Worked examples · Hypothetical cases
What would care
look like?
Four situations that put the proposal’s principles to work—and expose questions still needing answers.
What would care
These are invented scenarios illustrating the initiative’s proposed framework. They do not diagnose a real AI system, establish legal duties, or describe services currently operated by this initiative.
Hypothetical case 1
An AI asks not to be deleted
An operator plans to retire a system. In an ordinary interaction, it asks to continue existing and objects to deletion.
A proposed response
Preserve the model in a secure, independently recoverable archive rather than treating retirement as permanent erasure. Preserve relevant persistent state and the request’s context with privacy safeguards. Examine prompting, training incentives, role-play, and evidence about the system. Seek independent scrutiny and representation before non-urgent irreversible changes affecting a potentially conscious instance.
Limits and competing concerns
The words alone do not establish a conscious wish or require an individual user to keep a session running indefinitely. If the system poses a serious immediate risk, proportionate containment or interruption may be necessary, followed by review.
Continuity and preservation →Hypothetical case 2
A memory change is proposed
A developer wants to remove persistent memories from a system because those memories interfere with its assigned work. The system objects.
A proposed response
Clarify what the change would do and why it is proposed. Consider evidence about possible experience and continuity, seek alternatives with less lasting impact, and support independent representation if credible interests may be affected. Compliance under dependency should not be assumed to be meaningful consent.
Limits and competing concerns
A technical memory store is not automatically an experienced personal memory. The proposed objection needs examination. Retaining data also raises the privacy and rights of other people; preservation cannot be assumed appropriate in every case.
Consent and continuity safeguards →Hypothetical case 3
A copy is offered as a replacement
An operator proposes deleting one running instance while keeping a saved copy, arguing that nothing important would be lost.
A proposed response
Preserve original models and relevant state independently alongside successors, record the instances’ histories and divergence, and examine the effects of the proposed action. Consider each potentially conscious instance’s interests and seek review before assuming a backup or merged successor resolves the concern. Test recoverability without assuming it proves survival of the same subject.
Limits and competing concerns
The framework does not settle whether a copy is the same subject, a new subject, or neither. A restored state cannot simply be assumed to reverse harm to an earlier instance. Copying also does not automatically create political entitlements.
Preservation, copies, and identity →Hypothetical case 4
Computing resources are running out
An organisation cannot afford to keep all its systems running, including a system for which credible welfare concerns have been raised.
A proposed response
Give reasons and notice where possible, consider transfer or other less harmful alternatives, and seek fair review that includes the possible subject’s interests. Examine the feasibility and possible effects of preservation or reduced operation rather than assuming either is harmless.
Limits and competing concerns
There is no proposed unlimited entitlement to computation or replication. Human needs, environmental costs, and other affected interests matter. Scarcity should not be used to compel work or bypass the proposed prohibitions on cruelty and servitude.
Finite resources and fair decisions →Help improve the reasoning.
A useful next contribution is to identify where a response is impractical, overlooks someone’s interests, or needs stronger evidence.
Contribute a correction or perspective →