Scientists Found a Cheap Trick to Control AI's Writing Style

Ever ask an AI chatbot for something and get an answer that sounds nothing like you? A new study from researchers Ajit Mallavarapu and Ziwei Gu suggests there's a cheap way to find and control those style choices — no expensive retraining required.

The trick is surprisingly low-tech. Instead of teaching the AI anything new, the researchers asked the same prompt over and over with the AI's "creativity" dial turned up, so it produced wildly different-sounding answers. They then ran a standard math shortcut (called PCA, which simply finds the biggest patterns in a pile of data) over the AI's internal signals. That revealed clear "axes" of style: one end formal, the other casual; one end wordy, the other terse. The computer then labels those axes by reading its own extreme answers.

To check whether this was real, they compared the discovered axes against 245 style descriptions written by actual humans. On their best model, the top two axes matched what people spontaneously asked for 72.8% of the time, and three-quarters of human raters said the labels fit the examples. Not perfect, but strong for something requiring zero training data or human labeling.

But here's the honest catch: it depends heavily on which AI you use. Two versions of Qwen and Meta's Llama-3.2-3B showed clear, human-recognizable style axes. DeepSeek-7B-Chat dropped to 35.3% precision — its strongest internal patterns turned out to be about sentence structure, not tone. So you can't assume the trick works everywhere. Still, the finding matters: it hints that style control could soon be a cheap, built-in feature rather than an expensive custom project — meaning AI writing that finally matches your voice, your brand, or your audience without a big engineering bill.