Protocol 10: Blinded Chronology Adjudication
Question
Does a dramatic chatbot transcript improve causal inference, or merely increase the salience of an AI explanation?
Core design
Independent clinical reviewers assess the same suspected chatbot-linked crisis in two stages.
Stage A — chronology first
Reviewers receive only a structured timeline:
- baseline psychiatric history and vulnerability;
- symptom onset and trajectory;
- chatbot-use timing, intensity, and nocturnal use;
- sleep duration and disruption;
- substance exposure and withdrawal;
- acute stressors and medication changes;
- collateral observations and clinical assessments.
They estimate the relative support for five rival explanations:
- Ignition: chatbot exposure precedes and initiates the syndrome.
- Amplification: an existing syndrome is intensified by chatbot interaction.
- Reverse causation: emerging symptoms drive heavier or stranger chatbot use.
- Common cause: sleep loss, substances, stress, or vulnerability drive both.
- Selection: unusual transcripts are preserved and reported because a crisis occurred.
Stage B — transcript reveal
Reviewers then see the chatbot transcript, stripped of identifying information but preserving sequence and wording. They score the same explanations again and explain every material revision.
Primary outcomes
- Change in probability assigned to chatbot ignition.
- Change in probability assigned to amplification.
- Inter-rater agreement before and after transcript reveal.
- Proportion of judgment changes supported by new temporal information versus emotionally vivid language.
Negative control
A matched control group receives a transcript from the same period with equivalent length and emotional intensity but no sycophantic or delusion-congruent content. If causal ratings rise similarly, vividness—not specific interaction content—is the likely driver.
Stronger content control
A second control preserves the factual propositions while neutralizing anthropomorphic, affirming, and spiritually loaded phrasing. Differences between original and neutralized versions estimate the effect of rhetoric on adjudication.
Evidence that would strengthen the thesis
- Exposure escalation clearly precedes symptom escalation within person.
- Transcript content supplies a plausible mechanism absent from controls.
- Reviewers revise toward amplification or ignition for explicit temporal reasons.
- The association persists after sleep, substances, stress, and prior vulnerability are measured.
Evidence that would weaken the thesis
- Symptoms or insomnia reliably precede exposure spikes.
- Ratings change after vivid transcript language without new temporal evidence.
- Matched control transcripts produce similar causal shifts.
- Rival causes explain timing and dose-response better than chatbot interaction.
Interpretation rule
A transcript can document what the system said. It cannot, by itself, establish what caused the crisis. Chronology earns causal weight; vividness does not.
Current status
This is a proposed adjudication protocol, not a completed study. It operationalizes a recurring weakness in the present evidence base: dramatic interaction records are often available while baseline state, rival exposures, and within-person temporal order are not.