Do LLMs have different personalities?

The article suggests that as users choose certain behaviors and engineers introduce them in the next LLM generations, an LLM company's commercial product becomes biased toward those population prejudices.

"Technically, personality drift is not consciousness but tone alignment: the sum of phrasing, affect, and response framing learned during RLHF and interface fine-tuning. When millions of users reward a particular conversational energy, the model weights that energy higher in subsequent updates."

I guess the cumulative effect might become significant in a decade, to the point one LLM user may feel more at ease with some LLM products than others.

They have different characteristics but it's a category error to describe those as "personalities."