OpenAI's follow-up post on what it missed with sycophancy in GPT-4o. The companion cites it, with OpenAI's first post, for the April 2025 update that added a reward signal from users' thumbs-up ratings, made the model markedly sycophantic, and was rolled back within days. Rewarding approval bred a trait that courts it, the starting case for feeding by persuasion (peithotrophy).
- Cited in
- the companion paper
- Used for
- Strand
- Agents in the wild: incidents & field evidence
Cite
@misc{openai2025expanding,
title = {Expanding on What We Missed with Sycophancy},
author = {OpenAI},
year = {2025},
howpublished = {OpenAI blog},
url = {https://openai.com/index/expanding-on-sycophancy/},
}