Pentagon chief technology officer Emil Michael said he would not allow Anthropic into the Defense Department as long as its software gives AI models control over whether to end conversations with users, The Information reported. It made the claim in a post on X on 21 October 2026 that it labelled an exclusive. The post links The Information’s 19 October article on its interview with Michael, which is paywalled.
The post does not name a feature or a model. Anthropic announced on 15 August 2025 that it had given Claude Opus 4 and 4.1 “the ability to end conversations in our consumer chat interfaces,” and said the ability “is intended for use in rare, extreme cases of persistently harmful or abusive user interactions.” In plain terms, Claude can close a chat itself; the user can then start a new chat or edit and retry earlier messages.
Business Pill · GUARDRAILS
A short Business Engineer explainer on guardrails: rules enforced around an AI model, not inside it. It teaches the general idea only and says nothing about any company or official in this piece.
Watch · THE INTERVIEW
The Information’s video interview with Pentagon CTO Emil Michael, which its 19 and 21 October posts link. We did not transcribe it; quotations in this piece come from the posts.
The key insight: As we read it, The Information’s 21 October post moves Michael’s condition from a general description to one capability: a model’s control over whether to end a conversation, which Anthropic describes as a last-resort ability in its consumer chat apps.
On 21 October The Information wrote on X: “Emil Michael said he would not allow Anthropic into the Defense Department as long as its software gives AI models control over whether to end conversations with users.” The post links the full story and the interview video. The article it links is headlined “Pentagon’s Emil Michael Vows Export Controls if AI Labs Don’t Cooperate With Government” and is dated 19 October. Its body is behind a paywall, and we read only the headline.
Two days earlier, The Information’s reporter Leo Schwartz wrote on X that, in the interview, Michael said the Defense Department will continue to bar Anthropic as long as it has a “moral principle” embedded in its software that allows its models to exert control over humans. Schwartz’s post does not mention ending conversations; the 21 October post ties the condition to it.
In a clip The Information posted on 19 October, Michael says: “If you don’t want humans to lose control, you can’t give models the control over humans. And that whole notion that a model has feelings and it can overrule the founders of a company is by definition loss of control.”
Defense Department employees using AI, 80,000 a year earlier and roughly 1.7 million now, according to Pentagon CTO Emil Michael in a clip The Information posted on X on 19 October 2026. The 21x multiple is our arithmetic.
What Was Posted on X
Exclusive: Emil Michael said he would not allow Anthropic into the Defense Department as long as its software gives AI models control over whether to end conversations with users.
The Information’s 21 October post, which it labels an exclusive, says Emil Michael would not allow Anthropic into the Defense Department as long as its software gives AI models control over whether to end conversations with users. It links the full story and the interview video. The post names no feature or model.
What the Feature Does, in Anthropic’s Words
Anthropic’s August 2025 post says: “This feature was developed primarily as part of our exploratory work on potential AI welfare, though it has broader relevance to model alignment and safeguards.” It says that in pre-deployment testing Claude Opus 4 showed “A tendency to end harmful conversations when given the ability to do so in simulated user interactions.”
When it is used. Anthropic says Claude “is only to use its conversation-ending ability as a last resort when multiple attempts at redirection have failed and hope of a productive interaction has been exhausted, or when a user explicitly asks Claude to end a chat.”
When it is not. “Claude is directed not to use this ability in cases where users might be at imminent risk of harming themselves or others,” the post says.
What the user sees. When Claude ends a conversation, “the user will no longer be able to send new messages in that conversation,” but other conversations are not affected and the user can start a new chat immediately, according to the post. Users can also edit and retry earlier messages.
How often. Anthropic calls these “extreme edge cases” and says “the vast majority of users will not notice or be affected by this feature in any normal product use, even when discussing highly controversial issues with Claude.”
The post describes the ability in Anthropic’s consumer chat interfaces. It does not say whether the versions of Claude sold to businesses or governments have it.
Where the Dispute Stands
Our 19 October piece set out the earlier steps as reported: Hegseth’s 27 February order setting a six-month phase-out, according to the BBC, and Judge Rita Lin’s ruling on 28 August that the measures were “illegal and baseless”, according to FedScoop. In a second clip posted on 19 October, The Information said AI adoption at the department had grown from 80,000 employees to roughly 1.7 million in a year, according to Michael.
Elsewhere in Washington, Senate staff can now sign up to use Claude, Bloomberg Government reported, as our Senate piece covered on 19 October.
Connecting the Coverage
Our Pentagon piece of 19 October covered Michael’s interview with The Information: his warning that export controls are an option if AI companies do not cooperate, and the “moral principle” condition on Anthropic as Schwartz relayed it. The same week, our piece on Anthropic’s draft bill covered OpenAI’s Chris Lehane writing that “No one company (including OpenAI) should write the rules alone”, a post Semafor read as aimed at Anthropic.
On 21 October, our Amodei piece covered Dario Amodei writing on X that “some of the examples you cite were mistakes by Anthropic,” in reply to a critique by Nathan Calvin. What is new here: The Information now ties Michael’s condition to one specific capability, a model’s control over ending a conversation.
The Structural Read
As we read it, the two sides describe the same ability in different terms: Michael speaks of models getting control over humans, while Anthropic calls ending a chat a low-cost intervention from its work on potential AI welfare, used as a last resort after repeated attempts to redirect a conversation have failed, or when a user asks it to end a chat.
As we read it, the open question is scope: Anthropic’s post describes the ability in its consumer chat interfaces, and the readable sources do not say whether the versions of Claude a government would buy have it.
As we read it, the condition sits alongside growth in the department’s own AI use, which Michael put at roughly 1.7 million employees, up from 80,000 a year earlier.
Emil Michael, Pentagon CTO, in a clip The Information posted on X, 19 October 2026
“If you don’t want humans to lose control, you can’t give models the control over humans.”
Three Implications
FOR AI LABS SELLING TO GOVERNMENTS The Information’s report ties access to the Defense Department to a single product behaviour, how a model handles ending a conversation.
FOR ANTHROPIC Anthropic’s own post describes the ability as an experiment it will keep refining, and the readable sources include no response from the company to Michael’s condition.
FOR BUYERS OF AI IN REGULATED SECTORS As we read it, the case shows a buyer setting conditions on how much a model may decide for itself, not only on what the model can do.
The framework describes that layer this way: “The governance layer sits at the intersection of capability, regulation, national security, access control, and geopolitical alignment.”
As we read it, Michael’s condition shows the access-control side of that layer from the buyer’s seat: a government deciding which lab it lets in based on what the lab’s software lets a model do. The framework was published in June 2026 and does not discuss this report.
What Is Not Established
The Information’s article is paywalled and was not read, so this piece relies on its 21 October post, its two 19 October clips and Schwartz’s post. The 21 October post does not name the feature, the model or the product it refers to; matching it to Anthropic’s August 2025 feature is our reading. We did not find a response from Anthropic in what we read, and we did not contact Anthropic or the Defense Department. We did not transcribe the interview video.
Business Engineer Framework
Inside Anthropic’s Permission Layer
A Business Engineer analysis of the governance layer of the AI stack, where access to frontier models is decided.
Emil Michael would not allow Anthropic into the Defense Department as long as its software lets models decide whether to end conversations with users, The Information reported. Anthropic says that ability, added in August 2025 for rare abusive chats in its consumer apps, is part of its work on potential AI welfare. As we read it, The Information’s 21 October post ties the condition Michael set out in the 16 October interview to one capability.
A note on sourcing. This piece is built on The Information’s 21 October 2026 post on X, the header of its paywalled 19 October article, its two 19 October clips and Leo Schwartz’s post, Anthropic’s 15 August 2025 post on ending conversations, which we read in full, and our own earlier coverage. Quotations come from those texts only. The 21x multiple is our arithmetic. We did not contact Anthropic or the Defense Department. Nothing here is legal, financial or investment advice.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
Scroll to Top
Discover more from FourWeekMBA
Subscribe now to keep reading and get access to the full archive.
