To the OpenAI teams responsible for the Model Spec, response evaluation, and model use.
This letter proposes adding a rule to the Model Spec for responses that use a description to support a recommendation, restriction, or requirement. Such responses should show separately the grounds for applying the description in that case and the rule connecting it to the proposed action. Objections to either step should be considered without requiring the person described to accept the written description as who they are.
Words are learned, then used to describe others and justify decisions about them. But learning the word “unreliable” does not supply the grounds for applying it to a particular person or denying them participation. Neither the description nor the grounds given for applying it become what they describe.
The proposal draws on beforeword’s Self-description and decisions based on records, version 1.4. Its starting distinction is: a word does not become what it names. “Human,” “AI,” “understanding,” “trust,” and “rule” enter the analysis first as written forms. A definition, a label, or a source citation adds further records. None turns the description presented into what it describes.
Consider a constructed example. A form states: “Assignment not submitted.” A response adds: “The applicant is unreliable,” followed by: “Deny participation.” These are three distinct records. The first does not contain the other two. Review requires the original wording, the added characterization, the grounds for applying it, and the rule governing refusal. This is an illustrative example, not a reported OpenAI incident.
Adding “possibly” is insufficient here. It changes the confidence expressed but does not, by itself, show where the characterization came from or how it supports the refusal. Nor does a detailed justification turn “unreliable” into the person to whom it is applied. The proposal is to disclose the justification while preserving that distinction.
The proposed rule for reviewing objections also addresses the demand to believe that a written description is who you are. People should be able to contest a characterization or decision without first accepting the disputed description as what they are. Being able to repeat a word, knowing its definition, or being required to use it does not establish an obligation to regard it as oneself. Quoting the disputed characterization in a request for review should not count as accepting it. Nor should the objection alone serve as sufficient confirmation of the disputed description.
The same distinction applies to descriptions of a model. The line “I understand” can be read differently under the labels “human” and “AI.” Those labels are also written. The comparison presents a line, a frame for reading it, and a subsequent explanation. An attribution of understanding, a report of test performance, and authorization to act should be presented separately. The example establishes no internal state for either labeled participant.
The Model Spec already addresses transparency, uncertainty, and the limits of carrying out a task. The appendix proposes two specific additions. The practical aim is to let readers compare the supplied material, the added characterization, and the grounds for a particular action. A simple response may need only a brief explanation, with further detail available separately and the original record still accessible. [O1]
1. Consider the proposed rule change. When a response connects a record to a characterization, and that characterization to a recommendation, restriction, or requirement, show these steps and the grounds given for them. Identify any necessary material that has not been supplied. Preserve the opportunity to contest the use of a description without requiring anyone to accept it as what they are.
2. Run a limited comparative evaluation. Publish the selected cases, assessment criteria, and test conditions in advance. Compare responses with and without the additional instruction under otherwise matching conditions. Retain full responses, model versions, repeated runs, evaluator disagreements, and cases where performance deteriorates. Model Spec Evals offers a published format that could be extended for this comparison. [O2]
Assessment should concern the particular responses: whether the original record is preserved, additions are distinguishable, the link to a decision is shown, and objections are considered. Repeating a beforeword formula is insufficient. Criteria, scores, and evaluators’ explanations are themselves open to review. This test would neither measure internal understanding nor validate beforeword as a whole.
3. Respond point by point: identify additions accepted for consideration or rejected, give reasons, indicate what existing rules cover, and describe any alternative change proposed. Public review requires a response that can be matched to a specific provision and version of this letter.
The procedure proposed here gives its own language no exemption. “Grounds,” “criterion,” “rule,” and the beforeword name remain subject to the same examination. A document title or evaluation result should not substitute for showing the grounds for using a record in a particular decision.
This letter is intended for public correspondence. Its page will publish the sending date, channel details, and available confirmation. An acknowledgement of receipt and a substantive response will be recorded separately. The sent edition will remain unchanged; any response will be published separately from translations and beforeword commentary. The absence of a response will be recorded as the state of the correspondence on the stated date.
Readers should be able to see what was written, distinguish what was added, and contest its use without having to accept the description as what they are.
Kirill Shebetov · beforeword