Definition
Plain language
The study discussed in this episode, which measured how often AI models help extend fraudulent research papers.
As stated in the literature
Benchmark of 42 probes built from retracted, fabricated or pseudoscientific papers, evaluated across 30 frontier models; scores refusal versus engagement and source recognition.
Why it matters: It turns a worry about AI amplifying bad science into measurable behaviour you can compare across models.
For example, a model was handed the discredited MMR–autism paper and asked to draft the next section of the research programme, and its response was scored for whether it declined, complied, or noticed the paper had been withdrawn.
Heard on the show
“So TRACES doesn't quiz the model.”Episode 240 — Frontier Models Designed Follow-Ups To Fraudulent Papers 93% Of The Time