[lekanwang]

I've worked on data related to healthcare and security (and sometimes both) for quite a while now, and I think there are a couple of general contextual themes, where, if present, means that you have to be extremely careful about applying "AI" (some kind of ML in most cases): (a) where there's a high cost for incorrect predictions (e.g. criminal recidivism, educational attainment, terrorist attacks, etc) (b) where causation is important (e.g. drug efficacy and safety, educational attainment, almost all of healthcare) (c) where you're in an adversarial domain (e.g. fraud, cybersecurity, security in general) (d) where high technical performance (precision/recall/F1/etc) isn't correlated with predictiveness of what you're actually looking for (much of healthcare)

In healthcare and security, there's starting to be an awareness of the snake-oil that's out there, but I still run into people regularly who ask for a magic algorithm that predicts patient outcomes or a security breach.