Recognising AI Sycophancy in Irish Higher Education: A Cross-Sectional Vignette Study of Verification, Critical Evaluation and AI Dependence

Abstract

Generative artificial intelligence is now used across many academic tasks, but frequent use does not show whether students can recognise output that agrees with them when stronger evidence calls for qualification or correction. This study investigated AI sycophancy, verification behaviour, self-reported evaluation and AI dependence in a cross-sectional survey of 180 students from six Irish higher education institutions. The instrument combined use measures, five multi-item scales and six paired vignettes in which students selected between a sycophantic response and an evidence-sensitive response. Academic AI use was frequent: 90.6% of participants reported using generative AI at least weekly. Only 23.9% had previously heard the term AI sycophancy, although 67.2% reported recognising the experience after a neutral definition. Mean vignette recognition was 3.58 of 6, equivalent to 59.7%. Across all vignette decisions, 31.5% of selections preferred the sycophantic response and 32.1% indicated an intention to use it. Verification was positively associated with vignette recognition (r = 0.52, 95% CI [0.40, 0.62]), while affirming-response preference was associated with AI dependence (r = 0.33, 95% CI [0.19, 0.45]). AI-use intensity was associated with dependence (r = 0.48) but not with vignette recognition (r = -0.03). Self-reported critical evaluation exceeded vignette performance by 10.19 percentage points on a common 0 to 100 scale, although the two measures are not psychometrically equivalent. The findings point to a calibration and verification problem rather than evidence of cognitive decline. Because the sample was purposive and the instrument was newly developed, the results should not be interpreted as national prevalence estimates or causal effects.