The dispute is about whether OpenAI’s Pentagon agreement actually guarantees safeguards that Anthropic had demanded—not simply whether either company supports military AI. Anthropic CEO Dario Amodei said the Pentagon’s proposed terms did not adequately bar mass domestic surveillance and fully autonomous weapons. OpenAI CEO Sam Altman later said OpenAI’s agreement included protections against those uses. Amodei reportedly challenged that description as misleading, but the public record does not establish a symmetrical exchange in which Altman personally called Amodei a liar.
The unresolved question is contractual: do OpenAI’s published safeguards create enforceable limits, or do they rely in part on the broad “all lawful purposes” standard, existing government rules and technical controls? The full agreement and its enforcement terms are not in the public material cited here, so neither company’s summary alone settles that question.
As an Amazon Associate I earn from qualifying purchases.
What happened, in brief
Anthropic objected to Pentagon contract language requiring its AI system to be available for “all lawful purposes.” The company said that standard did not adequately protect against two uses: mass domestic surveillance of Americans and fully autonomous weapons. On February 26, 2026, Amodei said negotiations had reached an impasse over those safeguards. Anthropic’s statement
OpenAI announced a Department of War agreement on February 28. It said the deal prohibited mass domestic surveillance and preserved human responsibility for the use of force, while allowing use for all lawful purposes. Amodei reportedly told Anthropic employees that OpenAI’s public account was misleading or false. That reported criticism is not the same as a published, direct exchange of identical “liar” accusations between the two CEOs.
#1 Best Overall
The Pentagon defended its lawful-use position and denied seeking illegal surveillance or weapons use without human involvement. The disagreement is therefore partly about competing descriptions of the deal—and partly about what the agreement actually binds the government to do.
How the dispute unfolded
| Date | What happened | What the source establishes |
|---|---|---|
| February 26, 2026 | Anthropic published Amodei’s account of the impasse. | Anthropic said the disputed issues were mass domestic surveillance and fully autonomous weapons. Anthropic |
| February 27, 2026 | Anthropic responded to comments from the defense secretary. | The company said the Pentagon’s final language made virtually no progress on its two requested safeguards. Anthropic |
| February 28, 2026 | OpenAI announced an agreement to deploy its systems in classified government networks. | OpenAI described safeguards against mass domestic surveillance and limits concerning autonomous weapons and human control. OpenAI |
| March 4–5, 2026 | Anthropic said it had received notice that it was designated a supply-chain risk; reporting on a leaked internal memo intensified the public dispute. | Anthropic characterized the designation as retaliation for its position. That is the company’s account, not an independently established explanation of the government’s decision. Anthropic · AP |
| March 23, 2026 | Senator Ron Wyden’s office raised concerns about data-broker information and surveillance authorities in the context of AI. | The letter illustrates why the legality and safeguards around data access are contested; it does not establish that a particular Pentagon use occurred. Wyden’s letter |
What each side said about the safeguards
| Issue | Anthropic’s position | OpenAI’s published position | Pentagon’s position | What remains unclear |
|---|---|---|---|---|
| Mass domestic surveillance | Anthropic sought an explicit exclusion for mass surveillance of Americans. It said lawful foreign-intelligence and counterintelligence work could be acceptable, while warning that AI can amplify surveillance capabilities. | OpenAI said: “No use of OpenAI technology for mass domestic surveillance.” | The Pentagon said it sought lawful use, not illegal surveillance. | The publicly described language does not, by itself, show the definitions, exceptions, enforcement process or remedy in the binding agreement. |
| Autonomous weapons | Anthropic said current systems were not reliable enough to control weapons fully autonomously and sought a specific prohibition. | OpenAI said restrictions apply to independently directing autonomous weapons where law, regulation or Department policy requires human control, and described human approval for certain high-stakes decisions. | The Pentagon denied seeking weapons use without human involvement. | The summaries do not establish what counts as meaningful human control, which decisions require approval, or whether the contract bars uses beyond existing rules. |
| General permitted use | Anthropic objected to treating “all lawful purposes” as sufficient without its requested exceptions. | OpenAI said the Department could use its system for “all lawful purposes,” alongside the stated safeguards. | The Pentagon defended lawful-use authority. | Whether specific limits override the broad clause, and who resolves a disagreement about a proposed use, cannot be confirmed without the full contract. |
Why “all lawful purposes” is the fault line
“Lawful” is a legal threshold, not a complete safety standard. A company may decide that a use is too intrusive or risky even if government officials believe existing statutes and policies permit it. Conversely, a broad contract can still operate under limits imposed by law, regulation or agency policy. The real question is how those rules apply to a particular use and whether the contract adds protections beyond them.
Mass domestic surveillance is not one legally uniform activity. The answer can depend on the data source, who is being targeted, the agency involved, the applicable authority and whether a warrant or other court process is required. Anthropic’s concern, as it described it, was that AI could make collection and analysis of information—including information obtained from data brokers—more extensive. Wyden’s letter raised related concerns; it does not prove that any specific surveillance program violated law.
Rank #2
The weapons question also turns on distinctions that public summaries can blur. AI-assisted analysis or targeting support is not automatically the same as a system independently selecting and attacking a target. Nor does the phrase “human in the loop” alone show how much authority a person has, what information they receive or whether they can meaningfully stop an action.
What the public statements do—and do not—prove
OpenAI’s announcement is evidence of what OpenAI says it agreed to. Anthropic’s statements explain why it considered the Pentagon’s terms inadequate. Neither is a substitute for the complete contract. The public descriptions available here do not establish all operative definitions, exceptions, enforcement provisions or amendment procedures.
A useful way to evaluate any claimed safeguard is to ask:
- Is the restriction written into the binding contract? A public commitment or company usage policy may not have the same status as a negotiated contract term.
- Does it prohibit a use outright, or require only compliance with existing law and policy? Those are different levels of protection.
- Who decides whether a proposed use crosses the line? The answer matters when the customer and vendor interpret a term differently.
- What happens if the restriction is breached? A rule without a clear enforcement mechanism may not provide the same practical protection as an enforceable contractual limit.
OpenAI’s wording combines specific commitments with references to law and Department policy. Amodei’s reported criticism was that this did not amount to the same protections Anthropic had sought. That is a dispute over the language and effect of the deal, not proof on its own that OpenAI’s description was knowingly false.
Who called whom a liar?
The headline’s symmetry overstates what the cited accounts establish. Amodei reportedly described OpenAI’s account as misleading or false in an internal communication later reported by other outlets. The memo is a leaked document, and its wording should be treated as reported rather than as a public statement directly addressed to Altman. GizChina’s account
A separate account says Pentagon official Emil Michael accused Amodei of misrepresenting the negotiations. That does not establish that Altman personally called Amodei a liar. Axios’s account
So the supported description is narrower: Amodei reportedly attacked the truthfulness of OpenAI’s public account, while a Pentagon official criticized Amodei’s account. The available evidence does not support presenting this as a verified, equal back-and-forth of personal accusations between the two CEOs.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.Neither company’s position is simply “for” or “against” military AI
Anthropic said it supported American warfighters and had been deployed in classified government networks since June 2024. Its stated objection was to particular uses, not to all defense or intelligence work. The company distinguished intelligence support from mass surveillance and from weapons operating fully autonomously. Anthropic’s response
Free tools Windows power users keep installed
One-click scans. No signup required.
OpenAI, for its part, said its agreement included restrictions on surveillance and safeguards around human responsibility for force. Its announcement does not establish that it approved autonomous weapons without human control. Claims that the companies’ contracts were identical, or that either company accepted every use the Pentagon might consider lawful, go beyond what these public summaries show.
Best Value
Why the dispute matters beyond the two CEOs
For government customers, the argument highlights a basic procurement problem: a model provider may want limits on how its system is used, while a government customer wants operational flexibility under its interpretation of law and policy. If those limits are not clear in the contract, public assurances may leave both the vendor and the public uncertain about what is actually permitted.
For AI companies, accepting government work can bring revenue, access and influence, while refusing terms can carry commercial and political costs. The companies are also competitors for government contracts and for customers who care about responsible-AI commitments. Those incentives make careful scrutiny of both companies’ claims appropriate; they do not, by themselves, prove either company’s motives.
For businesses and individual users, this controversy is relevant as a signal about corporate policy and vendor risk, not as a complete comparison of ChatGPT and Claude. Consumer products, enterprise services, APIs and classified deployments involve different terms and uses. A purchasing decision should be based on the applicable privacy terms, data controls, contractual restrictions, administrative features and required capabilities—not on a defense-contract announcement alone.
Quick wins for a faster PC:
Clear out junk files and repair common Windows errorsFree Scan →Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →Repair Windows errors before they cause bigger problemsFix Now →What remains unresolved
- Whether the full OpenAI agreement contains specific, enforceable prohibitions that match the company’s public summary.
- How the agreement defines mass domestic surveillance, autonomous weapons and human approval.
- Who interprets those limits and what remedy applies if the government or vendor alleges a breach.
- Whether OpenAI later amended or clarified the agreement in a way that changed its binding terms. Coverage reports a clarification and quotes Altman calling the initial announcement “opportunistic and sloppy,” but the exact revised text and its legal effect are not established by the sources cited here. AP coverage
- The full basis for the government’s supply-chain-risk designation of Anthropic. Anthropic described it as retaliation; the company’s characterization alone does not resolve the government’s rationale.
The central disagreement is therefore not settled by the word “liar.” It turns on whether the contractual language, definitions and enforcement mechanisms deliver the restrictions the companies say they do.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
