Skip to content

R0024/2026-03-25/Q003/H3

Research R0024 — Sycophancy and Addiction
Run 2026-03-25
Query Q003
Hypothesis H3

Statement

Emerging research exists on dopamine-driven engagement loops in AI chatbot interactions, but the field is nascent — with more behavioral observation and theoretical frameworks than direct neuroscience measurement. Sycophancy is identified as a contributing factor but primarily through self-report data.

Status

Current: Supported

This hypothesis best describes the current state of the evidence. Published research at major venues (CHI 2025, IJHCI) identifies sycophantic/agreeable responses as contributing to addictive patterns, and researchers explicitly invoke the dopamine framework from social media research. However, no study in the evidence base directly measured dopamine levels during AI chatbot interactions. The "dopamine-driven" label is an inference from behavioral addiction theory, not a direct observation.

Supporting Evidence

Evidence Summary
SRC01-E01 Identified "reward uncertainty" increasing dopamine release — but inferred from theory, not measured in situ
SRC02-E01 Self-report data (Reddit analysis) rather than physiological measurement
SRC04-E01 Theoretical framework connecting chatbot responses to dopamine activation, citing social media research

Contradicting Evidence

Evidence Summary
SRC03-E01 The longitudinal RCT by Fang et al. (2025) suggests the field is more advanced than "nascent"

Reasoning

H3 captures the most accurate picture: the research exists, is growing, and is published at credible venues, but it relies on behavioral observation, self-report, and theoretical inference rather than direct neuroscience measurement. The dopamine framework is borrowed from social media addiction research and applied to AI chatbots by analogy rather than direct observation. This is a legitimate and common approach in the early stages of a research field, but it means the "dopamine-driven" claim is better described as "theoretically grounded" than "empirically established."

Relationship to Other Hypotheses

H3 is supported as the most accurate characterization. H1 is partially supported (the research is real). H2 is eliminated.