[alancalverd]

A cautionary tale from my first (and probably last) professional encounter with AI.

I've been asked to estimate the risk to a patient from a research program that involves four CT examinations of the skull. Not a problem as we have standard conversion factors from scan parameters → absorbed dose → effective dose → risk.

If you ask Bing for "dose from a skull CT examination" it returns 0.07 millisievert, one thirtieth of the correct value, based on its AI search.

The reason seems to be that it hasn't distinguished between computed tomography and 2D ("film") x-ray imaging, which does indeed result in the lower dose.

So if you aren't an expert, don't rely on AI, and if you are, don't waste your time with it!

[Bored chemist]

Is AI a useful guide to anything (yet)?

[alancalverd]

"Computer says yes." Else what is the point?

I think this underlines the point I was making in a discussion with HY elsewhere in this forum: it doesn't matter where you got the advice from, ultimately every decision is made by a human, for the benefit of a human. Hence the need for a signature on the risk documentation, and a neat phrase culled from an aircraft technical log book: "could you defend your signature at an inquest?"   

[Bored chemist]

"could you defend your signature at an inquest?"   

That's quite a common maxim.

[vhfpmr]

"could you defend your signature at an inquest?"

If I ask whether it's worth spending ?50k assessing the risk of a ?500 radio killing one in a million users, the answer from the boss will be no, and the answer from the barrister with the benefit of hindsight, yes.