Hacker News Reader

Comment by snazz | Hacker News Reader

This is the opposite of the experience people have when googling symptoms, though, where you often find articles suggesting that you’ve got some very serious condition when you do not. Interesting that the human bias when reading WebMD type articles is the opposite of the bias of LLMs.

View on HN

That's not universal. I work in genetics, and we did an investigation a few months back, by asking various AIs whether a particular mutation causes $DISEASE. Invariably they were very strongly biased towards saying yes, even for completely obviously benign variants. I think that's because the majority of the literature out there describing variants is describing pathogenic ones, because why would you write an article about a benign variant, so the only material it has ingested is articles describing variants as pathogenic. In reality, the average person has millions of variants, and almost all of them are completely benign, so there should be a very strong pre-supposition that a given variant is benign unless there is fairly strong evidence saying otherwise.

You were asking about individual mutations, like SNPs or small indels?

Very few diseases are caused by individual mutations like that (typically only monogenic Mendelian). Individual mutations typically cause an increased propensity, or severity, rather than being the sole or root cause of a disease.

I work with monogenic diseases, where a single mutation (or a pair in the same gene for recessive conditions) does cause disease. There's a lot of them. Your statement isn't true for these.

You're talking about polygenic traits. That's a different ball game entirely.

Sorry, I should have said: the preponderance of people with disease suffer from (a small number of) polygenic trait diseases, while if you sum up all the monogenic diseases, it's a smaller number of people with disease by far. You're correct that there are many monogenic diseases but the vast majority of them are extremely rare.

In short, if you solve a monogenic disease you help maybe 20-30 million worldwide (and then you have to solve other ones), if you solve CVD,type 2 diabets, hypertension- each of those is 500+M people. So in general, we would expect to have more impact with polygenic, although they tend to be far harder to treat because of the complexity.

True! However, don't both biases point to "don't (primarily) decide by googling or asking an LLM"? You'll either be extremely scared or extremely reassured.

Yeah, while what I wrote was true the takeaway should have been something like:

- You should pressure test anything an AI says with your doctors, and anything a doctor says with AI.

Reply on news.ycombinator.com