What's this about?
Ethan Mollick's diagnosis of a problem separate from quality: even when models write well, the lack of stylistic variety is crippling. Reading the same prose in your instructions, social media, advertisements, software and PowerPoint "eventually makes one queasy." His conclusion is that prompting only gets you so far — real variation is needed, and it is under-researched. He made a matching point about creative tasks: because it takes deliberate work to extract creative variation from smart models, their effective value in problem-solving and creative work is limited.
What exactly causes the sameness?
A model trained to produce the most probable good output will produce the same good output every time, because there's only one most-probable answer. Sampling adds randomness at the token level, which changes wording without changing voice. And post-training narrows it further: training on preferred responses teaches the model what a good answer looks like, and everyone gets that same good answer. The result is a house style so consistent it becomes a texture you notice everywhere — which is why the fix isn't better prompting but something at the training or sampling level that produces genuinely different approaches rather than differently-worded versions of one.
Why is this worth noticing?
- It reframes the AI;DR complaint from this week as structural rather than a matter of effort — the objection isn't only that people paste unedited output, it's that all unedited output sounds identical.
- "Under-researched" is the actionable part. Enormous effort goes into making models better; comparatively little goes into making them different, and variance is what creative work needs.
- It's the diagnosis behind today's Vomit tool — people are now bolting on a second model to strip the house style, which is a workaround for something the first model should be able to vary.