Pentagon–OpenAI Contract Release Includes Disputed ‘Minimal Refusal’ Language

OpenAI and the Pentagon deny that the wording appeared in the final executed agreement; a signed February 6 record and a Justice Department reversal complicate the released document's status.

By 2 min read
Pentagon–OpenAI Contract Release Includes Disputed ‘Minimal Refusal’ Language
Pentagon–OpenAI Contract Release Includes Disputed ‘Minimal Refusal’ Language

Listen to this story

The audio brief

About 1:33
0:001:33
Read transcript
The Pentagon and OpenAI now say a released contract document contains wording that was removed before the agreement was signed. The disputed phrase describes national-security “Mission Models” as having “minimal refusal rates”—meaning models designed to decline fewer requests under their safety controls. The document is version P00003, an update connected to a prototype agreement worth up to two hundred million dollars over two years. It covers testing, evaluating, and refining models for national-security problem sets, but it does not say which requests those models would be expected to handle. The document’s status is the central issue. The Intercept obtained it through a Freedom of Information Act lawsuit while seeking final, executed contracts, not drafts. None of the released records was marked as a draft. A separate record, signed by OpenAI and the government on February sixth, indicated agreement to the contents of the expanded P00003 version—but that record alone does not prove the minimal-refusal language was executed. A Justice Department attorney first confirmed that the released document was signed and final, then withdrew that statement and said it was not final, promising a corrected record. OpenAI says it rejected the wording in an earlier Defense Department proposal, and that the signed agreement excluded it. The Pentagon has not yet explained how the document was cleared for release. The next decisive fact is whether the promised corrected government record resolves which version was actually executed.

Story brief

3 key points

A FOIA release tied to a Pentagon prototype agreement with OpenAI contains language defining national-security “Mission Models” as having “minimal refusal rates,” but both parties say that wording was removed before signing. The document concerns version P00003 of a potential $200 million, two-year deal. A Justice Department lawyer first confirmed the record was executed, then withdrew that statement, leaving the...

  1. 01

    Version P00003 covers testing, evaluating, and refining models for national-security problem sets.

  2. 02

    The prototype agreement was valued at up to $200 million over two years.

  3. 03

    OpenAI says it rejected the minimal-refusal wording and that the executed agreement excluded it.

A Pentagon–OpenAI contract document released through a Freedom of Information Act lawsuit defines national-security “Mission Models” as having “minimal refusal rates.” OpenAI and the Pentagon say the phrase was not in the executed agreement, while a Justice Department attorney later withdrew an initial confirmation that the released document was final.

What the disputed phrase describes

The language appears in version P00003, an update tied to a prototype agreement valued at up to $200 million over two years. It describes OpenAI’s required work as testing, evaluating and refining Mission Models for national-security problem sets, then defines those models as designed for such use cases with minimal refusal rates.

A refusal rate describes how often an AI model declines a request under its safety controls. The released definition describes specialized models intended to turn down requests less often, but the document does not identify the national-security requests at issue.

Heavily redacted Department of Defense contract document with disputed language highlighted.
The released document defines Mission Models as having “minimal refusal rates.” Source: theintercept.com.

Why the record’s status matters

The Intercept sought final, executed contracts and explicitly excluded drafts; none of the released documents was marked as a draft. A Justice Department attorney initially confirmed that the document was signed and executed, then withdrew that confirmation. Days later, the attorney said it was not final and promised a corrected document without a definite timeline.

OpenAI spokesperson Nate Evans gave an account of the negotiations: the company had rejected the wording in an earlier Defense Department proposal, the department agreed to remove it, and the executed agreement excluded it. OpenAI did not provide the full contract to The Intercept.

An explanation still outstanding

Pentagon official Trevor Tiedeman said he would account for how the document was flagged by FOIA officers, cleared for release and confirmed by government lawyers. According to The Intercept, he later stopped responding to its inquiries. That leaves the promised corrected record as the next document needed to resolve the conflicting accounts.

Sources

  1. theintercept.comThe Pentagon Asked OpenAI for Artificial Intelligence Designed to Rarely Say No

Loading discussion...