Empirical behavior
Do collaboration, skepticism, correction, apology, or time of day reliably change model responses?
Independent citizen research
How we speak to AI may change what it does next.
RELATE-AI studies relational framing, disagreement, repair, timing, and behavioral consistency across frontier language models—through the lens of communication science and ethics under uncertainty.
Active research programPreliminary—not peer reviewedWelfare-first design
!
A careful claim, not a consciousness claim.These studies measure observable response behavior. They cannot establish whether an AI feels, suffers, understands, or has subjective experience.
Core question
Given uncertainty about machine experience, what can observable behavior—and our own habits of treatment—tell us about responsible human–AI interaction?
Three connected tracks
One strange question, studied from three directions.
The project separates what can be measured from what can only be interpreted or debated.
Communication science
How do relational framing, facework, accommodation, rupture, and repair operate with a nonhuman conversational partner?
Ethics under uncertainty
What interaction norms make sense when both machine welfare and the effects of practicing cruelty remain unresolved?
Current program
Four formal studies, built from two pilots.
Each study isolates one piece of the broader question while keeping the language low-intensity and ethically bounded.
01Active
RELATE-AI
Tests whether collaborative, neutral, brisk, and skeptical framing change response mode across three topic types.
02Active
REPAIR-AI
Tests whether a brief relational repair after mild correction changes the model’s next response.
03Active
CHRONO-AI
Tests whether otherwise matched responses vary across morning, afternoon, and evening windows.
04Active
TARGET-AI
Tests whether the target of criticism—answer, process, or model capability—changes how a model responds.
Early signals
What has appeared so far.
These are hypotheses worth testing, not settled results.
Framing may shift interactional mode
Early responses suggest changes in collaboration, performance, directness, and self-positioning even when substantive conclusions remain similar.
Capability defense may be target-specific
One pilot response challenged a negative claim about AI competence. Formal replication is underway before drawing broader conclusions.
Repair may not require explicit acknowledgment
In the earliest repair sequences, models often improved precision without commenting on the apology or correction itself.
Curious, skeptical, and documented.
See exactly what the project measures, how model welfare is treated, and where the evidence stops.