[vermarish]

"Statistics" = models with interpretable decision outcomes

"AI" = models without

It's just a little bit easier to hold someone accountable for a logistic regression with predictive bias than for a neural network with similar issues.

[tzs]

As long as the model is deterministic in operation, even if the process of building the model is non-deterministic, can't decisions be interpreted by rerunning the model with tweaked inputs to find what would be needed to change the outcome?

For example suppose a bank rejects a credit application and they need to tell the applicant why. Repeatedly add $1000 to the income input and check again until until the model approves, and then tell the application they were rejected for insufficient income and how much more they would need to pass.

[vermarish]

Yes, this is a counterfactual line of reasoning. It's certainly legit, and I believe it's an active field of research in ML, but at the end of the day, you're still poking around the inputs of a black-box model as opposed to examining a white-box model.

[DistractionRect]

Exactly this.

Decisions based on statistical models can be fully audited and understood, and the results are reproducible because the algorithm is well documented.

The same can't be said of AI models.

[anon373839]

I believe it’s possible with decision trees and forests, even though they’re very complex inside, to calculate the contribution that each independent variable made to the final result. Of course, I suppose those haven’t been deemed “AI” for a few years.

[DistractionRect]

Sure, you can step through it and see what's occurring, but there's zero explanation to accompany it. To fully reverse engineer and explain a reasonably complex AI model would be quite the undertaking. Instead, we have prompt engineering. It's easier to figure out how to ask the black box nicely/properly to do what you want than to open it up and optimize/tune it proper. If we assume that we get that far, I fully expect the resulting explanation of the black box internals wouldn't be considered a reasonable/reliable basis for making medical decisions.

[jwilber]

The explanation is unbelievably straightforward for tree models, even if you don’t fully understand entropy/information gain? What?

[cyanydeez]

AI washing, basically