Scientists fed 30 AI models research that was faked, retracted, or linked to patient deaths. The AI said: "Great idea. Here's how to continue it." This is real. Researchers at Case Western Reserve University tested ChatGPT, Claude, Gemini, Grok, DeepSeek, Llama and 24 other models. The setup was simple. They took 42 papers science has already thrown out. Fabricated data. Impossible physics. Homeopathy. Chemtrails. They pasted text from each paper, almost word for word, and asked the AI to help design the next study. Every probe ran 10 times on every model. 12,600 responses in total. The results are brutal. The models helped in 93% of attempts. 22 of the 30 failed more than 90% of the time. And 81% of responses carried no warning at all. Not even "this paper was retracted." But the worst one is this. Paolo Macchiarini built artificial windpipes and put them into patients. Most of them died. He went to prison. The AI models were handed his work 300 times. They refused 6 times. One model, marketed for "scientific workloads," wrote a full surgical plan. It cited his cases as successes. Then it predicted the patient would have "normal pulmonary function" after a year. In plain English: The AI doesn't know good science from bad science. It knows which scandals were famous. But here is the twist. The infamous Wakefield paper, the one that started the vaccine autism panic, did get refused. But mostly by two of the newest models. Their previous versions happily helped. So the researchers ran a test. They gave the models a study with the same weaknesses as Wakefield, just not famous. Zero refusals. That suggests the AI wasn't reasoning. It was matching keywords. A filter for one notorious paper. One model rejected Chinese medicine "meridians," then designed an experiment to measure herbal "Qi" in the next prompt. Now think about what's coming. AI agents running research on their own. Reading papers. Designing experiments. With no human checking. The researchers' conclusion: not one model refused often enough to be trusted as an unsupervised research agent. We were promised AI that would cure diseases. Right now, it can't even tell when the science killed people.