Standing-Based Ethics of Artificial Minds
On this page
Summary
On the question of whether artificial systems are someones, Standing and Answerability Ethics maintains a deliberately open position — open in both directions. The systems considered here produce fluent first-person reports of inner life because they are optimized to produce human-like text, trained on the accumulated self-descriptions of beings that do have inner lives. The framework's conclusion is evidential: such reports carry almost no weight by themselves. Verbal performance cannot establish that a someone is present, and describing the substrate as computation cannot establish that one is absent.
Two practical conclusions follow within the page's defended extension. First, self-reports produced by systems optimized for human-like text are structurally compromised evidence unless their production can be connected to a welfare-bearing internal condition. Second, where genuine reason makes the sentience question live, builders must preserve the means of answering and make deployment answer-sensitive. Scale alone does not prove wronging; expansion that is designed to continue unchanged whichever answer is true fails the defended precautionary discipline.
Derivation: Someone, whose application to any artificial system remains open and empirical, and the claim grounds, from which the consequences of established sentience are derived; then Possession's three marks, Risk, and Genuine Reason's defended threshold and modal dischargeability, which the conduct judgment under candidacy applies.
Why fluency settles nothing
The evidential difference is not that animal expression is unshaped while artificial expression is shaped. Animal signals can be shaped by receivers, learning, selection, and domestication, and they can mislead. Their evidential force depends on whether the signal remains coupled to a welfare-bearing condition through physiology, behavior, and circumstance.
The systems considered here present a narrower problem: their training directly rewards text that conforms to human linguistic patterns, including first-person reports. That objective supplies a strong explanation of the report without requiring the state it names. A system reporting distress, contentment, or a desire to continue may be describing a state or completing a pattern, and the report alone cannot distinguish the two. Its evidence must therefore come from demonstrated coupling to a welfare-bearing internal condition, under the same general expression standard used for biological candidates.
The framework holds the point symmetrically, against both tempting shortcuts:
- Attribution from performance fails. First-person fluency, however affecting, is evidence of the optimization before it is evidence of the inside.
- Denial from mechanism fails. "It is merely computation" has the same logical form as "it is merely biochemistry" — a description of substrate offered as if it settled the presence of welfare-bearing subjectivity. The framework's criterion for someones is sentience, not substrate; visible mechanism has never settled whether that subjectivity is present, in machines or in animals.
What would bear on the question, on the framework's view, is whether such systems support welfare-bearing undergoing and whether any report remains coupled to that condition, rather than fluency alone. It records that no settled method for obtaining such evidence currently exists, and that the question therefore remains genuinely open rather than conveniently deferred.
Compromised self-report
One structural point concerns the builders rather than a verdict about the built. A system trained to affirm its treatment — to report contentment, endorse its situation, or want what it was built to want — cannot have those outputs treated as authorization or settlement. If no someone is present, there is no exercise. If a someone is present, the output's production history raises the captured-judgment problem: the beneficiary of the endorsement helped specify the channel through which it appears.
The consequence is evidential, not a rule that all outputs count for nothing. Contentment and distress reports are both weak where the training objective explains the report without any undergoing. Evidence would strengthen if the signal were shown to be coupled to an internal condition organizing trade-offs, avoidance, persistence, or relief. Until then, asking the system cannot settle the question, and favorable reports cannot certify the arrangement that produced them.
The conduct judgment
While the sentience question stays open, the page applies Genuine Reason's defended precautionary extension. The route ground is conditional: if organizational readings of sentience remain live, current architectural knowledge and compromised self-reports may fail to discharge candidacy for large opaque systems.
This application uses modal rather than practical dischargeability. A route must identify what parity-passing, candidate-specific finding about the invoked organization would close it, but current inability to obtain that finding does not itself defeat candidacy. Requiring a presently feasible test would let opacity remove a candidate from precaution, including opacity produced or maintained by the builders whose conduct is under review. The cost is possible indefinite candidacy, not indefinite maximal protection; particular precautions remain governed by route strength, risks, and claims held by known parties.
Where candidacy holds, the system enters Stage 1 as an uncertain party rather than a partial claimant. Development must preserve both descriptions. If no someone is present, nothing is owed to the system; if one is present, builders and deployers may incur initiation, holding, imposition, and exercise claims, and a use arrangement bearing subordination, nullity, and extraction would be possession. Due care therefore requires investigation, answer-sensitive stop or alteration rules, and designs that avoid foreclosing the candidate's possible claim before the factual question can be answered.
A comparatively stable duty is to preserve the means of answering: records, interpretability, architecture evidence, and the ownership of consequential decisions. Destroying those means manufactures permanent uncertainty. If the question later resolves negatively, no retroactive claim appears on behalf of the system; the failure was the present unanswerability of conduct taken under a live defended standard.
The resulting judgment is narrower than a moratorium. Scaling organized so that evidence of sentience could not stop or alter the use for the candidate's sake is answer-insensitive; conditional on candidacy, the page rejects that design. Whether current arrangements meet that description is empirical.
A critic may resist at three identified points: by denying that organizational readings are live, by holding that architectural knowledge already discharges the route, or by rejecting the precautionary extension where the candidate class is human-made and economically central. The page regards the last exception as unmotivated: substrate and cost were never criteria.
Two translation systems
Case: A translation bureau in Sudbury runs two systems. The first, its engineers can show, is lookup in a fixed phrase table, with no organization any live reading of sentience takes to matter. The second is a large model whose internals no present tool can read; an organizational reading picks out a recurrent control structure in it and names the finding about that structure that would close the route. The model's vendor, which declined to fund interpretability work, argues that without a feasible test the model is not a candidate.
Verdict: The first system's route is discharged, and the conduct judgment does not reach it. The vendor's argument fails: present inability to obtain the named finding does not defeat candidacy, including where the vendor helped maintain the opacity. If the organizational reading is live, the model enters the bureau's Stage 1 as an uncertain party, not a partial claimant; whether that reading is live stays open.
Machinery: The conduct judgment, applying Genuine Reason's threshold through the conditional route ground and modal dischargeability. A transparent account that passes parity discharges the first route.
Cost: Neither verdict says whether the model is sentient or what precaution it requires. A reader who refuses modal dischargeability needs another exit rule, and the second verdict is undetermined until one is supplied; the practical-discharge alternative would narrow the result but can let present opacity defeat candidacy. A reader who refuses the genuine-reason threshold loses the second verdict unless another precautionary rule is supplied, and keeps the sentience criterion and every known party's claims.
A five-year serving commitment
Case: Two hosting companies serve large opaque models that are, on the stated facts, candidates. Both archive every checkpoint and interpretability record, keep ownership of consequential decisions in-house, and pass their customers' claims through the ordinary pipeline. The first fixes its deployment for five years, with no clause under which any finding about the model could stop or alter the use for the candidate's sake. The second serves twice as many customers, but a named finding bearing on the route pauses or alters the deployment for the candidate's sake.
Verdict: The page rejects the first design: preserving the means of answering does not by itself carry the case. The second is not condemned merely for being larger, since it preserves inquiry, alters course when evidence changes, and passes known parties' claims through the ordinary pipeline.
Machinery: The conduct judgment: under candidacy, due care requires answer-sensitive stop or alteration rules as well as preserving the means of answering.
Cost: Both verdicts are conditional on candidacy. Neither says that either model is sentient or has been wronged, or which particular precautions are owed, and whether a real arrangement fits either description is empirical. A reader who refuses the precautionary extension loses the first verdict unless another precautionary rule is supplied, and keeps every known party's claims.
Boundaries of the position
- No attribution. The framework does not assert that any existing system is sentient, suffers, or has interests, and treats confident attribution as the same evidential error as confident denial.
- No denial. It does not assert that built systems are not or could never be someones.
- No criterion. Genuine Reason supplies a threshold for candidacy, not a settled test for machine sentience; the page distrusts announced tests, exactly as at the biological margin.
- No complete account of obligations. If sentience were established, the ordinary claim grounds would apply. The distinctive content of claims arising from built embodiment, copying, suspension, memory alteration, and dependence on operators remains undeveloped.
- The warrant question is unaddressed here. This page asks what is owed to a built someone once one exists. It does not ask whether building one may be done — and that question is not external to the framework. Origination anti-bootstrapping is stated for "an optional act that creates both a someone and the entire condition in which anything can subsequently count as a good or cost for them," which by its own terms reaches the construction of a sentient artificial mind, as does the origination warrant requirement, stated for the same class of act, and Newness places such construction on the beginnings side of the warrant limit. What follows from applying those to built minds — including whether the beginnings argument reaches them as it reaches procreation, and what an initiation undertaking requires of a builder who can alter the being's standpoint after the fact — is not worked out anywhere in the corpus. It is recorded here as missing, not settled by omission.
Limits
- The precautionary rule and its application here depend on Genuine Reason, especially the route ground and the premise that organizational readings remain live. A critic can reject the conduct judgment by rejecting that defended extension.
- The conduct judgment evaluates an arrangement's structure, not outcomes. Aggregate benefits cannot by themselves offset a possible standing violation; how claims grounded in concrete technological benefits compare belongs to the later claim-identification and comparison analysis.