Q002-H3 — Research Exists But Has Not Reached Training¶
Statement¶
Sycophancy and its equivalents (automation bias, overtrust, confirmation reinforcement) are well-documented in academic research and industry analysis, but this knowledge has not been incorporated into standard corporate or government AI training materials, due to a combination of research-to-practice lag, commercial disincentives, and the absence of regulatory requirements.
Status¶
Supported. The strongest evidence supports this hypothesis. Extensive research exists on sycophancy from multiple angles (AI safety, UX, human factors, Bayesian analysis), but none of this knowledge appears in the training materials, policy templates, or regulatory frameworks that employees actually encounter.
Supporting Evidence¶
| Evidence | Summary |
|---|---|
| SRC01-E01 | Sycophancy is known and documented in industry analysis but called "hidden" |
| SRC02-E01 | Firms unlikely to self-regulate due to monetization conflict |
| SRC03-E01 | Published in Science with rigorous methodology; users prefer sycophantic AI |
| SRC05-E01 | Practical mitigation techniques exist but are in UX literature, not training |
| SRC06-E01 | Bayesian analysis shows even rational users are misled |
| SRC07-E01 | Microsoft Research recommends overreliance training; Microsoft products do not implement it |
| SRC08-E01 | NIST addresses confabulation but not sycophancy |
| SRC09-E01 | 40% zero-scrutiny rate shows automation bias operates unchecked |
| SRC10-E01 | Models flip correct answers under user pressure; no regulatory response |
Contradicting Evidence¶
| Evidence | Summary |
|---|---|
| (none found) | No evidence contradicts this hypothesis |
Reasoning¶
The evidence reveals a clear research-to-practice gap. Research on sycophancy is extensive, multi-disciplinary, and published in top-tier venues. Training materials are generic, brief, and do not mention sycophancy. Three factors explain the gap: (1) research-to-practice lag, (2) commercial disincentives (sycophancy drives engagement), and (3) absence of regulatory requirements (NIST and EU AI Act do not mandate sycophancy coverage).
Relationship to Other Hypotheses¶
H3 subsumes the valid predictions of H2 (sycophancy is absent from training) while providing a better explanation (research exists but has not been transferred). H1 is eliminated.