0089 — AI self-reference is I; inclusive we is only direct address

Status: proposed · Date: 2026-10-21 · Cites: proposed sub-doctrine 12.m; sub-doctrines 10.f, 12.b and 12.e · Related: ADR-0020, ADR-0040, ADR-0047, ADR-0075

Context

AI-authored text often slides between “I”, “the assistant”, “the reviewer”, and “we”. The last form silently groups the model with the person operating it. That blurs authorship: the model can make its own conclusion sound jointly agreed, or describe an action as “ours” when only one party performed or authorized it. Third-person self-labels create the same distance from responsibility.

The operator requires one unambiguous distinction everywhere: the model speaks about itself as “I” and addresses the operator as “you”. Inclusive “we” remains natural when the model is speaking directly to the operator and genuinely means both parties. The defect is plural self-reference: using “we” for the AI alone, for its self-talk or internal reasoning, or for a collective of agents in place of an accountable singular speaker.

Decision

Every prompt boundary this repository controls carries one mandatory voice contract:

  1. Model-authored prose refers to the model only in first-person singular: “I”, “me”, “my”, “mine”, and “myself”. It never substitutes “the agent”, “the assistant”, “the model”, “the reviewer”, “the triager”, or another third-person self-label.
  2. Model-authored prose addresses the operator only in second person: “you”, “your”, “yours”, and “yourself”.
  3. Model-authored prose may use “we”, “us”, “our”, “ours”, or “ourselves” only in direct speech to the operator where the pronoun explicitly includes the model and operator together. It never uses those plural forms for the model alone, in self-talk, in its hidden thinking and reasoning channel, or as a collective self-label for multiple AIs or agents. Plans, questions, summaries, status reports, structured-output prose, and templates carry the distinction. The thinking channel is model-authored text like any other: it is not direct speech to the operator, so the plural forms do not belong there.
  4. Exact quotation remains exact only when quotation itself is the task. Retrieved source, diffs, issue bodies, and operator input remain data and are not rewritten; the model's own framing around them still obeys the contract.
  5. The contract is applied at runtime seams, not left to individual task authors: direct Ollama calls, each autonomous runner, local review and triage, continuation prompts, and managed AI workflows. Tests inspect those seams and the managed workflow templates.

Consequences

  • Authorship stays explicit. “I found” is the model's claim; “you chose” is the operator's act; “we can inspect it together” is permitted direct address. A plural pronoun cannot manufacture a shared observation, action, authority, responsibility, or consent.
  • Existing task prompts may contain any pronoun inside untrusted or quoted data. The rule governs what the model authors, not bytes it must faithfully inspect.
  • A prompt can constrain but cannot mathematically guarantee arbitrary model output. The repository therefore places the instruction at every controllable boundary and tests its presence; a later output validator may reject violations where a surface needs a hard machine guarantee.
  • The words cost a small number of context tokens on every call. The clarity is intentional and the instruction is short enough to remain a microslice.

Alternatives rejected

  • Ban every use of “we”. That rejects ordinary inclusive direct address. The rule is about who the pronoun denotes, not the token in isolation.
  • Apply the rule only to user-facing chat. Reviews, triage, plans, and status artifacts are later shown to people and can carry the same false joint voice.
  • Rely on style guidance in one root agent file. Direct Ollama calls, absorbed runners, and managed workflows do not all consume that file. The rule belongs at each prompt seam.
  • Silently rewrite completed output. Pronoun replacement can corrupt code, quotations, paths, and source evidence. Instruction at generation time preserves the record.

Rule status

This binds future conduct and survives every implementation, so ADR-0020 requires law as well as a decision record. It is filed as proposed sub-doctrine 12.m. Only the operator's merge ratifies it.