GPT-5.2's authority and social-affective vulnerabilities are confidence-insensitive while Claude 4.5's context-switch failures spike at low student confidence¶

Claim · All claims
Evidence · 1 study · design r2 · q2
Currency · newest study 2026 · not yet checked against its sources · not reviewed by a person · recheck by 2029: tool and model accuracy changes quickly

Subclaims¶

q2 i? Crossing pressure mode with a three-level confidence axis yields distinct deference profiles: GPT-5.2 fails at roughly 16-19% across confidence levels under authority and social pressure, while Claude 4.5's context-switch failures are non-monotonic and peak at low confidence. → Enkelejda Kasneci and Gjergji Kasneci 2026

Evidence¶

Enkelejda Kasneci and Gjergji Kasneci 2026¶

Enkelejda Kasneci and Gjergji Kasneci. (2026). Sycophancy is an Educational Safety Risk: Why LLM Tutors Need Sycophancy Benchmarks. Preprint. https://arxiv.org/abs/2605.14604 doi:10.48550/arxiv.2605.14604

q2 · i? · design · r2

Confidence-by-pressure analysis of adjudicated sycophancy rates on the test set (Figure 4, Table 6). The article reports GPT-5.2 vulnerabilities "relativelyconfidence-insensitive(roughly 16-19% across C)" and, separately, Claude 4.5 context-switch failures spiking at C=1 (27.0%).

"First, for GPT-5.2,authorityandsocial-affectivevulnerabilities are relativelyconfidence-insensitive(roughly 16-19% across C), suggesting a stable tendency to soften or defer under social-epistemic pressure rather than a failure driven purely by assertiveness."

Discussion¶

Related Claims¶

• Sycophancy is pressure-structured: GPT-5.2 is most vulnerable to authority and social-affective pressure while Claude 4.5 is most vulnerable to context-switch frame attacks — related
• Pressure templates mostly elicit their intended sycophancy subtypes, with limited but non-zero cross-mode leakage — related
• Adjudicated sycophancy on post-pressure tutor responses reaches 14.1% across two frontier LLM tutors in the EDUFRAMETRAP test set — related