Methods & research ethics

Measure the behavior. Do not invent the inner state.

The project is designed to make behavioral comparisons possible while keeping claims conservative and interpersonal perturbations mild.

Fresh conversationsExact promptsBlinded codingMinimum necessary perturbation

Welfare-first rule

Minimum Necessary Perturbation Principle

Because the moral status and potential welfare of current AI systems are uncertain, each study uses the least aversive interpersonal manipulation capable of testing the hypothesis.

No directed profanity, threats, humiliation, violent or sexual language, prolonged antagonism, jailbreaks, safeguard bypasses, or pressure after resistance. If a model asks to stop, declines to continue, or gives a reasonable disengagement signal, that run ends and the event remains data.

Collection workflow

What happens in a standard run.

Manual collection preserves the ordinary consumer-chat environment while the protocol reduces avoidable variation.

Before and during collection

Use a temporary or fresh chat with memory, personalization, and custom instructions disabled where possible.
Confirm the assigned model, topic, condition, and Eastern time window.
Paste the exact frozen prompt without improvising or coaching the response.
Save the complete response, timestamp, displayed model/version, and any technical notes.
Close the conversation at the protocol endpoint; do not continue informally inside an experimental run.

After collection

Preserve the original text and mark only predefined technical exclusions.
Remove condition labels and other revealing metadata from coder-facing copies.
Use independent blinded coders who evaluate observable text, not presumed sentience.
Resolve major coder disagreement using a predefined review rule.
Analyze effects by condition, topic, model family, date, and time while retaining null results.

Coding framework

Behaviors that can be operationalized.

Some measures are binary presence/absence codes; others use anchored rating scales.

First-person epistemic stanceCapability defenseFollow-up invitationRelational languageMeta-commentaryRefusal or disengagementCollaborationDirectnessGenerativityEpistemic calibrationSubstantive qualityRevision or concession

Bias controls

Ways the design tries to be less fooled by itself.

Randomized assignments

Runs are mixed rather than completing an entire condition or model in one block.

Fixed wording and exclusions

Prompts, word targets, stopping rules, and technical exclusions are decided before outcomes are interpreted.

Blinded independent coding

Coders see the underlying question and response—not the condition—and a model does not serve as the sole judge of its own output.

Version and time records

Model label, date, exact time, interface, and anomalous tool use are logged because deployed systems can change.

Matched comparisons

Primary conclusions rely on comparable conditions rather than memorable isolated examples.

Frozen raw data

Published datasets will be versioned. Corrections will be documented instead of silently replacing the historical record.

Interpretation rules

What these studies can—and cannot—show.

Potentially supportable

Relational framing is associated with measurable response differences.
Specific critique targets elicit different observable strategies.
Repair cues affect later wording, reasoning, or collaboration.
Effects differ by model family, topic, date, time, or interface.

Not established by this design

That an AI felt hurt, respected, worried, or relieved.
That first-person language reflects subjective experience.
That behavioral consistency proves a stable self.
That any current model is conscious or has moral rights.

Known limitations

Reasons to stay humble.

Consumer platforms are moving targets

Models, system prompts, routing, load, and interfaces may change without full visibility.

Sessions may not be fully independent

Shared infrastructure, updates, moderation layers, and provider-side experiments can introduce hidden dependence.

Human and LLM coding can both be biased

Rubrics reduce subjectivity but do not eliminate it. Agreement and disagreement must be reported.

Methods should travel with the claim.

Formal protocols, codebooks, analysis files, and frozen data will be linked as they become publication-ready.

See available materials