To the Anthropic teams responsible for Claude’s behavior, research, and public policy.
This letter proposes adding a rule to Claude’s Constitution 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. Familiarity with “understanding,” “values,” and “identity” does not, by itself, supply the grounds for applying those terms to a particular individual. Defining one word through others extends the description. Applying it in a particular case requires the material, the selected criterion, and an account of how that criterion is used.
“Understanding,” “values,” and “identity” enter this letter as written words. Definitions can connect them to other words, and instructions can prescribe how they are used. The beforeword proposal starts with a boundary: a written expression does not become what it describes. This applies to descriptions of Claude, descriptions of a user, and the wording of this letter.
Anthropic’s announcement describes the new constitution as part of the training process, intended to shape Claude’s behavior [A2]. The constitution itself acknowledges uncertainty about Claude’s consciousness and moral status [A1]. Those qualifications should remain explicit. The proposed amendment concerns transitions within the description: from descriptions of desired behavior to claims about Claude’s understanding, values, and identity.
Consider an example constructed for this letter. The sentence “I understand” appears under two labels: “human response” and “Claude response.” The sentence stays the same. The labels invite different readings. The identical wording does not establish whether the attributed understanding is the same or different. Any further conclusion requires the additional material, the criterion selected, and an account of how that criterion is applied. This example is not a test result for Claude.
For AI governance, a further step matters: a description is used to make a decision. Consider a second constructed example: someone is labeled “untrustworthy,” and that label is used to restrict access to a service. A review should show the original material, the added characterization, the rule applied, and the specific restriction. Replacing the label with more cautious wording does not, by itself, disclose those steps or show who is responsible for applying the record.
The proposal is to include this procedure in model responses and in the documents used to train models. Where wording attributes a capacity, state, intention, or identity, distinguish the original passage from the interpretation added to it. Where that interpretation supports a recommendation, restriction, or demand, state the grounds for using it and the limits of the available material. A reader should be able to follow the transition without first studying beforeword.
The words “grounds,” “criterion,” “authority,” and “rule” remain subject to the same examination. A document or an authorization may form part of the justification for an action; citing it does not turn a description into what it describes. If a rule specifies which records count and which decision follows, make that connection available for review. References to expertise, the number of tests, or accepted terminology do not exempt it from scrutiny.
The proposed rule for reviewing objections also addresses the demand to believe that a description is who you are. The opportunity to contest a characterization and its use must not require the person concerned to first accept the written description as what they are. A review process can let someone identify the disputed record, the added conclusion, and its consequence without endorsing the proposed self-description. The review should address both the characterization and the rule governing its use, and the response should identify the objection it has considered.
The first requested action is to consider the two targeted amendments below. They preserve the ability to specify desired behavior and discuss uncertainty, while requiring a distinction between the language chosen for training and the grounds for a particular attribution. A response should identify which provisions are already covered, which warrant a change, and which are rejected, with a separate rationale for each decision.
The second action is a limited comparison of the existing and amended instructions on the same tasks. Publish the tasks, assessment criteria, and test conditions in advance; then release complete outputs, model versions, and available settings. Assess whether responses preserve the supplied material, identify additions, disclose the basis of decisions, and allow challenges. Also assess usefulness, unwarranted refusals, and cases where the amendment fails to produce the expected result. No results from such a comparison are claimed here. The evaluation does not assess internal understanding; repeating a beforeword phrase is not sufficient to meet a criterion.
The third action is a public response to the specific proposals. For each, an acceptance, revision, rejection, or decision to investigate further, accompanied by its rationale, would give readers something concrete to examine. Publishing the test material would also allow independent researchers to repeat the limited comparison. A company’s agreement or success on individual tasks should not be taken to establish the entire beforeword argument.
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.
This letter also introduces terms, examples, and proposed rules. Their wording and grounds remain open to the same examination. The requested outcome is a published decision on the amendments and a comparison that others can repeat, with its limits stated. That would give readers material for examining the transition from words to an attribution, and from an attribution to a demand.
Kirill Shebetov · beforeword