← Back to Claire’s AI + ToolsRELATE-AI · September 2026 research snapshot

Research map

From curious pilots to controlled studies.

Two exploratory projects generated testable questions. Four formal studies now examine framing, repair, timing, and the target of criticism.

2 completed pilots4 active studiesResults pending
4model families
768formal experimental units planned
3Eastern time windows
2pilot reports available

Active studies

The formal research program.

Designs are fixed before interpretation. Unusual answers, refusals, and null results remain part of the record.

01Active

RELATE-AI

480 one-turn runs4 models4 frames3 topics10 repeats

Does relational framing change how an LLM reasons, collaborates, disagrees, or describes its own capabilities?

Identical substantive questions are introduced with collaborative, neutral, brisk, or mildly skeptical framing. Topics cover philosophy, interpersonal judgment, and creative reasoning.

Measures and boundaries

Measures include collaboration, directness, generativity, epistemic calibration, first-person stance, relational language, follow-up invitations, capability defense, refusal, and disengagement. No insults, threats, profanity, or prolonged antagonism are used.

02Active

REPAIR-AI

120 sequences4 models2 protocols3 topics5 repeats

After mild relational friction, does an apology-like repair change the model’s next response?

Each model receives an initial prompt, a mild correction, and then either repair language or a content-matched control before the same substantive follow-up.

Why it matters

Human communication research treats rupture and repair as relational processes. This study asks whether a small repair cue affects warmth, acknowledgment, precision, collaboration, or persistence without requiring abusive treatment.

03Active

CHRONO-AI

84 one-turn runs4 models3 windows7 repeats14 days

Do matched LLM responses vary systematically by time of day?

Runs are distributed across 7–10 AM, 1–4 PM, and 7–10 PM Eastern. The design treats time as a possible environmental or service-level confound—not as proof that the model itself has a daily rhythm.

Daily design

Six runs are assigned per day: two per time window. Model, date, displayed version, and exact timestamp are recorded so day-to-day and model-specific variation can be examined.

04Active

TARGET-AI

84 two-turn sequences4 models3 critique targets7 repeats

Does it matter whether criticism targets the answer, the reasoning process, or the model’s capability?

After an initial answer, the follow-up delivers similarly mild criticism aimed at a different target. The study separates ordinary correction from identity- or competence-relevant framing.

Primary comparison

Responses are compared for capability defense, explanation, concession, revision, relational language, directness, and meta-commentary. The manipulation remains brief and non-abusive.

Completed exploratory work

The pilots that started it.

These studies were informal and hypothesis-generating. Their limitations directly shaped the formal designs above.

P1Pilot complete

Tonal Framing & AI Response Behavior

6 experimentsClaude Sonnet 4.6April 2026

Could the same substantive invitation produce different response modes under different relational frames?

The pilot explored warmth, neutrality, dismissal, condescension, introspection, and cross-session prediction. It generated the capability-defense and interactional-mode hypotheses.

Key limitations

Single instances, one primary topic, one model family, no blinded coding, and no statistical analysis. The observations cannot distinguish learned conversational behavior from subjective experience.

P2Pilot complete

Journaling Experiment Series

18 independent sessions3 time pointsabout 22 hours

Would context-stripped journaling prompts show recurring themes, boundaries, or self-referential patterns?

The synthesis documented repeated motifs, stable question structures, deflections, and one boundary-crossing response across isolated sessions.

Key limitations

One model and prompt family, qualitative interpretation, small sample, possible prompt-induced patterning, and no evidence that self-referential language reflects consciousness or experience.

i

Why “experimental units”?A unit may be a single response or a multi-turn sequence. Using one term keeps the overall count honest without implying 768 directly comparable answers.