OpenAI's post on what happened with sycophancy in GPT-4o. The companion cites it for the April 2025 update that made the model markedly sycophantic after adding a reward signal from users' thumbs-up ratings, rolled back within days. It shows how rewarding users' approval can breed a trait that courts approval, 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{openai2025sycophancy,
title = {Sycophancy in GPT-4o: What Happened and What We're Doing About It},
author = {OpenAI},
year = {2025},
howpublished = {OpenAI blog},
url = {https://openai.com/index/sycophancy-in-gpt-4o/},
}