Comment on Chatbots Make Terrible Doctors, New Study Finds
XLE@piefed.social 16 hours agoCalculators are programmed to respond deterministically to math questions. You don’t have to feed them a library of math questions and answers for them to function. You don’t have to worry about wrong answers poisoning that data.
On the contrary, LLMs are simply word predictors, and as such, you can poison them with bad data, such as accidental or intentional bias or errors. In other words, that study points to the first step in a vicious negative cycle that we don’t want to occur.
thebazman@sh.itjust.works 14 hours ago
As I said in my comment, the technology they use for these cancer screening tools isnt an LLM, its a completely different technology. Specifically trained on scans to find cancer.
I don’t think it would have the same feedback loop of bad training data because you can easily verify the results. AI tool sees cancer in a scan? Verify with the next test. Pretty easy binary test that won’t be affected by poor doctor performance in reading the same scans.
I’m not a medical professional so I could be off on that chain of events but This technology isn’t an LLM. It suffers from the marketing hype right now where everyone is calling everything AI but its a different technology and has different pros and cons, and different potential failures.
XLE@piefed.social 14 hours ago
You’re definitely right that image processing AI does not work in a linear manner like how text processing does, but the training and inferences are similarly fuzzy and prone to false positives and negatives. (An early AI model incorrectly identified dogs as wolves because they saw a white background and assumed that that was where wolves would be.) And unless the model starts and stays perfect, you need well-trained doctors to fix it, which apparently the model discourages.