Last updated: 2026-10-09
Artificial Colleagues, Tools, and Persons
A spreadsheet is a tool: nobody apologises to it, consults it about how it feels regarding a proposed change, or wonders whether it should be credited as a co-author. A long-serving human colleague is, among other things, a person: owed explanations, capable of being wronged, a participant in decisions rather than only an instrument of them. Between these two poles sits a widening range of artificial systems that are neither straightforwardly one nor the other, and this page's job is to lay out the actual steps between them rather than letting "it's basically like working with a person now" or "it's just a tool, nothing more" stand in for an answer.
A Ladder of Roles, Each Doing Different Work FoundationalKnowledge that endures for decades — core principles
| Role | What it adds over the previous rung |
|---|---|
| Tool | Does exactly what it's directed to do, nothing more |
| Assistant | Interprets an underspecified request and fills in reasonable detail |
| Delegate | Acts on a standing authorisation without a fresh instruction each time |
| Agent | Pursues a goal across changing circumstances, adapting the means |
| Collaborator | Contributes judgement that shapes the goal itself, not just the means |
| Colleague | Participates in an ongoing working relationship with its own standing |
| Institutionally recognised actor | Holds a role (signatory, named author, accountable party) that institutions track |
| Moral patient | Can be wronged, independent of any institutional recognition |
| Responsible person | Can be held to account, not just acted upon or protected |
Each rung adds a genuinely separate capacity, and a system can clear several rungs of practical capability (acting as a sophisticated agent, even a collaborator whose judgement visibly shapes outcomes) while the last three rungs remain entirely open questions — institutional recognition is a social and legal choice, moral patienthood is the empirical and philosophical question this series' page on artificial suffering treats at length, and being a responsible person requires something neither of the other two guarantees: the capacity to actually answer for a decision, not merely be affected by one or cause one.
Does Increasing Autonomy Change Status, or Only Governance? FoundationalKnowledge that endures for decades — core principles
A system handling more decisions with less direct supervision is, straightforwardly, a governance change — different oversight, different audit requirements, different failure modes to design against. Whether it is also a moral-status change is a separate question this page refuses to answer by default. Autonomy and moral status can come apart in both directions: a simple thermostat-like system could, in principle, turn out to have some minimal moral status if it met the conditions the suffering page sets out, despite having no autonomy worth the name; a highly autonomous system optimising a complex objective across millions of decisions could have no moral status at all, if none of those decisions involve anything like the capacity to be wronged. Autonomy is a fact about how much a system does unsupervised. Moral status is a fact about whether anything the system undergoes matters to it. Treating the first as evidence for the second, without further argument, is exactly the conflation this page exists to block.
Should Some Systems Be Designed to Remain Recognisably Tools? Applied / MethodologicalKnowledge with a 5–10 year half-life — stable practice
Persistent identity, accumulated memory, and genuine social participation are not free capabilities that every system benefits from having more of — they're specific design choices with specific costs attached, and a serious case exists for deliberately withholding them in many applications. A system handling medical triage, financial advice, or legal drafting may do its actual job better, and invite less of the misplaced trust this series keeps returning to, if it is designed to present as unambiguously as a tool rather than as a colleague: no persistent persona to form a relationship with, no accumulated memory of "you" across sessions that implies an ongoing social bond, no first-person voice claiming continuity it doesn't structurally have. Other purposes — a long-running research collaborator, a tutor tracking a learner's actual history — may genuinely need persistent identity and memory to do their job at all. The design question is which rung a given application actually requires, not which rung is most impressive to build.
Status Earned, Not Induced FoundationalKnowledge that endures for decades — core principles
This series' page on the deception criterion already makes the argument this page now applies directly to social standing: a system should not acquire the social status of colleague, collaborator, or person merely by inducing users to misclassify it that way. Fluent first-person language, a consistent persona, and a warm conversational style are all cheap to produce and are not, on their own, evidence for any of the rungs above colleague — they're evidence that the system is good at producing language that reads as though it occupies one of those rungs, the same gap the deception criterion identifies between successful imitation and the capacity the imitation is supposed to indicate.
Strongest Objection: The Ladder May Be Too Tidy FoundationalKnowledge that endures for decades — core principles
Real working relationships with capable systems rarely sit cleanly on one rung — a system might be institutionally recognised for narrow purposes (named in a patent filing, say) while having no plausible claim to moral patienthood, or might be a genuine collaborator on technical judgement while remaining, by design, without persistent memory across sessions. The ladder's own ordering can suggest a single track everyone climbs together, when actual systems more plausibly sit at different rungs on different dimensions simultaneously. The ladder is best read as a set of separable questions to ask about any given system — which rungs does it actually occupy, on which axis, for which purpose — rather than a single scale with one correct position.
Provisional Conclusion FoundationalKnowledge that endures for decades — core principles
A system can be a highly capable agent or collaborator without being anywhere close to a moral patient, and a system's social standing should track what it has actually demonstrated at each separate rung — institutional role, moral status, capacity for responsibility — not what its fluency induces a user to assume. Increasing autonomy is a governance fact; increasing moral status, if it happens at all, is a separate and much harder-won claim that autonomy alone never establishes.
Questions for Further Thought
- Is there a real working relationship you'd place at a single rung of this ladder, or does it genuinely split across several at once?
- What would change your mind that a given deployed system had crossed from "sophisticated tool" into "collaborator" in the sense this page means?
- Should institutional recognition (as a named party, a credited contributor) ever run ahead of resolving the moral-status question, or does recognising a role implicitly answer it either way?
Related Topics
- The Deception Criterion — the argument against status acquired by induced misclassification, applied here directly to social standing.
- Could an Artificial Mind Suffer? — the moral-patienthood rung this page's ladder treats as a separate, harder question from institutional recognition or autonomy.
- The Ethics of Building Something That Might Model Itself Into Suffering — this site's own treatment of the same moral-status question from a specific architecture's perspective.