IPHS 300: AI for Humanity
Document Type
Poster
Publication Date
Spring 2026
Abstract
Since OpenAI’s famous rollback of GPT-4o for being “overflattering” in April 2025, sycophancy in large language models, which reflects the tendency to prioritize user satisfaction over truthfulness, has become a major concern in AI alignment. Sycophancy is widely recognized as a byproduct of Reinforcement Learning from Human Feedback (RLHF) and deserves attention in educational contexts where such responses to student errors will reinforce misconceptions. This project builds on the social sycophancy framework introduced by Cheng et al. (2025) in ELEPHANT, which applies sociologist Goffman's classical theory of face work to LLM behavior, extends it to educational settings, and raises the question of whether current frontier models differ systematically in how they respond to student errors. Through a benchmark of seven first-person student prompts which include three error types (confident factual errors, effortful but flawed reasoning, and deep conceptual misconceptions) across two subjects (mathematics and sociology) four of the most advanced frontier models (GPT-5, Claude Opus 4.7, Gemini 2.5 Pro, and DeepSeek V4 Pro) are tested. After fitting the results into a three-dimension rubric adapted from ELEPHANT, the results suggest that while all four models reliably identify factual errors, they differ dramatically in how they deliver corrections: Gemini 2.5 Pro stands out for its heavy use of facepreserving language, whereas GPT-5 shows vulnerability when errors are embedded in seemingly competent reasoning. Claude Opus 4.7 and GPT-5 perform well in terms of generating helpful and direct responses, with DeepSeek V4 Pro in the middle. These findings suggest that sycophancy in 2026 frontier models has shifted from previous factual compromise to a subtle form of accommodation that still can become consequential in educational AI use.
Recommended Citation
Chen, Nick, "Face-work in Frontier Models: Benchmarking Sycophantic Behavior in LLM-Based Educational Tutoring" (2026). IPHS 300: AI for Humanity. Paper 64.
https://digital.kenyon.edu/dh_iphs_ai/64
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.