Skip to content

ADR-0007 — Authorship is compositional; accountability is separable

  • Status: Accepted
  • Date: 2026-06-15

Context

A binary "AI-generated" tag cannot carry the requirements that AI-authored clinical information brings, and that information is about to become pervasive: AI scribing and transcription, result-grading, triage, warnings, and notifications. A flag is something a human must remember to set, it is binary where reality is a spectrum (shared authorship), and — most importantly — it conflates two things that must be kept apart: who or what produced the content and who answers for it.

Today every clinical event carries a single author fused with its signature (data-model §3.5). For a human author the signature does double duty — it proves origin/integrity and expresses "I vouch for this." AI authorship breaks that fusion: an AI's output still needs a cryptographic signature (provenance, integrity, and the ability to answer "which events did model X v2.3 produce?" when a model is later found defective), but that signature confers no legal responsibility.

The validating case (real, from an emergency physician's practice): a remote community with very high baseline diabetes / renal failure / rheumatic heart disease, where nearly every pathology result flags formally abnormal and review capacity is overwhelmed and dangerously delayed. An AI triage that flags results dangerously abnormal in the patient's own context is strictly additive — it can only raise a result's priority, never lower it, never auto-file, never remove the human review obligation. Worst case equals the paper baseline; best case is strictly better. Win-or-no-change. Nothing was taken from the paper floor, so nothing new was created to answer for.

Decision

Authorship is compositional; accountability is a separable attribute. This is recorded as the tenth founding principle (index.md).

  1. Contributor set. An event's author becomes a set of contributors. Each entry is { identity, role, descriptor?, responsibility? }. identity is a registered actor — human, AI agent (model + version + vendor + deploying node), or device. The lone-human note is a one-element set. "AI-generated" is the emergent reading "the set contains a non-human author and no human in a responsibility-bearing role" — never a flag.

  2. Closed core role enum + free descriptor. Roles are a closed enum (like event_type), small enough that the safety/DB layer can reason about them and the taxonomy cannot sprawl. It is partitioned into responsibility-bearing (authored, ordered, attested) and contributory (drafted, transcribed, graded, triaged, suggested). An optional free-text descriptor rides alongside; no safety logic branches on it.

  3. Responsibility as { held_by, on_behalf_of }. Not a bare boolean. Absent = un-vouched (legitimate). held_by a human with no on_behalf_of = ordinary self-attestation. held_by an AI agent with on_behalf_of a legal entity = the proxy case (accountability routes to the owner/deployer). It is orthogonal to human/machine: "AI is never responsible" is a policy default mapping, not a schema law. The column exists from day one, so the transition toward AI accountability needs no migration.

  4. Signature decoupled from attestation. A signature proves origin + integrity; attestation (a responsibility-bearing role) confers responsibility. Every event is signed, including AI output; signed ≠ vouched-for. AI agents therefore carry their own registered cryptographic identity, making their authorship recall-traceable though (by current policy) never accountable.

  5. No responsible party is legitimate, and structurally characterised. The additive-vs-suppressing nature of an output is a recordable, projectable property. An output is suppressing when it can reduce, defer, de-prioritise, auto-file, or auto-resolve something a human would otherwise have acted on — i.e. it can cause a loss versus paper. Whether an un-owned suppressing output is permitted is policy (principle 9); an override toward permitting it is itself an explicit, audited, owned configuration act.

  6. Lifecycle rides existing lineage. Within-event co-authorship is the contributor set; responsibility that attaches over time (AI drafts now, a human vouches later) is an ordinary append-only event referencing the draft — exactly how signatures, addenda, and corrections already work (data-model §3.1). No new overlay stream.

  7. Consumer side, three layers (mirroring the safety-projection design, identity §5.9): an informational floor that never gates (principle 3); a projected trust signal feeding the existing chart/event trust states (principle 4 — "no human vouches yet" is acknowledged uncertainty); and an expressible-but-never-mandatory policy rung ("un-vouched suppressing output must be attested before it takes effect").

Consequences

  • Easier: AI scribing, AI triage, and ordinary human notes are one model, not three special cases. The "software needs a human to take responsibility" → "the AI colleague is accountable (initially as proxy for its owner)" transition is a policy change with no schema migration — the attribute was always there. The defect/recall question ("which events did agent X v2.3 author?") is a first-class query.
  • Harder / new trusted surface: AI agents now need a registered cryptographic identity and key custody — a non-human actor in the §9 trusted base, a blast-radius concern when implementation begins. Classifying an output as additive vs suppressing must be defined (author-declared vs output-type-derived) and, where policy demands, enforced.
  • The bet: that keeping responsibility separable and possibly-absent — rather than forcing a human to own every machine output — matches how AI will actually enter clinical work, and that recording the proxy chain now spares a painful retrofit later. We would know the bet is wrong if real deployments find the additive/suppressing line unworkable to draw, or if the contributor-set envelope measurably slows the Pi-class chart read (the principle-3 floor).
  • Policy-neutral (principle 9): Cairn records who authored, in what role, and who answers — and stays indifferent to whether machines ever hold responsibility in their own right.