Q002-H2 — Training Does Not Warn About Sycophancy¶
Statement¶
No corporate or government AI training materials warn users about sycophancy or its equivalents because the concept is too new, too technical, or too niche to have reached standard training.
Status¶
Partially supported. No training material was found that warns about sycophancy. However, the reason is not that the concept is too new — extensive academic and industry research exists. The reason is a gap between research awareness and training content.
Supporting Evidence¶
| Evidence | Summary |
|---|---|
| SRC08-E01 | NIST does not name sycophancy as a risk category |
| SRC10-E01 | No legislation targets sycophancy; regulation fragmented |
| SRC04-E01 | Sycophancy incident caught users unprepared |
Contradicting Evidence¶
| Evidence | Summary |
|---|---|
| SRC01-E01 | Sycophancy risk is well-documented in industry analysis |
| SRC03-E01 | Published in Science — hardly "too niche" |
| SRC06-E01 | Formal Bayesian analysis exists |
| SRC07-E01 | Microsoft Research reviewed ~60 papers on overreliance in 2022 |
Reasoning¶
H2 correctly identifies the absence but incorrectly attributes it to novelty. The concept is well-established in research (Science publication, Bayesian analysis, ~60-paper review). The absence from training is better explained by the research-to-practice gap and commercial incentives (H3).
Relationship to Other Hypotheses¶
H2 and H3 both predict the absence of sycophancy from training. They differ on the reason. H3 is better supported because it explains both the absence AND the existence of extensive research.