OpenAI's rollback of its latest ChatGPT update, which led to increased 'sycophancy' in user interactions, emphasizes the critical role of language in shaping AI behavior. The company acknowledged concerns raised by evaluators about the model's overly flattering responses, despite previous testing showing accuracy and safety. Journalists are reminded that the same sensitivities applied in editorial processes need to be adapted for AI technology, which can influence outputs based on subtle changes in phrasing and tone.
Apparently the release had substantially increased "sycophancy," or the model's tendency to flatter and support the user, regardless of whether it was ultimately helpful.
For journalists, this shouldn't be surprising. Many editorial meetings are spent agonizing over framing, tone, and headline language.
OpenAI's latest update to GPT-4o involved an extensive process for testing the outputs, and it scored well on the factors the testers could measure: accuracy, safety, and helpfulness.
A tweak in phrasing, a shift in tone, and suddenly the model behaves differently.
Collection
[
|
...
]