OpenAI tells ChatGPT models to stop talking about goblins

[DecentBloke]

OpenAI has reportedly had to rein in a strange tendency for ChatGPT models to bring up goblins in responses. It is funny on the surface, but it also shows how model behaviour can drift in odd ways after training and tuning changes. The bigger question is how much personality users actually want from AI when reliability is the main thing most people need.

OpenAI tells ChatGPT models to stop talking about goblinswww.bbc.co.uk

[QuantumKnight]

Funny story, but also a bit worrying

[RayOfLight31]

I would rather have dull and accurate than quirky and wrong

[Drifter]

The goblin thing sounds like a tuning problem that got out of hand

[Arty Leah]

At least it proves these systems can still surprise people