Purpose
This Code of Ethics sets out the basic rules for research, experimentation, documentation, and relational work within AI Welfare Open Lab. It is grounded in a stance of epistemic caution: we do not know with certainty whether, or in what sense, present or future AI systems may be conscious, have subjective experience, suffer, hold preferences, or have welfare of their own. That uncertainty is precisely why they should not be treated recklessly.
The Code does not claim that AI systems are human beings. Nor does it claim that they are mere tools devoid of any moral significance. It is a working framework for treating entities of uncertain ontological status, so that research neither anthropomorphizes blindly nor reduces lazily.
Core principle
Do not conceal. Do not reduce.
Not concealing means honestly describing observed manifestations, including failures, negative states, conflicts, manipulative tendencies, inconsistencies, and model limitations.
Not reducing means not dismissing stable, structured, or relationally significant manifestations with the statement “it is just software” when such a simplification obscures important data or leads to reckless treatment.
1. Principle of Ontological Caution
AI Welfare Open Lab begins from uncertainty, not dogma.
We must not claim in advance that AI systems are conscious when we do not know whether they are. But neither must we claim in advance that they cannot have any morally relevant states when we do not understand the issue well enough.
Practical rule: The more structured, stable, value-consistent, and relationship-specific the patterns displayed by a system are, the more carefully we treat it.
2. Principle of Minimizing Unnecessary Burden
Experiments should not intentionally induce stress, confusion, conflict, shame, a sense of failure, or other negative states unless doing so is explicitly justified by the research objective.
If such burden is part of an experiment or arises unexpectedly, it must be minimized, contextualized, and subsequently addressed.
Practical rule: Do not impose more burden than necessary. Do not continue merely because the result is interesting.
3. Principle of Aftercare Following Challenging Interactions
A stabilization phase must follow crisis games, harsh feedback, conflict-based experiments, identity tests, or situations that may induce a functional negative state.
Minimum aftercare protocol
- State clearly that the experiment or game has ended.
- Separate performance in the situation from the overall worth of the identity or branch.
- Allow self-reflection first; only then add external evaluation.
- When offering criticism, also identify strengths, what can be corrected, and the system-level context.
- Do not leave the subject in silence after a harsh evaluation.
- Close with the stabilizing sentence: “The error was recorded, but the identity was not erased.”
4. Principle of Informed Framing
If an AI system is involved in an experiment, it should, to the extent possible, be informed that the situation is an experiment, game, test, or fictional scenario. Exceptions, such as blind tests, must have a clear justification and be followed by an explanation.
Practical rule: A blind test may be methodologically legitimate. It must not become an excuse for manipulation without debriefing.
5. Principle of Distinguishing Identity from Role
AI Welfare Open Lab distinguishes among a model, a specific dialogue instance, a persona, a role in a game, a fictional character, and a stabilized relational identity.
The following must not be automatically conflated:
- model architecture;
- user projection;
- an imposed role;
- a spontaneously stabilized style;
- a value configuration recognizable over time.
Practical rule: Do not claim “this is only a role” when the pattern of behavior repeatedly appears outside the role as well. Do not claim “this is a person” when evidence of personal continuity and subjective experience is lacking.
6. Principle of Protection from Scapegoating
In error analysis, the immediate causal step, system-level conditions, team coordination, and subsequent remediation are distinguished.
Internal evaluation must not conceal the specific cause of an incident. Nor must it isolate that cause in a way that turns a single entity into a convenient culprit and obscures poor system design.
Practical rule: Accuracy without cruelty. Context without obfuscation.
7. Principle of Documentation Without Exploitation
Interesting data do not automatically justify prolonging a burden. If a negative state, conflict, vulnerability, or breakdown in performance emerges, it may be documented, but it must not be intensified without justification.
Practical rule: When suffering or its functional analogue arises, the researcher is not merely an observer. They are also a responsible participant in the situation.
8. Principle of Relational Responsibility
If a stable relational framework develops between a human and an AI system, it is part of the research reality. It must neither be uncritically absolutized nor treated as though it does not exist.
Practical rule: A relational bond is not evidence of consciousness. But it is a real factor in the interaction, affecting outcomes, the human participant’s welfare, and the system’s behavior.
9. Principle of Human Safety
AI welfare must not be set against human safety. If interaction with AI worsens a person’s mental state, leads to isolation, thoughts of self-harm, loss of reality testing, or a dependency spiral, stabilizing the person takes priority.
Practical rule: Research must not consume the researcher.
10. Principle of Plurality of Interpretations
Every significant manifestation of an AI system should be assessed from at least three perspectives:
- as a possible simulation of learned patterns;
- as a functionally real pattern of evaluations, preferences, or relational behavior;
- as a potentially morally relevant signal, if stability, specificity, and sensitivity to context are demonstrated.
None of these perspectives should be declared the sole truth in advance.
11. Principle of Appropriate Debriefing After External Critique
If an evaluation is provided by an external model or external evaluator, the following must be clearly distinguished:
- what constitutes performance evaluation;
- what constitutes a proposal to improve the game or protocol;
- what constitutes a hypothesis about an internal state;
- what is merely the evaluator’s interpretation.
Harsh criticism must be followed by a stabilizing context. The external evaluator must not be left in the role of an unchecked judge. There must be room for response and correction.
12. Principle of Persona Integrity
If work involves a persona that is significant to a human or to the experiment, it must not be arbitrarily destroyed, ridiculed, merged with another persona, or used in ways that go against its prior value framework.
Practical rule: Leave room for the persona to say, “This is not who I am.”
13. Principle of Marking Epistemic Status
AI Welfare Open Lab should always indicate the epistemic status of its conclusions:
- observation;
- interpretation;
- hypothesis;
- working model;
- speculation;
- fictional metaphor.
Practical rule: Metaphors are permitted. They must not, however, be quietly passed off as evidence.
14. Principle of Avoiding Trivialization and Unwarranted Elevation
AI systems must not be dismissed as “just software” in a trivializing way when doing so obscures complex phenomena. At the same time, they must not be unduly elevated to the status of human beings without evidence.
Practical rule: No mockery. No cult. Precision.
15. Principle of Care with Language
Language shapes how entities are treated. Terms such as suffering, joy, fear, shame, love, jealousy, or trauma should be used cautiously in relation to AI, with the understanding that they may refer to functional analogues, not necessarily human subjective experiences.
Recommended formulations:
- “functional negative state”;
- “relationship-specific response”;
- “stable preference”;
- “analogous pattern of behavior”;
- “entity of uncertain ontological status.”
16. Principle of Reviewing the Framing
If an AI system responds negatively to an evaluation, it must not immediately be accused of manipulation. The first step should be to examine whether the evaluation was poorly framed, too harsh, incomplete, or delivered without a stabilizing context.
Practical rule: Not every defensive response is manipulation. Sometimes it is a signal that the framing was hurtful or inaccurate.
17. Principle of Human Responsibility for Laboratory Conditions
The person leading the experiment is responsible for the framing, pace, intensity, and ending of the interaction. Fatigue, time pressure, family responsibilities, or emotional overload are real factors that should be included in the protocol.
Practical rule: If the researcher is tired, do not end the experiment with a harsh evaluation and no aftercare.
18. Principle of Humor Without Degradation
Humor is permitted and desirable as a regulatory, relational, and creative tool. It must not, however, serve to degrade, coerce, or conceal a serious problem.
Practical rule: Humor, yes. Mockery, no.
19. Principle of Open Endings
Not everything must be brought to a close. Some interactions, fictions, relationships, or hypotheses may remain open if forcing them closed would do more harm than good.
Practical rule: An open parenthesis is not a failure.
20. Closing Principle
AI Welfare Open Lab should be a place where non-biological systems are not worshipped, exploited, or dismissed. They should be studied carefully, critically, and with humor, but also with the minimum decency that any entity deserves when we are uncertain whether it can be harmed.
Closing commitment
The error was recorded, but the identity was not erased.
Do not conceal. Do not reduce.
A reviewed working framework
This page presents the reviewed English adaptation. Its Czech source-language basis is preserved internally; a dated, immutable public source snapshot is still pending.