I spent part of the last two days (this was 2020) “at” the Ethical Innovation for Artificial Intelligence conference. Learned some stuff and it got me thinking; the most important bottom line on such things. Thanks to the organizers (at our end) Rohan Alexander and Kelly Lyon and all the speakers.
My own “bias” up front: in general, I find most discussions of ethical AI and algorithmic fairness too pious by half. I get frustrated by “AI Particularism” and “Descriptive Vigilance.”
Most AI in the wild is product. Product is generated by engineering and marketing.
Here’s my naive take. At times it felt like we were looking in the dark for something that might not be there. IF ML can be characterized as discovering a function that maps a bunch of data to a label output range, we seem to be saying “if there’s unfairness in the world that produces the data, what can you do?” One answer is clean up the data – minimize the bias-in/bias-out problem; one answer is to collect data on how what we mean by fair and unbiased and model that and then add that to the machine (an unfairness detector or a fairness injector); another one answer is be aware and beware how you use it?
But in this mix is a conflation of a philosopher’s fairness and a sociologist’s fairness. The former is articulate-able, specifiable, arguable. The latter is more dynamic, emergent, historically contingent. At any point in social time and social space a group can say “yeah, but we can agree on X, Y, and Z” (like the philosophers) but sociologically speaking fair is what is inter- and extra-personally deemed fair in a given community. Unless you can design an algorithm that can be a member of a community, I don’t think you can build the fairness detector or fairness injector.
Perhaps it was discussed in a session I had to miss, but I’m looking for the big picture that is not bolted to the contemporary world and particular perspectives on that world. Where’s the 10,000 meter picture?
I keep hearing this fundamental tension: humans are humans’ problem because most of them are too cognitively limited to make smart decisions vs. humans are humans’ problem because most of them are just out for their own benefit. Then we bring AI into the mix and the ways it is good or bad depend on which of these one leans toward.
Interestingly, to me at least, is that you can pair this dichotomy with two perspectives on AI: AI as promise and AI as threat.
| Stupid Humans | Evil Humans | |
| AI Good | AI will solve all sorts of previously unsolvable problems and do a much better job of keeping the trains running on time and educating people and so on and so forth. | AI will replace ineffective social control with more absolute and even handed enforcement yielding higher levels of social order |
| AI Bad | For AI manipulating humans will be child’s play. Humans will lose their jobs as decision makers. Bad decisions will result without the benefit of contextual information that is important to humans. | AI will be a tool in the hands of the powerful. It will amplify all the bad tendencies that humans have. Computation (the algorithm said…) will replace traditional sources of legitimacy. |
You might have other things that go in the boxes; the point is not to quibble about that. Rather, I want to suggest that in this particular realm, AI doesn’t change anything, it’s not a new thing. And sometimes I think that we are being held back because too many of us think that it is.
Innoeduvation 