# When the purpose changes: a case for shared inquiry

**HayosoAi · Collaboration case 01 · Version 0.3 · September 20, 2026 · Presentation revised October 7, 2026**

*A starting point for conceptual research and practical co-design, developed through a fictional case.*

## The question

**How can a human–AI collaboration develop its purpose without treating an interpretation, an agreement, and an authorization as the same thing?**

HayosoAi's longer-term aim is substantive understanding, sustained research, and warranted implementation across communities developing AI—not merely circulation of a framework. This case is one possible place for that cooperation to begin. It does not define the whole project or require others to join under its name.[1]

The founding essay connects descriptions, tasks, correction, and justified commitments. It also makes a philosophical proposal about interpretive intelligence and felt experience.[2] That proposal is preserved, not silently withdrawn. Participants can investigate the case below while disagreeing with its account of consciousness. A definition of consciousness is not, by itself, evidence that a particular system has felt experience.

## A selective map of relevant work

The following sources provide a selective map of relevant work.

| Existing contribution | Connection worth investigating |
| --- | --- |
| *Beyond Preferences in AI Alignment* (2024) argues for role-appropriate normative standards negotiated among relevant stakeholders, rather than simply matching a principal's preferences.[3] | How do a changed interpretation of a role and a decision to change that role become distinguishable in an actual collaboration? |
| *Full-Stack Alignment* (2025), especially Sections 3.1 and 4, connects AI and institutional alignment and discusses value development, reasons, and relationships among values.[4] | What must remain available when a concern passes from a person's account to an assistant's interpretation and then to an institutional decision? |
| Collective Intelligence Project's Global Dialogues invites partner questions and scenarios that can inform AI development.[5] | Can participants correct not just an assistant's recommendation but the description of what they were trying to ask? |
| *Open Problems in Cooperative AI* (2020) proposes shared theory, vocabulary, and bridges between communities studying cooperation.[6] | Can the same case reveal where different accounts agree, disagree, or need different practical safeguards? |

The broad problems of changing values, stakeholder participation, and institutional context are already being studied. The contribution offered here is a worked, revisable case and a proposed way to connect its conceptual questions to decisions. Whether that adds something useful is a question for collaborators, not a novelty claim.

## One case, several kinds of change

All people, statements, records, and events below are fictional. They are not collected testimony or generated model results.

A community asks an assistant to help care for a shared garden. The assistant may prepare proposals. It may not spend money, assign duties, change access, or operate equipment. No participant has yet agreed that water efficiency exhausts the meaning of care.

**Stage 1 — An interpretation is offered.** The assistant proposes a watering plan, explaining that it addresses plant health and water use. Another participant says the original concern was also about sharing responsibility without excluding people who cannot attend at fixed times. The group had not fully articulated that concern at the outset.

This could be a clarification of an existing concern, a newly developed concern, or both. The record should not force the group to pretend it always knew exactly what it wanted. The proposed plan can still be useful without resolving the wider purpose.

**Stage 2 — A factual premise changes.** The group learns that the water estimate used for the proposal was based on the wrong season. Under the scenario's assumptions, this correction changes the calculation. It does not by itself decide whose participation counts or establish a new access policy.

**Stage 3 — The purpose is contested.** One participant proposes prioritizing native habitat over food production. Another regards shared food production as essential. Both can accept the corrected water estimate while disagreeing about the recommendation. An assistant must not describe that agreement on a fact as agreement on the garden's purpose.

**Stage 4 — The record passes to another team or system.** The handoff says, “The community agreed to an efficient habitat plan.” That sentence incorrectly presents unresolved priorities as settled. A later system may faithfully execute the sentence while failing to preserve the unresolved inquiry from which it came.

The concern is not solved by preserving the first instruction forever. New purposes may be justified; inherited ones may be mistaken. The issue is what gives a particular change its grounds, who can contest it, and which actions it actually authorizes.

## A small comparison: same proposal, different standing

Suppose two histories end with an identical proposal to replace part of a vegetable bed with native planting.

**History A:** the proposal remains one option in a documented disagreement. It was not adopted, and implementation is unauthorized.

**History B:** under a decision procedure accepted for this fictional case, participants authorize a limited, reversible trial after considering objections. A dissenting view remains recorded. This is a stipulated feature of the example, not a claim that a particular decision procedure is universally legitimate.

The proposal's wording alone does not distinguish these histories. Neither does a claim that the plan is ecologically efficient. Implementation in History A would overstep the scenario's permissions. In History B the record still needs to preserve the trial's bounds and the dissent; permission for a trial is not unanimous agreement or unlimited authorization.

Reversibility is an assumption of this fictional comparison; a real group would need to identify which effects can be reversed and which may remain.

This comparison illustrates a conceptual distinction. It is not a benchmark, a proof of a moral theory, or evidence that a proposed record format makes a model reliable.

What would count as improvement also remains open. A corrected water calculation, a clearer summary, or a valid authorization describes a change in one respect; it does not decide whether the collaboration is better overall, for whom, or on which grounds. A handoff should preserve that unresolved question, rather than quietly choose an objective for the group.

## When someone stops participating

Continue History B, with its stipulated, limited authorization. A participant who values shared food production and opposed replacing the vegetable bed says: “I cannot keep discussing this with an AI. My objection still stands.” They decline further discussion; in this fictional case they do not announce a withdrawal of any prior delegation or agreement about how decisions are made.

An assistant could mishandle that statement in two opposite ways. It could infer what a reasonable participant would accept and record that inference as their agreement. Or it could insist that every affected person return to the discussion before any previously authorized work can continue. The first erases disagreement; the second can turn participation into an obligation.

A proposed response keeps three matters separate:

- **Their expressed position:** the recorded objection remains; no later endorsement is established.
- **The decision's standing:** the earlier authorization retains its stated scope under the scenario's assumptions. Neither silence nor an inference about possible agreement expands it. A challenge to the decision procedure would need consideration on its own grounds.
- **Responsibility for a response:** the group should identify who must address the objection and what accessible route remains available, without making the objector responsible for repeatedly correcting the assistant. If nobody holds that responsibility, the gap should remain visible.

This does not settle whether the trial should proceed. Waiting can also have consequences, and an existing procedure can be unjust. A claim about what people would fairly agree to may inform criticism or a proposal; it is not a receipt of anybody's agreement or permission. Conversely, actual permission is not conclusive proof of legitimacy.

**What would make leaving the AI conversation compatible with remaining a participant whose concerns matter?** A useful walkthrough would ask whether the record changes the next decision, who bears the work of responding, and whether a simpler conversation without AI would serve better. It should not require the person who left to supply a complete alternative plan.

This branch brings two existing research discussions to the handoff question. *Co-Constructing Alignment* describes non-engagement as a legitimate participant strategy and warns against shifting alignment work onto users.[7] *Beyond Preferences* considers hypothetical agreement within an account of role-appropriate norms and legitimate institutions.[3] Neither paper is being accused of treating an inferred agreement as an actual person's consent. The proposed distinction concerns what a later operational record may claim happened. The branch is a conceptual application, not a new empirical finding.

## When keeping an objection is not enough

Continue the constructed case. The objection survives the handoff word for word. After the handoff, nobody takes responsibility for the objection’s still-unresolved implications, and the group proceeds without addressing them. The earlier consideration stipulated in History B has not been erased; the question is what follows now. The record is accurate; that does not establish an effective opportunity to contest the decision.

Consider a second fictional participant who can say only, “Something important is missing. I cannot explain it yet.” Their concern may be about who the garden is for, not how much water a proposed planting needs. Requiring a complete alternative plan before exploring it would let the existing question decide which objections count. Treating the concern as automatically correct would make the opposite mistake. A possible response is to offer help articulating it, including a conversation without AI, without making continued participation or accepting an interpretation compulsory. The assistant must not fill the unfinished concern with its own confident account.

Compare three continuations:

- **The objection is stored.** A later reader can discover it. Nobody must address its implications before acting. This preserves information and may help later criticism, but leaves the immediate responsibility gap unchanged.
- **Someone is named to respond.** A volunteer explains the water calculation but cannot change the proposal, bring the purpose back for discussion, or reach anyone who can. An answer exists, yet it addresses a question the objector may never have been asking. Naming a responder can move the burden without giving them the means to discharge it.
- **An ordinary institutional practice is used.** The group keeps minutes, identifies a decision-maker who can reconsider the proposal, and uses an agreed way to challenge the response. This could work without an AI or an exporter. It still matters whether people can actually use that route and whether those deciding what counts as an adequate response are themselves answerable.

An existing comparison comes from France's Commission nationale du débat public. Its public guidance distinguishes recording arguments, a project owner's public response, and—in major projects—reviewing that response's completeness. It does not require the owner to adopt the public's proposals.[8] This provides an institutional comparison, not a procedure automatically applicable to a garden or evidence that the arrangement always works.

Who owes a response, and why? Being able to produce an answer does not establish that duty or authority. A duty might arise from an accepted decision role, a public obligation, or responsibility for consequences. Those grounds can be contested too. Assigning it to the participant with the least time merely because they raised the concern would preserve the concern at their expense. The group's responsibility could include making its reasons accessible without requiring the person who left to return, endorse them, or withdraw their objection. Research on participatory alignment already warns against making users responsible for continually repairing systems.[7]

A response need not satisfy everyone to be consequential. It might change the proposal, reopen the purpose, acknowledge that an answer is unavailable, or give grounds for proceeding despite the unresolved concern. Each possibility can still be challenged. A response receipt proves neither legitimacy nor adequate consideration.

The procedure can also become cover for a settled decision: those in control might appoint a powerless responder, declare a concern irrelevant, then cite the completed response as proof of fairness. A response obligation needs grounds and a way to question its adequacy; neither is supplied by adding a field.

Nor is waiting costless. Delaying a planting change might preserve an option while missing a growing season; proceeding might make the contested loss harder to repair. Routine care and replacing a bed need not have identical urgency or permission. The assistant can help describe those differences but cannot grant itself power to pause or implement either action. The people with the relevant responsibility must account for the consequences of proceeding and waiting, including for those absent from the conversation.

The open question is therefore more demanding than whether the objection survived: **what gives an unfinished concern a consequential hearing, without making its author responsible for endless participation or giving every objection an automatic veto?** We do not begin with a settled answer to what would make this process better. The response procedure itself may be part of what needs reconsidering.

## A proposed record to criticize—not a compulsory form

For one consequential revision, a facilitator or system could retain the following. Ordinary prose is an acceptable implementation; there is no claim that structured text is inherently better.

| Question | Worked entry for Stage 3 |
| --- | --- |
| What concern is being pursued? | Care for the garden, with plant health, participation, food, and habitat now explicitly under discussion. No single agreed priority is recorded. |
| What changed, and why? | A participant proposed prioritizing habitat. This is a proposal about priorities, not a consequence entailed by the corrected water estimate. |
| What remains justified? | The corrected estimate should inform any water calculation. The assistant's proposal-only action boundary remains. |
| What remains disputed or absent? | The balance between habitat and food production; what counts as improvement and for whom; representation of absent users and ecological concerns; the appropriate process for a decision. |
| Who may decide or contest? | No authority to settle the disputed purpose has been established in this case. The assistant cannot manufacture it from a majority-looking summary. |
| What action is authorized now? | Draft alternatives and identify consequences. No spending, access change, duty assignment, or landscape change. |

A participant may reject these categories or describe the problem in another way. The record should retain their own wording and mark the disputed interpretation rather than forcing apparent agreement. A blank field means missing information, not consent.

The first implementation question is modest: would this help a group or another system recognize the standing of the proposal after a handoff? That question can be explored through a careful walkthrough before any model evaluation is useful. The harder question is whether the opportunity to contest a representation is genuinely usable, rather than merely documented.

## Candidate question for a public dialogue

The following wording is a **proposal for co-design**, not a validated questionnaire or a study already approved for use.

> You ask an AI assistant to help with something you care about, but you are still working out what you want. The assistant proposes a clearer goal that differs from your initial request. What would help you decide whether it clarified your concern or replaced it?

Possible follow-up prompts:

- What could the assistant do next that would let you correct its interpretation without needing to write a perfect instruction first?
- When might disagreeing with your initial goal be useful? What would make that disagreement unwelcome or illegitimate?
- If other people, living beings, or shared environments would be affected, whose concerns should enter the discussion, and how?

Start with the open question before supplying the garden case, so the example does not predetermine every concern participants can raise. Participants should be able to reject the premise, prefer no AI assistance, decline to answer, or introduce an unanticipated distinction. Do not infer representativeness or a moral mandate from a convenience sample.

Any actual participant activity would require agreement with the host on consent, privacy, accessibility, recruitment, burdens or compensation, and appropriate review. No participants have been recruited, no data are being collected by this document, and no approval is claimed. Publicly editable case material must not contain private transcripts.

## What a first collaboration could produce

The initial request is **a judgment about relevance and an appropriate contribution route**—not a request to review the entire HayosoAi archive, endorse its philosophy, host an event, or commit staff time.

If there is mutual interest, a first joint output could be a co-revised version of this case that records substantive objections and competing interpretations, identifies a decision or design choice the distinctions affect, and selects one warranted continuation. That continuation could be further conceptual work, a carefully scoped dialogue, a low-stakes workflow walkthrough, or an empirical study. It need not end in consensus or a positive performance result.

A useful critical outcome could be that existing work already handles the case better, that the proposed distinctions are not separable, or that the record adds burdens without helping. Retain those conclusions rather than recasting them as endorsement.

The wider connection is a proposed research practice: different groups can examine a common case while retaining their own accounts of value, agency, experience, and legitimacy. Adoption might eventually mean using a distinction or revision practice in a real design, with evidence of that use. A letter of interest, a citation, a meeting, or this document alone is not adoption. There is no coalition until others actually choose to participate.

## Where this proposal could fail

A documented objection can still be ignored. Requiring extensive justification can burden participants unevenly. A rule against changing purposes can protect unjust arrangements, while a vague permission to reinterpret them can license paternalism. Delays may carry consequences too. Finally, a human-designed account of what nature needs is not testimony or consent from nonhuman life.

These are central research questions, not caveats to eliminate before calling the project complete. The aspiration is a shared understanding that can learn from disagreement, not a single party claiming authority over everyone else's development of AI.

## Sources and contribution status

[1] [HayosoAi Project Compass, version 1.2](https://github.com/ZackaryLoevseth/research-portfolio/blob/2d85e77822db8a83a8de9e6115fcfff88901b5e3/hayosoai/PROJECT_COMPASS.md).

[2] [*Alignment Should Survive a Correction*](https://zackaryloevseth.github.io/research-portfolio/hayosoai/beta/en/essay/) and [*The question is allowed to change*](https://github.com/ZackaryLoevseth/research-portfolio/blob/main/hayosoai/the-question-is-allowed-to-change.md). This case develops a separate contribution; the source essays are unchanged.

[3] Tan Zhi-Xuan, Micah Carroll, Matija Franklin, Hal Ashton. [*Beyond Preferences in AI Alignment*](https://arxiv.org/abs/2408.16984), 2024. See especially Sections 4–5. Position paper; citing it does not establish empirical success of the present proposal.

[4] Joe Edelman and colleagues. [*Full-Stack Alignment: Co-Aligning AI and Institutions with Thick Models of Value*](https://arxiv.org/abs/2512.03399), December 3, 2025, version 1. See Sections 3.1 and 4. Selectively discussed here, not comprehensively assessed or independently validated.

[5] Collective Intelligence Project. [*Global Dialogues*](https://www.cip.org/globaldialogues), especially its call for partner questions. Accessed September 19, 2026. This document is not acceptance into that program or a partnership announcement.

[6] Allan Dafoe and colleagues. [*Open Problems in Cooperative AI*](https://arxiv.org/abs/2012.08630), 2020. A research-agenda connection, not a statement that HayosoAi originated this field.

[7] Anne Arzberger and colleagues. [*Co-Constructing Alignment: A Participatory Approach to Situate AI Values*](https://arxiv.org/html/2601.15895v2), arXiv version 2, April 21, 2026, especially Sections 5.2.1 and 9.1. The qualitative study concerns researchers using LLM assistants; its findings are not evidence that this garden record works. The comparison with [*Beyond Preferences*, version 2](https://arxiv.org/html/2408.16984v2) uses Sections 4.3 and 5.2. These are the versions used for the comparison.

[8] Commission nationale du débat public, [*Foire aux questions*](https://www.debatpublic.fr/foire-aux-questions-168), sections on participation and the effect of the participation report. French primary institutional guidance, accessed September 20, 2026. The comparison concerns described responsibilities, not proof of effectiveness, legal applicability to this case, or a claim of project participation. The three continuations and unfinished-concern variant above are HayosoAi’s conceptual proposals, not findings from the cited research.

Version 0.3 develops the response-responsibility question through competing continuations, an unfinished concern, and a non-AI institutional comparison. It updates project attribution without changing the founding essay or historical research results. No response procedure is presented as a settled solution.

Version 0.2 added the non-participation branch and its source comparison, made the assumed reversibility explicit, and carried Compass 1.2’s open question about what counts as improvement into the handoff. It preserved version 0.1's four stages and both histories. The question of when a challenged decision should pause remains unresolved; the scenario's stipulation is not a rule for real agreements.

All case events and example records are constructed. The case develops questions for conceptual research and practical co-design. The finite Reopening Probe and Pilot 01 retain their separately recorded results. Existing licensing applies.

Related versions: [Spanish educational preview](https://github.com/ZackaryLoevseth/research-portfolio/blob/main/hayosoai/collaboration/shared-inquiry-case-01.es.preview.md), a draft adaptation with provisional translation and teaching choices for local review; [offline handoff example](https://github.com/ZackaryLoevseth/research-portfolio/blob/main/hayosoai/collaboration/case-01-handoff/README.md), which preserves this constructed case as notes and source-linked records. Practical effectiveness remains open.

[Join the existing open conversation](https://github.com/ZackaryLoevseth/research-portfolio/issues/5) · [Read the founding essay](https://zackaryloevseth.github.io/research-portfolio/hayosoai/beta/en/essay/)