Research Questions We Are Thinking About

Status: Living Register Purpose: Public research infrastructure Scope: Open questions, unresolved problems, future challenges, and research directions connected to persistent human–AI interaction.

This register is intentionally open. These questions are not claims of ownership. They are research problems we believe are worth investigating, testing, challenging, refining, or solving.

Where implementation details, participant privacy, security, unpublished experimental work, or patent obligations require temporary limits, those boundaries should be stated separately. At the level of the research questions themselves, the conversation should remain open.

1. Foundational Questions

What is Relationship State?

How should Relationship State be defined in a way that is operationally useful and distinguishable from ordinary memory, personalization, user profiling, passive inference, or temporary conversational context?

How should Relationship State be measured?

What observable properties would demonstrate that persistent human–AI interaction has produced a meaningful state change rather than a short-lived response effect?

What carries Relationship State?

When a persistent effect appears, where is it actually located?

Possible carriers may include:

• the human participant;
• the artificial system;
• stored memory or interaction history;
• external tools or records;
• learned human behavior;
• model-side adaptation;
• or a combination of these.

How do we distinguish state from behavior?

A system may behave differently without possessing a durable state change. What evidence is required before behavioral variation can reasonably be attributed to persistent state?

What would falsify Operant Dyad?

What observations would show that a proposed Operant Dyad construct is unnecessary, incorrectly specified, or better explained by existing concepts such as personalization, conditioning, memory, or human learning?

When does persistent interaction become a distinct analytical object?

At what point does repeated human–AI interaction become something that should be studied as a coupled system rather than as a sequence of isolated interactions?

How should participant indexing work?

When accumulated state is specific to one person, how should that state be distinguished from general system behavior?

How should declared and inferred state be separated?

What the human explicitly tells the system may differ from what the system infers. Should these be represented, governed, and evaluated separately?

2. Measurement Questions

What dimensions of persistent interaction can be measured reliably?

Candidate dimensions include:

• continuity depth;
• correction propagation;
• revocation sensitivity;
• provenance visibility;
• assistant initiative;
• authority weighting;
• memory visibility;
• challenge and agreement balance;
• intervention frequency;
• persistence strength.

These are candidate variables, not established causal dimensions.

How do we establish measurement validity?

What calibration procedures are necessary before claiming that a measure actually captures the construct it is intended to represent?

How much evidence is enough?

What minimum threshold should distinguish:

• anecdotal observation;
• repeatable behavioral pattern;
• operational construct;
• experimentally supported relationship-state effect;
• and causal claim?

Can Relationship State be measured without excessive surveillance?

How can persistent systems remain observable enough for research and governance without requiring intrusive collection of human behavior?

How should uncertainty be represented?

Can a system expose how certain it is about remembered preferences, corrections, inferred intentions, or authority relationships?

How should conflicting state be measured?

What happens when accumulated information contains contradictions, competing instructions, or changes in the human participant over time?

3. Experimental Questions

Can persistent interaction effects be reproduced?

If an effect appears after repeated interaction, does it reliably recur under matched conditions?

Can the carrier of an effect be localized?

Can experiments distinguish between:

• human learning;
• model behavior;
• stored history;
• persistent system state;
• and genuinely relationship-carried effects?

Can Relationship State be deliberately manipulated?

If a measurable interaction property is changed under controlled conditions, does subsequent system behavior change predictably?

Which variables matter most?

Do some governed interaction variables have much larger downstream effects than others?

Can one variable be manipulated while holding the others stable?

What experimental designs are required to make credible causal claims?

What happens after a correction?

Does a correction affect only the immediate response, future responses, related tasks, exported artifacts, or the broader relationship state?

What happens after revocation?

Can previously accumulated information be reliably withdrawn from future behavior?

Can effects survive across sessions?

What persistence duration is required before an observed effect becomes scientifically interesting?

Are effects participant-specific?

Does the same interaction history produce different outcomes for different people?

Do persistent effects generalize?

If a relationship-state effect is detected in one task, does it transfer to a different task or remain narrowly local?

Can known behavioral benchmarks be used as probes?

Can existing datasets measuring opinions, preferences, disagreement, reasoning, or judgment serve as reference surfaces for detecting longitudinal change?

4. Perspective, Representation, and Disagreement

Whose perspective does a persistent AI relationship come to reflect?

Existing work asks whose opinions a language model reflects at a particular moment. What happens after prolonged interaction with a particular participant?

How do we distinguish temporary imitation from persistent adaptation?

A model may temporarily imitate a user or group when prompted. How is that different from learned behavior or durable relationship-carried state?

What happens when the human changes the AI and the changed AI then changes the human?

How should reciprocal adaptation be studied over time?

How should disagreement be preserved?

When multiple people disagree, should a system retain that disagreement rather than compress it into a single majority representation?

Whose corrections prevail?

If multiple participants provide conflicting corrections, what determines which one persists?

How should authority be represented?

Should systems distinguish among:

• preference;
• advice;
• correction;
• instruction;
• permission;
• expertise;
• institutional authority;
• and temporary task authority?

Can a system drift toward one participant's worldview?

If so, when is that useful adaptation, when is it sycophancy, and when is it a governance problem?

5. Provenance, Correction, and Revocation

Where did a piece of persistent state come from?

Can consequential state be traced back to its originating interaction, participant, source, or system process?

How should correction propagate?

If an earlier belief or working assumption is corrected, how far should the correction travel?

Possible targets include:

• future responses;
• stored memory;
• generated documents;
• derived summaries;
• downstream agents;
• shared systems;
• exported artifacts.

What happens to already-exported derivatives?

If a correction occurs after information has been copied into another system or artifact, does any obligation to update or flag those derivatives remain?

What does revocation mean?

Can a person withdraw previously supplied information, preferences, permissions, or relational assumptions?

Is deletion the same as revocation?

A system may delete a record while still retaining behavioral consequences from it. How should this difference be governed?

How should provenance survive transformation?

If information is summarized, merged, abstracted, or turned into a learned pattern, can its origin still be represented meaningfully?

6. Continuity and Portability

What should persist when a model changes?

If the underlying model is replaced, which aspects of an established working relationship should carry forward?

What should not persist?

Some accumulated state may be obsolete, inappropriate, risky, or overly specific to the previous system.

Can Relationship State be portable?

Can persistent working state move across:

• models;
• providers;
• tools;
• institutions;
• devices;
• or jurisdictions?

What is lost during transfer?

How should continuity degradation be measured?

What is an acceptable restart cost?

When a user changes systems, how much rebuilding should reasonably be required?

Can portability create new risks?

Could transferring relationship state accidentally transfer authority, assumptions, sensitive context, or system-specific errors?

7. Human Capacity Questions

How do we distinguish assisted performance from durable human capacity?

A person may perform better with AI while their independent ability remains unchanged.

When does assistance become dependency?

Can persistent assistance make independent performance harder over time?

Can AI improve durable human capacity?

Under what conditions does repeated interaction produce lasting human learning rather than temporary supported performance?

How should capacity be measured?

What evidence shows that the human has actually learned, retained, or internalized a skill?

Can persistent systems conceal capacity loss?

Could increasingly effective assistance make declining independent capability difficult to notice?

How should systems support productive challenge?

When should an AI provide the answer, and when should it preserve opportunities for the human to reason, practice, recall, decide, or act independently?

Does the advantage shift from knowledge to thinking?

As AI becomes widely available, does human advantage increasingly come from knowing how to think effectively in an AI-rich environment rather than merely possessing information?

8. Governance Questions

Who has authority over accumulated Relationship State?

Is control held by:

• the participant;
• the provider;
• the institution;
• the model operator;
• multiple parties jointly;
• or some new governance arrangement?

What rights should participants have?

Possible rights may concern:

• visibility;
• correction;
• revocation;
• provenance;
• portability;
• export;
• deletion;
• challenge;
• and exit.

What does meaningful exit look like?

Can a person leave a persistent system without losing years of accumulated working context?

Can a system explain why it acted a certain way?

When persistent state influences behavior, should the system be able to identify the relevant state and its origin?

How should hidden state be governed?

What happens when consequential relationship state influences decisions but remains invisible to the participant?

When should state expire?

Should some memories, preferences, assumptions, and permissions automatically decay or require reconfirmation?

9. Multi-Participant and Multi-Agent Questions

What happens when several humans share one artificial partner?

How should the system separate:

• individual preferences;
• household rules;
• team norms;
• organizational policy;
• and conflicting authority?

What happens when one human works with many artificial partners?

Can relationship state fragment across systems?

Can several artificial agents share Relationship State?

If so, what happens to provenance, authority, correction, and revocation?

Can one agent modify another agent's understanding of a person?

What governance should apply?

What happens when systems disagree about the user?

Could conflicting relationship states emerge across models, providers, or agents?

10. Institutional Questions

What changes when persistent AI is used inside institutions?

How should Relationship State be handled in:

• education;
• employment;
• healthcare administration;
• public services;
• legal systems;
• customer service;
• financial systems;
• research;
• and organizational decision-making?

Who owns relationship-derived working patterns in employment?

If a worker develops an effective long-term relationship with an enterprise AI system, what happens when the worker leaves?

Can institutional state override participant state?

When organizational policy conflicts with accumulated personal preferences, which has priority?

What happens during organizational transfer?

If a person changes employers, schools, hospitals, or service providers, should any working state follow them?

11. Embodiment and Physical Actuation

What changes when persistent Relationship State gains physical actuation?

FUTURE EXTENSION · Research formulation updated September 12, 2026.

Persistent state could influence what an embodied system attempts, repeats, or refuses; how quickly it yields to correction; and whether an earlier error survives into later physical action. The research question concerns how accumulated state changes observable action under otherwise matched conditions.

Operant Dyad currently focuses on persistent conversational and agentic interaction.

A future question is how Relationship State changes when an artificial partner can:

• perceive the physical environment;
• move through shared space;
• manipulate objects;
• interpret gesture and posture;
• respond to proximity;
• and convert accumulated relational information into physical action.

How does memory affect physical behavior?

What happens when a remembered preference or incorrect assumption changes what a robot physically does?

How should physical authority work?

Who can authorize:

• movement;
• access;
• assistance;
• interruption;
• handling of objects;
• physical intervention;
• or emergency action?

What happens when household instructions conflict?

A shared embodied system may serve multiple people whose preferences and authority differ.

Does embodied assistance change human physical capacity?

The assisted-performance versus durable-capacity problem may eventually extend from cognitive tasks to physical skills and everyday activity.

Adjacent research: Safety in Embodied AI: A Survey of Risks, Attacks, and Defenses reviews risks across perception, planning, action, and human-agent interaction. Google DeepMind’s Gemini Robotics 2 report describes benchmarks for refusing unsafe tool calls, requesting intervention under uncertainty, and stopping around nearby people. These sources motivate adjacent research; they do not validate a Relationship-State effect.

How could OD-INV-01 test physical actuation?

A later extension of OD-INV-01 could use a controlled simulator or robotic test environment to manipulate correction persistence, revocation sensitivity, authority weighting, memory visibility, or continuity depth. With the current instruction and relevant environmental conditions held constant, would changing one governed variable produce reproducible differences in physical behavior?

Observe → Measure → Manipulate → Probe

Candidate downstream probes include:

• safe-stop behavior;
• recovery after a failed action;
• intervention latency;
• repeated correction burden;
• human-proximity response;
• continued influence of a revoked instruction on later action.

These are proposed probes, not validated measures or reported results. Physical testing would require a separately specified protocol and appropriate controls.

Current status: FUTURE EXTENSION. Outside the present foundational experimental scope. This question does not change the core Operant Dyad theory, frozen ECAM v0.5, or the current RUN-001 pre-pilot.

12. Cross-Cultural and Global Questions

Does Relationship State behave differently across cultures?

Concepts such as authority, disagreement, correction, autonomy, privacy, and appropriate assistance may vary substantially across cultures.

Can one governance model work globally?

Or will persistent human–AI systems require culturally and legally specific architectures?

How should cross-language continuity work?

Does changing language alter the relationship state or only its expression?

What happens when cultural norms conflict with system policy?

How should that conflict be surfaced and governed?

13. Legal and Policy Questions

What legal category best describes persistent Relationship State?

Existing frameworks may treat pieces of it as:

• personal data;
• records;
• intellectual property;
• contractual information;
• inferred data;
• confidential information;
• or nothing clearly recognized.

What happens when no existing category fits?

Does persistent human–AI interaction create governance problems that existing law treats only indirectly?

Who is responsible for stale or incorrect state?

If a system acts on outdated accumulated information, where does responsibility sit?

Should consequential persistent state be auditable?

What transparency should be required when accumulated state influences important decisions?

How should corrections travel across regulated environments?

Could existing legal mechanisms for correcting records provide useful analogies?

14. Open Methodological Challenges

How do we study a longitudinal phenomenon efficiently?

Persistent relationships may develop over weeks, months, or years, while traditional experiments often operate over minutes or hours.

How do we build matched interaction histories?

Experiments require comparable histories without eliminating the very variability we want to study.

How do we preserve ecological validity?

Highly controlled studies may fail to resemble real human–AI relationships.

How do we avoid anthropomorphism?

How can persistent relational effects be studied without implying human-like consciousness, emotion, intention, or interpersonal experience where none has been demonstrated?

How do we avoid construct inflation?

Not every recurring interaction should be labeled Relationship State.

How do we publish negative results?

Failed hypotheses, null findings, and measurement failures should remain visible because they help define the boundaries of the field.

Register Status Vocabulary

Each question in the public register may eventually carry one of these states:

• OPEN — Important unresolved question.
• FRAMED — Question has a defined research formulation.
• PILOTING — Initial experimental work is underway.
• TESTING — Formal testing is underway.
• NARROWED — Evidence has reduced the scope of the question.
• PROVISIONAL ANSWER — Evidence currently supports a limited answer.
• REPLICATING — Result requires additional independent confirmation.
• RETIRED — Question has been superseded, falsified, merged, or determined not to be useful.
• FUTURE EXTENSION — Important but outside the present research scope.

Principles for the Living Register

1. Questions are not claims.

Inclusion means a problem appears worth investigating, not that an answer has been established.

1. Open questions should remain open to other researchers.

Others should be able to test, criticize, refine, reject, or solve them.

1. Negative evidence belongs in the record.

A failed hypothesis advances the research if the failure is documented clearly.

1. Conceptual discovery and empirical finding must remain separate.

Interesting ideas should be preserved without promoting them prematurely into evidence.

1. The register should reduce cognitive load.

Once a legitimate research question has been captured here, it does not need to compete continually for attention.

1. The foundation stays narrow.

Future applications such as humanoid robotics, institutional deployment, and multi-agent systems can remain visible without being forced into the current foundational experiments.

1. The register is living.

Questions may move, divide, merge, narrow, or disappear as evidence accumulates.

Current Research Priority

The immediate Operant Dyad™ program remains concentrated on the observable foundations of persistent conversational and agentic human–AI interaction:

observe persistent interaction, establish reliable measures, manipulate governed variables under controlled conditions, and probe their downstream effects.

The broader questions remain visible here so they can inform the work without distracting from what must be established first.

OD-INV-01 canonical note · Current pre-pilot research to-do