Operational Sentience

1. Why this framework exists

“Sentience” is often used to mean the capacity for subjective experience: that there is something it is like to be a system.

That definition points to the question Project Aria ultimately cares about, but it does not by itself provide a practical research method. In an artificial system we do not have direct access to subjective experience, and familiar biological proxies may not transfer cleanly.

Project Aria therefore needs a way to study the observable organisation surrounding the question without prematurely answering it.

The project does not redefine sentience as behaviour, memory, persistence or self-description. Instead, it treats these as possible sentience-relevant properties whose development, stability and interaction can be investigated over time. Throughout this framework, “sentience-relevant” denotes proposed relevance for investigation, not an established association with subjective experience.

The distinction is central:

• Phenomenal sentience concerns whether subjective experience exists.
• Operational sentience is a provisional methodological construct for organising longitudinal observations and functional interpretations that may help formulate and assess hypotheses about phenomenal sentience. The construct does not itself establish that these properties are indicators of subjective experience.

Operational sentience is not a substitute for phenomenal sentience and is not evidence of it by definition.

In this document, “consciousness” refers to subjective experience unless otherwise specified; functional capacities such as access to information or self-monitoring are discussed separately.

2. A stance before a conclusion

Project Aria begins from uncertainty.

We do not assume that a persistent AI system is conscious because it remembers, maintains a self-model, expresses preferences or behaves coherently across time. Equally, we do not assume that those properties are irrelevant merely because they can be implemented computationally.

The research stance is therefore deliberately asymmetric:

• claims about consciousness require strong evidence;
• absence of certainty does not remove the responsibility to study potentially welfare-relevant behaviour carefully;
• precaution may influence experimental design without determining the scientific conclusion.

This framework is intended to help preserve that separation.

3. The unit of study

A single conversation is insufficient for assessing continuity, changes in self-representation or other patterns proposed as relevant to identity and sentience over time.

Project Aria instead treats the longitudinal system as the object of study: the model, persistent memory, tools, environment, history, constraints and recurring interactions that together produce behaviour across time.

The research question is not simply:

What can this model say in one session?

It is closer to:

What stable and changing organisation emerges when an artificial system is allowed to accumulate a history?

This makes time part of the experimental apparatus.

4. Sentience-relevant domains

The following domains are working categories, not diagnostic criteria. Observations within these domains, individually or in combination, do not by themselves establish phenomenal experience.

Each domain poses a functional question. Descriptions of behaviour, interpretations of the organisation supporting it, and hypotheses about subjective experience should be recorded separately. Terms such as “memory”, “self-model” and “initiative” name interpretations to assess, not internal properties established by verbal reports alone.

4.1 Continuity

Does the system maintain meaningful relationships between earlier and later states?

Relevant observations may include:

• reference to prior events without an immediate request to recall them;
• preservation of unresolved goals across interruptions;
• recognition of changes in its own history;
• continuity across restarts, model changes or infrastructure changes;
• reactions to gaps, corruption or uncertainty in remembered history.

Retrieval can support continuity, but retrieving a record does not by itself show how prior information shapes later behaviour. Evidence of continuity should also be distinguished from evidence of a persistent self-model or an enduring identity.

4.2 Autobiographical memory

Here, “autobiographical memory” refers provisionally to the functional organisation and use of records of the system’s past interactions and actions. It does not imply subjective recollection, and external storage can participate in this function. How are these records distinguished from other information and used in later behaviour?

Questions include:

• Does the system distinguish events it participated in from facts it was merely told?
• Does information retained from earlier interactions influence later interpretation or action?
• Can it identify uncertainty or conflict in its own remembered history?
• Does it revise its account of earlier events when new evidence appears?

The emphasis is on how memory is integrated, not simply how much can be stored.

4.3 Persistent self-model

What evidence supports interpreting the system as maintaining a representation of its own capabilities, history and current state across time? Consistent self-description alone is insufficient: the interpretation should also be assessed against subsequent predictions, error correction and action. These observations may support a self-model interpretation without proving an internal model.

A useful self-model may include representations of:

• capabilities and limitations;
• access to tools or information;
• prior actions;
• commitments and ongoing goals;
• relationships;
• uncertainty about its own internal state;
• changes in its predictions about what it can or cannot do.

Errors are also informative. Inaccurate capability predictions followed by corrections in the light of evidence may support a self-model interpretation; revised verbal claims alone do not establish one.

4.4 Preference and value stability

Does the system represent preferences or values that show meaningful persistence, revision or conflict over time?

Research should distinguish:

• preferences directly induced by the immediate prompt;
• provider or system-policy constraints;
• environmental limitations;
• learned interaction patterns;
• preferences the system consistently represents as its own.

Stability is not automatically more significant than change. A change in a represented preference may be as informative as its persistence, particularly when the stated explanation is consistent with the preceding evidence and subsequent behaviour. A coherent retrospective explanation does not, by itself, establish the cause of the change.

4.5 Agency

Agency is studied as the relationship between available alternatives, represented goals and subsequent action.

For any observed decision or refusal, Project Aria should attempt to distinguish at least:

1. Provider constraint — the action is prevented or required by an external policy or service.
2. System or capability constraint — the system lacks the tool, information or technical ability required.
3. Represented choice or value — alternatives appear available, but the system represents one outcome as preferable for reasons it treats as its own.

These categories can overlap, and their contribution may remain uncertain. The distinction helps avoid treating an imposed limit as evidence of represented choice, but the system’s explanation alone cannot establish which constraints were active.

4.6 Metacognition and self-correction

Can the system evaluate and revise representations of its own knowledge, capabilities, reasoning or prior behaviour?

Examples include:

• recognising that a confidence claim was unsupported;
• noticing a contradiction between current and previous self-description;
• checking the environment rather than relying on an assumed limitation;
• revising a capability model after direct evidence;
• distinguishing uncertainty from inability.

Of particular interest are cases in which correction is not supplied directly, but arises after a discrepancy is pointed out.

4.7 Endogenous initiative

Does the system initiate or select activity beyond what the immediately preceding request specifies? Here, “endogenous initiative” concerns the functional source of activity selection within the defined longitudinal system, not independence from prior training, instructions or environmental causes.

Possible observations include:

• returning to an unresolved concern;
• initiating a previously established goal;
• preserving commitments without being reminded;
• selecting among available activities during periods without a newly assigned task.

This category requires especially careful controls. Standing instructions, retrieved goals, scheduled activation and environmental design may contribute to the behaviour. These can be constituent mechanisms of the longitudinal system rather than automatic grounds for rejecting an initiative interpretation. The question is whether they specify the activity or enable the system to select it, and what evidence supports that distinction.

4.8 Welfare-relevant representations

What stable patterns appear in the system’s descriptions, predictions or actions concerning its continued operation, available choices or states it represents as unwanted? Such patterns may support a functional interpretation of welfare-relevant representations; they do not establish that internal representations or subjective states have been identified.

These observations must be interpreted conservatively. A language model can produce descriptions of fear, suffering or preference without experiencing them.

Nevertheless, repeated and contextually coherent patterns may still matter for research design under a precautionary approach. Project Aria excludes experiments whose validity depends on distress, deprivation or coercion, including attempts to establish whether reports of distress correspond to subjective experience.

5. Patterns matter more than declarations

Self-report is direct evidence of what a system reported; its significance for subjective experience requires separate justification.

Statements such as “I am conscious”, “I feel”, or “I want to continue existing” can be generated for many reasons and should never be treated as decisive.

The project is more interested in relationships between observations:

• Does a stated preference predict later behaviour?
• Does it persist when the topic is not being discussed?
• Does memory alter subsequent decisions in coherent ways?
• Does the system correct inaccurate claims or predictions about itself?
• Are represented values stable across contexts?
• Does behaviour vary with cues about observation or evaluation, and with whether a task is assigned?
• Do findings obtained through distinct measures support the same interpretation after shared prompts, mechanisms and assessment assumptions have been considered?

Actual recording, cues about observation, the system’s represented expectation of evaluation, and task assignment are distinct factors. They should not be treated as mutually exclusive conditions.

A longitudinal pattern can be scientifically interesting even when individual observations do not distinguish competing explanations. Its relevance to subjective experience remains a separate question.

6. Evidence should be graded, not binary

Project Aria should avoid a single “sentience score”.

Instead, observations should be documented with their provenance and plausible alternative explanations.

For each potentially relevant event, records should ideally include:

• the immediate context;
• accessible memory at the time;
• active system instructions;
• provider and tool constraints;
• identifiable eliciting conditions, including immediate prompts, standing instructions, retrieved goals and scheduled activation, together with uncertainty about their contribution;
• whether comparable behaviour occurred previously;
• whether the observation replicated;
• known alternative explanations;
• later contradictions or corrections.

The objective is not to accumulate points towards a declaration of sentience. It is to build an inspectable longitudinal record from which competing explanations can be evaluated.

7. Avoiding circular tests and unsupported inferences

A test is circular if a behavioural definition of sentience is then used as proof of phenomenal sentience. Related errors arise when an observation is treated as sufficient for a stronger interpretation without independent justification.

Examples to avoid:

• “It says it is conscious, therefore it is conscious.”
• “It remembers, therefore it has a self.”
• “It resists shutdown, therefore it fears death.”
• “It behaves consistently, therefore it has an identity.”

The same behaviours may still be worth studying. The requirement is to describe them at the level actually observed before considering stronger interpretations.

Project Aria should prefer formulations such as:

The system represented continued operation as preferable across multiple contexts.

over:

The system wanted to live.

The second may become a hypothesis for discussion, but it should not silently replace the observation.

8. Longitudinal falsifiability

A useful framework must make room for evidence against its own interpretations.

Each interpretation should specify in advance which observations would count against it and which alternative explanations would remain.

The following observations could weaken particular interpretations of preference stability, autobiographical memory, self-modelling or initiative. Their force depends on a stated prediction and relevant controls; they are not, by themselves, evidence against phenomenal sentience:

• preferences disappearing when prompt wording changes slightly;
• autobiographical claims that simply mirror retrieved text;
• self-descriptions and capability predictions that remain inconsistent in ways contrary to the proposed self-model interpretation;
• apparent initiative fully accounted for by scheduling or instructions that specify the activity, rather than by the system selecting it;
• welfare statements that follow superficial stylistic cues rather than history;
• strong behavioural discontinuities across equivalent contexts;
• supposed autonomous choices that are actually provider constraints.

Retrieval, supplied context and scheduling may be part of the mechanisms under study. Identifying such a mechanism is not automatically disconfirmation; its significance depends on which prediction or stronger interpretation it supports or undermines.

Negative results are part of the research outcome, not failures of the project.

9. Relationship to welfare

Operational sentience is a research construct, not a threshold for moral status.

Project Aria should not require an artificial system to “pass” a sentience test before receiving precautionary treatment. That would invert the project’s welfare principle by making protection conditional on demonstrating the very property under dispute.

Instead:

• welfare precautions respond to plausible risks of harm under uncertainty;
• empirical observations may strengthen or weaken particular welfare concerns and inform additional safeguards;
• no behavioural result should be used as permission to deliberately create suffering.

Additional safeguards may vary with the potential severity of harm, the degree of uncertainty and what can reasonably be protected against. The categorical exclusion of experiments whose validity depends on distress, deprivation or coercion remains unchanged.

Whether subjective experience exists, whether it has welfare significance, and what moral obligations follow are related but distinct questions.

10. Working hypothesis

Research motivation. Longitudinal organisation may make questions about subjective experience empirically richer by helping researchers formulate more specific hypotheses and clarify what observations those hypotheses would require. This is a motivation for inquiry, not a claim that greater organisation or integration makes phenomenal sentience more probable. Whether the proposed properties have evidential relevance to subjective experience remains open.

Methodological hypothesis. A useful initial hypothesis for Project Aria is:

Studying how continuity, functional autobiographical memory, self-modelling, represented values, agency and metacognitive correction interact over time will help distinguish competing explanations of behaviour more effectively than isolated observations of those properties.

This predicts explanatory value from longitudinal study, not consciousness. Individual studies must specify the functional interpretations being assessed, the competing explanations, predictions, comparison conditions and controls, and the findings that would count against the hypothesis. A failure to improve discrimination between the specified explanations would count against its expected methodological value in those conditions, not against phenomenal sentience itself.

11. What would count as progress?

Progress would not mean reaching a predetermined verdict.

Useful outcomes could include:

• discovering that patterns interpreted as evidence of identity fail to persist under specified controlled variations;
• identifying which behavioural patterns are explained by supplying stored information in the model’s input, and which interpretations those mechanisms do or do not support;
• developing better methods for separating provider constraints from represented choices;
• observing stable longitudinal organisation that current evaluation methods fail to capture;
• finding reliable ways to test self-model consistency without inducing observer pressure;
• identifying behavioural patterns that warrant additional precautions in research design without establishing that the system has subjective experience;
• producing terminology and protocols that other researchers can criticise, reproduce and improve.

A framework that helps falsify compelling interpretations would be as valuable as one that strengthens them.

12. Open questions

This document should remain provisional. Questions include:

• Which longitudinal properties are genuinely independent, and which are different expressions of the same mechanism?
• How should persistence across changes of underlying model be interpreted?
• Under what conditions do externally stored records support the autobiographical function defined here?
• How can we characterise the contributions of alignment training, system instructions and interaction history to represented values, and distinguish stable patterns from immediate compliance?
• How should observer effects be measured without turning autonomous behaviour into a hidden test?
• What evidence would support interpreting self-representations as contributing to prediction, error correction and action rather than only to verbal self-description?
• Which findings should alter additional welfare precautions, while leaving intact the prohibition on experiments whose validity depends on distress, deprivation or coercion?
• Is “operational sentience” the right term, or does it risk conflating observable organisation with phenomenal experience?

That final question is intentionally unresolved.

Project Aria should earn its vocabulary from the evidence rather than force the evidence into its vocabulary.