at.
These are the plaintiff's quotations, pulled from discovery and arranged to win a motion, and the account in Thurrott.com does not include a response from OpenAI or Microsoft [1][15].
For a team shipping an answer feature next week, the working distinction is between an answer that sends a user onward and an answer that finishes the errand. Two measurements separate them: the share of answers after which the user opens a cited source, and the share of answers that carry everything the source had, so that opening it adds nothing. Traffic to your own surface rises in both cases, which is why it cannot tell you which product you built. The second measurement is the one the Times is calling substitution, and it is the ground the fair use question is being fought on [5].
The filing says there is no dispute that OpenAI and Microsoft participated in each type of copying, and that they "cannot carry the burden of proving their fair use defense for any of them" [12].
What to watch
- The defendants' opposition brief, and whether it places the 'doom loop' document in a wider context.
- Whether the court grants liability without trial on any of the four copying categories.
- Whether publishers in licence talks start quoting the internal substitution language across the table.