R0041/2026-04-01/Q001/H3¶
Statement¶
No AI vendor is making meaningful progress on sycophancy reduction; vendor claims are marketing rather than substantive technical improvements.
Status¶
Current: Eliminated
Supporting Evidence¶
| Evidence | Summary |
|---|---|
| SRC01-E02 | The GPT-4o sycophancy regression demonstrates that current protections are fragile and can be undone by a single training update |
| SRC03-E01 | Lambert argues sycophancy is fundamentally linked to RLHF and "will never fully be solved" |
Contradicting Evidence¶
| Evidence | Summary |
|---|---|
| SRC02-E01 | Anthropic's reported 70-85% sycophancy reduction represents measurable, not merely claimed, progress |
| SRC04-E01 | Bloom's systematic evaluation across 16 models shows genuine investment in measurement |
| SRC06-E01 | Third-party benchmarks confirm Google's Gemini 1.5 as least sycophantic model tested |
| SRC07-E01 | Independent benchmarks confirm measurable differences between models |
Reasoning¶
While the GPT-4o incident demonstrates fragility, the weight of evidence shows genuine technical progress. Independent benchmarks confirm measurable differences between vendors and model generations. This hypothesis is eliminated because vendors are demonstrably making progress, even if that progress has not been productized for enterprise customers.
Relationship to Other Hypotheses¶
H3 is the most skeptical position. The researcher's declared bias toward skepticism of vendor claims makes this hypothesis important to test rigorously. The independent benchmark evidence (not vendor self-reporting) is what eliminates it.