I like the MIT AI Risk Repository. It collects AI risks from existing frameworks into one place, and it is valuable.
But I do have one criticism, which is that it is pretty thin in the human factors category. So are most of the frameworks it draws on.
I think I understand why. Risk frameworks highlight categories that could directly lead to harm. They leave out the ones that might just create a bad system (doesn’t satisfy business objectives well) but doesn’t actively harm humans. It may be hard to label these as risks. But my point is that it’s not necessarily a dichotomy but a spectrum.
For example, yesterday I wrote about latency. It’s a fundamental characteristic of both UI/UX and information systems. But AI can introduce latency in systems in a variety of ways, many of which may be undesired, but harmless, but some may be harmful. AI, as it is opaque, may also have its inputs clouded or overridden by the latency in the system.
Factors like latency will arise in new systems in unexpected ways, and therein lies the danger.