AI flags more than 250,000 suspicious cancer research papers

Remember when AI was supposed to cure cancer? Instead, it's been busy catching the people faking it. A machine learning system just chewed through 2.6 million cancer papers and flagged over 250,000 as suspicious — roughly one in ten. That's an impressive detection rate, but it leaves a rather uncomfortable question: what exactly has been informing oncology decisions for the past decade? Because if even a fraction of those 250,000 are paper-mill products, the evidence base for some treatments might be built on sand. And the best part? The system identifies "writing patterns similar to those from fraudulent paper mills" — so the AIs are busting humans for writing like AIs. The irony is almost too delicious.

Of course, nobody is rushing to retract those 250,000 papers. That would require admitting a systemic failure that publishers, universities, and funders have all enabled. Instead, we'll likely see a new arms race: paper mills will train their own AIs to produce texts that pass this detector, while the detection AI gets better, and the cycle continues. Meanwhile, legitimate researchers — the ones who actually run experiments and peer review — will have their work crowded out by fabricated data. The machine learning system is a clever piece of engineering, but it's a band-aid on a haemorrhaging wound.

Worse still, this finding comes at a time when AI itself is being integrated into medical diagnosis and treatment planning. If the training data includes fraudulent cancer studies, then the models we're building today are being poisoned from the start. So congratulations to the researchers who built this detection tool — you've done a heroic job of documenting the wreckage. Now if only someone would do something about it.