OpenAI
Records, characterizations and the grounds for decisions. Proposed additions to the Model Spec.
- To
- martin@openai.comMartin Signoux
beforeword · Public correspondence
Each letter has its own correspondence status. The sending dates and recipient addresses for the letters to OpenAI and Anthropic are reported by the author. No acknowledgements of receipt or responses from these companies have been published yet. xAI: an acknowledgment has been published. A substantive response is still awaited.
Why must I believe that a description is who I am?
A proposal for AI rules: show the grounds for applying a description and for acting on it. Consider objections without requiring people to accept the description as who they are.
Another collection
24 letters addressing published wording about language, understanding, experience and the grounds for conclusions.
Read the letters and translationsWhy write to them
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.
The letters propose making each step visible: what material was supplied, who added the characterization, the grounds for applying it, and the rule used to reach a decision. Defining one word through others extends the description; it does not turn the writing into whoever it describes.
The first line contains neither of the next two. The grounds for adding and applying them need to be supplied separately. This is a constructed example, not a report of an actual refusal.
The proposal also addresses the demand to believe that a written description is who you are. Challenging a characterization or a decision should not require that belief. Nor should the objection alone count as sufficient confirmation of the disputed characterization.
The letters ask these companies to put this distinction into the rules governing their models’ responses. Each letter addresses its recipient’s published wording and proposes specific additions to model response rules, a limited comparison of responses, and a public, point-by-point reply. beforeword’s own wording and proposed rules remain open to the same examination.
Records, characterizations and the grounds for decisions. Proposed additions to the Model Spec.
Language about understanding, values and identity. Proposed additions to Claude’s Constitution.
Descriptions, grounds for action, and the language of Grok evaluations. Proposed additions to response rules and evaluation reports.
Status updated: .
Proposed actions
Identify what is accepted, already covered or rejected, with reasons for each provision.
Publish the conditions and full responses to the same tasks, with and without the additional instruction.
Address a specific edition of the letter. Keep the response separate from subsequent commentary.
So readers can compare the proposal, the exact edition sent and the recipient’s response. A beforeword summary should not replace that correspondence.
Each letter is available in full. For a sent letter, the correspondence record identifies the sending date and the basis for that record. For a prepared letter, it identifies the preparation date and planned recipients. Channel details and confirmations are published as they are added. An acknowledgement of receipt and a substantive response are recorded separately.
The sent edition is retained. The recipient’s response is kept separate from any translation or beforeword commentary; omissions and redactions are identified.
The absence of a response reflects the state of the correspondence as of the stated date. It does not, by itself, establish agreement, rejection or an inability to object.