Maxim (Max) Topaz, Visiting Professor in Health Informatics at University of Maribor, shared on LinkedIn:
“Nature just covered a study claiming that nearly 90 percent of new biomedical papers show signs of AI writing.
I read the preprint. The evidence is words like ‘these‘ and ‘potential.’
The machinery is more careful than that sounds.
The team tracked 379 ordinary English words, fit each word’s frequency trend through 2018 to 2022, extended the line past ChatGPT, and attributed any excess to AI.
‘These‘ appeared in half of December 2025 abstracts; the trend line said a third. Stack hundreds of those gaps into a mixture model and out comes 89 percent.
The trend is real, and I am inside it: I use AI for grammar and polish daily, and this method would flag every paper I have written this year, along with most written by the colleagues I respect.
The precision is softer. The same excess appears if no AI edited anything and three years of reading AI text simply taught humans to write like it.
The authors acknowledge that possibility, assume it is small, and move on. So 89 percent may be the true number, or the sound of everyone’s vocabulary drifting a little.
Either way it measures the wrong thing.
Last week I wrote that detection can tell you a machine made the words but not whether a human stood behind them.
This preprint is the same verdict at the scale of the entire literature: vocabulary shows AI touched almost everything and stays silent on who did the thinking, the only question that decides whether a paper can be trusted.
By this measure, are any of your papers still human?”
Other articles about AI on OncoDaily.