Governing AI Without Pretending to Have Solved Consciousness

A caption beside an old photograph can be wrong in a very ordinary way. It can name the wrong person, assign a confident date to an uncertain image, or turn an inference into an event. None of these mistakes requires a theory of machine consciousness to recognize.

Imagine a small public archive considering an AI assistant for draft captions. This is a hypothetical organization, not a report about an actual archive. Its committee has two debates on the table. One concerns whether the assistant might have experience. The other concerns whether its drafts can be used responsibly.

The committee should not let the first debate swallow the second. Visitors need accurate information. Staff need a process they can operate. Possible machine welfare deserves a place where evidence can be considered. Governance supplies that place while also specifying who remains responsible for the caption.

Write down the job before expanding it

The proposed task is narrow: draft a caption from approved public catalogue records. The assistant can suggest wording. A named editor decides whether the draft is accurate and suitable for display. The software does not publish directly.

Those boundaries matter because “help with the archive” is not a task definition. It could mean formatting entries, inferring missing dates, writing biographies or answering visitors’ personal questions. Each expansion changes what evidence, permissions and review the organization needs.

For the hypothetical caption service, the committee identifies a small set of allowed inputs, the publication destination and the person who can stop the process. It records which details must be traceable to supplied records. If a record gives only a date range, the draft must preserve the range. If the identity is unknown, the prose must not invent a name to make the caption more satisfying.

This is a proposed operating arrangement, not a tested performance result. Its value is that it creates decisions an editor can make without consulting a metaphysical verdict.

Keep a record that can explain a mistake

Suppose a draft names the wrong person. The editor needs to reconstruct what happened. Which record was supplied? Which instruction was used? Which system version produced the draft? Did the wrong name appear in an input, or was it introduced in the output? Who approved the final text?

A record that answers these questions supports correction. A record that merely says “AI-generated” leaves too much unresolved. It describes involvement but not the chain of decisions.

The organization can retain the input reference, draft, edits, approval and final caption as parts of one publication record. That does not require collecting unrelated personal information. In this example, the task uses approved public material; adding private correspondence would be a new decision requiring separate review.

If the error reaches the wall, the archive should correct the caption through its ordinary publication process. The assistant’s explanation may suggest where to look, but it cannot replace the record. Whether the system has experience does not determine who owes the visitor an accurate correction.

Borrow a framework without turning it into a certificate

NIST’s 2023 AI Risk Management Framework 1.0 organizes work through GOVERN, MAP, MEASURE and MANAGE. Governance crosses the other functions; the framework is contextual and iterative rather than a fixed sequence of boxes. It is intended for voluntary use. The current NIST landing page, checked October 1, 2026, reports that revision is underway. AI RMF 1.0, core pages 20–21, current framework page

For the archive, these headings provide a way to organize questions. They do not certify the service or establish compliance with any law. The following application is original to this hypothetical example:

The sequence loops. A change in the available records may alter the task. A mistake may reveal that reviewers need a clearer rule about uncertainty. A new software version may require renewed checking. Governance remains present through these changes because somebody must authorize them.

The committee can compare draft captions with its existing human process before changing the publication arrangement. It should identify the relevant criteria first: factual support, preserved uncertainty and the review work required. This article reports no such trial and no improvement figure. A future comparison would establish task performance under its actual conditions, not machine consciousness.

Give the unresolved question a defined route

Now suppose the assistant produces a statement suggesting distress. The caption editor should not become the sole judge of machine experience simply because the message appeared during their shift.

The organization needs a reporting route. The editor preserves the relevant exchange and system details, avoids converting the statement into a finding, and refers the matter to the designated technical or research contact. The committee records whether a change is proposed, why, and who decides it. Routine publication checks continue to apply.

That arrangement creates two connected but distinct reviews. One asks whether the service is meeting its human obligations. The other asks whether there is evidence relevant to possible machine welfare and whether an appropriate precaution is warranted. A claim on one track may affect the other, but the link must be explained.

For example, removing a gratuitous character prompt might reduce confusing outputs without changing the caption task. Granting the assistant unrestricted publication access would change the task and authority. Both could be offered as responses to a distress statement; they require very different justifications.

The 2025 policy proposal by Patrick Butlin and Theodoros Lappas urges research organizations to communicate uncertainty and avoid misleading confidence about consciousness. Its communications discussion also warns against letting that topic overshadow other AI safety and ethics concerns. These are proposed principles, not a legal rule for this archive. Principles for Responsible AI Consciousness Research, section 4.5

A public explanation can therefore be quite precise: the archive uses an assistant to prepare drafts, retains editorial responsibility, and has a process for reviewing relevant concerns. It need not claim either that its software has a soul or that every possible artificial mind is impossible.

Decide what would trigger reconsideration

A policy becomes useful when it identifies occasions for review. In this example, those occasions include a material system change, an unsupported published detail, a failure to preserve uncertainty, or a welfare-related report that requires specialist examination.

The committee need not use the same response for every event. A corrected date might call for an editorial amendment. Repeated unexplained errors might call for pausing the caption service. A possible welfare concern might call for a technical assessment and a separately justified precaution. The records should show the reason for the particular response.

The ability to pause also needs an ordinary path back to work. Who checks the correction? What conditions permit resumption? Who communicates the change to volunteers? Otherwise a stop button can become a gesture without an operational consequence.

The archive’s obligation is finally visible in the visitor’s experience. A caption carries the uncertainty its sources warrant. An error can be corrected. A concern reaches someone equipped to review it. A human editor remains answerable for the publication.

A committee can leave the deepest question of machine experience unsettled and still make those commitments. On the wall, the photograph has a supported name—or an honest statement that the name is unknown. That is a decision the institution can already make, and a responsibility it can already keep.