Plan on Failure with AI

Mon, 01 Jan 0001 00:00:00 +0000

Using AI is a great way to improve your systems, but there is one very important aspect to think about: all models fail. Every AI model has some points where it will return either a false positive or a false negative. If you don’t account for this fact, then you will face this in production in very unexpected ways.

To combat this, you should realize and plan on these two failure modes: false positive (it thought it was providing a correct positive answer, but it was wrong), or a false negative (it thought it was providing a correct negative, but it was wrong). Your business use case may be much more acceptable of one type over the other. In the general case of medical testing, a false positive is much more acceptable than a false negative. It’s not great to tell a patient that they might have a diagnosis (when they don’t - false positive), but it is much, much worse to tell a patient that they don’t have a diagnosis when they really do (false negative).

Ai on Stephen Shary

Plan on Failure with AI