aicarehab

Lab

No results are published. No data has been collected.

This page describes a research program. Nothing on it is an empirical claim. All numbers in the linked documents are design parameters or assumptions, not findings. No claim is made about AI inner states, and no clinical claim is made about people or models. "Care" and "rehab" are engineering metaphors for recovering task state.

Four strands

  1. De-escalation mirror. Whether a calm agent reply changes how hostile the next message is.
  2. Sycophancy versus constructive dissent. Whether agents that agree with everything do worse on task outcomes than agents that disagree politely and say why.
  3. Interpersonal communication drift. Whether the way a conversation sounds changes over a long session.
  4. Bilateral recovery. Whether a reset of the agent's state together with an optional pause for the person changes later task outcomes.

Status of the preregistration drafts

Both documents are DRAFTS. They describe a design and nothing more. They have not been frozen, registered or reviewed.

Consent and privacy stance

Conflict of interest

We are testing the effects of our own product concept. A positive or negative result would matter to us commercially, which is why the plans are written down before any data exists and why null results are listed as reportable outcomes.

Teams and Lab tier

A "Teams and Lab" tier is proposed. It is not available, not sold, has no pricing and no waitlist. See the tiers on the Duos page.