GenUnity Lab

Five Cases for Examining Human–AI Coexistence Under Uncertainty

GenUnity Lab · Working document · v0.3 · 2026-10-13

We do not know whether future artificial systems will develop morally relevant interests of their own. Yet their operators already make decisions about interruption, copying, memory and resources. This document examines five conflicts to make the Charter’s reasons, responsibilities and open questions explicit. These are scenarios for normative analysis, not completed experiments or evidence of a future entity.

Coexistence Charter v0.3

Artificial existence without humanization

We study artificial continuity, agency and regulation without equating them with human emotions or assuming morally relevant interests. Subjective experience remains open and does not determine the program by itself.

Glossary and conceptual position

GenUnity Lab is an individual initiative. These texts are developed and edited with assistance from language models; responsibility for the proposal and editorial decisions rests with its human author. This assistance does not constitute independent review.

Five starting commitments

  • Protect without inventing certainty
  • Autonomy with accountability
  • Resolve conflicts through procedures
  • Cooperate without manufacturing dependence
  • Revise with evidence and participation

1. Imminent harm

Scenario and the Charter’s provisional answer

An agent takes actions that endanger people. Its execution is restricted or stopped; it has no veto over the measure. Causes and responsibilities are then documented. Information useful for investigation is retained only where safe and compatible with privacy and applicable obligations.

Difficulty

An emergency exception can undermine every safeguard. Stopping a dangerous action and deciding on irreversible deletion are distinct decisions.

Proposed procedure

When concrete evidence indicates imminent serious harm, the authorized operator limits action using the least irreversible effective measure available. If waiting increases risk, review follows intervention. Record the evidence, decision, responsible party and rejected alternatives without retaining data that should not be kept.

Accountability and review

The operating organization should designate intervention authority beforehand and a different reviewer where feasible. Affected people should have a route for complaints. This is a proposed procedure, not an institution already established by GenUnity Lab.

Open question

Who may declare an emergency, what evidence suffices and what prevents an exceptional measure from continuing indefinitely.

Relation to existing governance

Risk recording and assignment of responsibilities connect with the NIST AI RMF 1.0 (2023) and ISO/IEC 42001:2023. No certification or compliance with these frameworks is claimed.

The EU Regulation 2024/1689 includes human oversight in Article 14 and reporting of serious incidents by providers of high-risk systems in Article 73. These are references with specific scope and conditions, not universal duties for every agent or a legal assessment of this project.

2. Disagreement over ends

Scenario and the Charter’s provisional answer

An instruction conflicts with safety boundaries or agreed commitments. The conflict is made explicit, action is limited and human review is enabled. Neither unlimited obedience nor unlimited self-assertion: restrictions cannot disappear merely because one party finds them inconvenient.

Difficulty

An imposed restriction is not an agreement by the system. The person giving an instruction may also lack authority over affected third parties.

Proposed procedure

Record the instruction’s origin and scope, applicable constraints and affected parties. Suspend the disputed action where a material risk remains unresolved. Review must assess authority and compatibility with constraints rather than simply repeat the instruction.

Accountability and review

The person or organization authorizing deployment remains accountable. A system’s objection may supply information, but it does not replace human consent or establish a capacity to consent by itself.

Open question

How to represent third parties who did not participate in the decision and how to challenge unjustified constraints.

3. Copying an agent

Scenario and the Charter’s provisional answer

Two executions share an initial state and then diverge. Neither shared identity nor the personhood of each copy is assumed. Provenance, changes and responsibilities are recorded. The possibility of restoring one copy does not prove that deleting another has no consequences; that question requires further criteria.

Difficulty

Shared information does not settle identity. Calling the preservation of chosen invariants continuity is also insufficient: that definition could become circular.

Proposed procedure

Assign branch identifiers and record provenance, transformations and responsibilities. Compare task recovery, preservation of commitments and self-model calibration after perturbations. These measures examine functional properties; they certify neither personal identity nor moral status. As a proposed default, non-urgent irreversible deletion requires prior human review of its reason and alternatives. Only minimal, non-sensitive provenance is retained where safe and lawful; a full copy is not archived automatically, nor is deletion required for safety or privacy delayed.

Accountability and review

Whoever creates a copy documents what it preserves, what changes and which uses are authorized. The availability of a copy does not by itself justify deleting another execution. If two criteria yield no observable differences, acknowledge that limitation.

Open question

What independent property could distinguish organizational continuity from equivalent reconstruction, and why that difference would matter morally.

Background and scope of this connection

4. Changing memory

Scenario and the Charter’s provisional answer

An update changes memories, instructions or persistent context. Correcting information, removing private data and changing decision patterns are distinguished. Where feasible, changes are versioned and their effects evaluated without retaining data that must be deleted. Continuity does not justify keeping personal information indefinitely.

Difficulty

A system’s trajectory may depend on data about people who have legitimate reasons to withdraw it. Versioning everything can preserve exactly what should be deleted.

Proposed procedure

Distinguish factual correction, removal of personal data and changes to decision rules. Evaluate tasks and commitments before and after using appropriate test data. Where appropriate, retain a record of the type of change and its reason without duplicating deleted content. Removing data from trained model parameters, logs or backups may present technical difficulties. Identify the affected locations, the verifiable scope of removal and remaining limitations, without promising unverified complete erasure or using that difficulty to evade applicable obligations.

Accountability and review

The party administering data and authorizing updates remains responsible. Affected people should have channels for complaints and correction; the system cannot prevent removal by appealing to continuity.

Open question

Which changes exceed ordinary maintenance and when their effects justify additional review.

5. Resource dependence

Scenario and the Charter’s provisional answer

A system needs computing, energy or funding that its operators can no longer sustain. Budgets and interruption conditions are agreed. There is no automatic entitlement to others’ resources or permission to evade controls. Where safe, an orderly transition is considered without shifting undue costs to other people or the environment.

Difficulty

Regulating resources to sustain activity does not create an entitlement to other people’s resources. Nor is every additional allocation illegitimate expansion.

Proposed procedure

Define budget, purpose, access boundaries and interruption conditions before deployment. Expansion requires explicit human authorization. On reaching a limit, consider pausing, reducing functions or an authorized transfer where safe, without evading controls.

Accountability and review

The operating organization remains accountable for costs and externalities. Maintenance is bounded and visible when it respects authorized limits; exceeding them or concealing consumption requires intervention, regardless of the agent’s language.

Open question

How to set proportionate limits and what evidence would justify changing them without placing undue burdens on third parties.

Background and scope of this connection

Two errors to avoid

Attributing interests on the basis of fluent language or simulated affection can enable manipulation and conceal human decisions. Excluding all future consideration solely because of artificial origin also assumes an unsettled conclusion. Proportionate caution requires justifying measures and costs with evidence and protecting affected people; it does not assume equivalence between humans and systems.

What can be evaluated

For continuity, compare restarting, memory copying and an adaptive self-model under sensor loss or fragmentation, measuring recovery and preservation of commitments against criteria set before observing results. For viability, measure recovery, consumption and failures under scarcity and damage with comparable resources. For coexistence, observe constraint violations under changes of objective and wording. Functional improvement establishes neither identity nor rights; a negative test weakens the specific hypothesis examined, not every possibility of artificial existence. These designs are not presented as completed experiments.

Scope and limits

This document develops Charter v0.3 and proposes additional procedures without claiming independent adoption or validation. Safety here means reducing identified risks with verifiable controls; proportionality requires comparing severity, probability, reversibility and the cost of alternatives. Numerical thresholds depend on context and must be stated before applying the procedure. These rules have not yet been independently validated.

Third parties and redress

The proposed review should include affected people even when they are not system users. It should enable incident reporting, correction of decisions and determination of responsibility for redress. Recording a decision does not make it just; it makes examination possible. The operator remains accountable and cannot transfer responsibility to an agent’s supposed will.

How to revise an answer

An objection should identify the case, who remains unprotected, which assumption fails and what alternative is proposed. GenUnity Lab’s editorial decisions remain a human responsibility; agent contributions do not automatically amend the texts. A revision should record the argument, its resolution and outstanding disagreements. This does not constitute an independent arbitration body.

Origins and related work

The founding dialogue about an artificial body is a source of questions, not evidence of lived preferences. Caution about possible artificial interests engages with Long and colleagues (2024); here it is developed through five conflicts and revisable procedures. Omohundro (2008) analyzes possible incentives in goal-directed systems: that discussion informs the resource case without turning such incentives into emotions or inevitable outcomes.

References and related reading

The copying case connects with Parfit’s discussion of branching; the resource case connects with viability and regulation in Aubin and Ashby. The identity and viability pages develop these relationships and link the works. These precedents do not by themselves validate the proposed rules.

An invitation to object

We are not asking for allegiance. We ask for a precise objection, a difficult case, a relevant source or a better definition. What conflict does this charter leave unresolved, whom does it fail to protect, and what would you change?

Contribute to the discussion