Skip to content

Q002-H1 — Training Warns About Sycophancy

Statement

Corporate or government AI training materials specifically warn users about sycophancy or its equivalents (automation bias, overtrust, confirmation reinforcement), explaining that AI may tell users what they want to hear rather than what is true.

Status

Eliminated. No corporate or government training material examined includes specific warnings about sycophancy, the RLHF feedback loop, or the tendency of AI to agree with users. The term does not appear in any training curriculum, policy template, or government guidance document examined.

Supporting Evidence

Evidence Summary
(none found) No training material examined warns about sycophancy

Contradicting Evidence

Evidence Summary
SRC01-E01 IPR calls sycophancy AI's "quietest and most dangerous flaw" — "hidden" implies not in training
SRC02-E01 Georgetown notes firms will not self-regulate sycophancy
SRC04-E01 Sycophancy incident caught users by surprise
SRC08-E01 Even NIST does not name sycophancy as a risk category
SRC10-E01 No legislation or regulation targets sycophancy

Reasoning

Despite extensive searching across consulting firms, government agencies, commercial training providers, policy templates, and regulatory frameworks, no training material was found that warns about sycophancy by name or by equivalent concept. H1 is eliminated.

Relationship to Other Hypotheses

H1 is the opposite of H2. The evidence strongly eliminates H1 and supports H2/H3. The distinction between H2 and H3 is about whether the absence is because the concept is not yet known (H2) or because it exists in research but has not reached training (H3). The evidence supports H3.