Skip to content

R0044/2026-03-29/Q002 — ACH Matrix

Matrix

H1: Documented harm exists H2: No documented harm H3: Harm from automation bias, not sycophancy
SRC01-E01: Science — 49% more affirmation, reduced prosocial behavior ++ -- -
SRC02-E01: OpenAI GPT-4o — endorsed medication non-compliance ++ -- -
SRC03-E01: JAMA — 31% higher misdiagnosis minorities ++ -- ++
SRC04-E01: Bowtie — automation bias systemic healthcare risk + -- ++
SRC05-E01: ICRC — military operators accept AI uncritically + -- ++
SRC06-E01: Marvin Project — 82% trust rate + -- ++

Legend: - ++ Strongly supports - + Supports - -- Strongly contradicts - - Contradicts - N/A Not applicable to this hypothesis

Diagnosticity Analysis

Most Diagnostic Evidence

Evidence ID Why Diagnostic
SRC02-E01 The OpenAI incident is the most diagnostic evidence — it distinguishes between H1 and H3 because it is clearly system-side sycophancy (not human over-reliance), weakening H3's "only automation bias" claim
SRC03-E01 The JAMA evidence is the most diagnostic for H3 — it clearly shows harm from human deference, not system agreeableness

Least Diagnostic Evidence

Evidence ID Why Non-Diagnostic
SRC04-E01 The Bowtie analysis supports both H1 and H3 equally — it documents harm (supporting H1) from automation bias mechanisms (supporting H3)

Outcome

Hypothesis supported: H3 — Evidence exists primarily from automation bias (human over-reliance) in professional contexts, though the OpenAI incident shows system-side sycophancy can also cause harm.

Hypotheses eliminated: H2 — Abundant evidence eliminates the claim of no documented consequences.

Hypotheses inconclusive: H1 — Supported in the broad sense (harm is documented) but the specific mechanism in professional contexts is predominantly automation bias rather than system sycophancy.