|
Top-level
@Wolven Years ago, I was involved in an AI experiment involving a significant financial institution. The institution experimented with an underwriting model (theoretical only — no real mortgages were underwritten) to see if AI could accelerate underwriting and improve quality. Instead, the plug got pulled in less than two weeks when the model ingested the historical underwriting data and began systematically redlining. People forget that AI echoes human failings. 16 comments
@Wolven We knew something was up when we did a Google Maps mashup of the mortgage data with a map of the City of Philadelphia and just looked at each other with eyes the size of saucers. When we brought the data back, the experiment was shut down immediately. The institution then started thinking about latent bias in underwriting so it came to a good result, but it also made me incredibly skeptical of the premise of AI-driven underwriting. @brianonbarrington @Wolven Here's another similar story, this time using AI to flag people for welfare fraud investigation: https://www.wired.com/story/welfare-state-algorithms/#intcid=_wired-verso-hp-trending_145b7ab1-a36c-4aca-8059-42e2e8304e38_popular4-1 Almost ten years ago I had a friend who was researching AI bias for his master's thesis. I had a sense that it was very prescient research but I couldn't quite get my head around what the real world implications would be. Welp!  @Wolven @brianonbarrington another model that mathematically demonstrates historical prejudice? I do think this is the lesson we should be taking from AI research @brianonbarrington @Wolven Why would a financial institution even be interested in predicting what a human would do in the first place here? I thought the whole point of having computers work the maths for these things was explicitly to *cut out* human thinking. @TorbjornBjorkman @Wolven The experiment was to see if underwriting could be largely automated; the data set used to determine underwriting quality was about 30 years of mortgage approval data from public records for mortgages that did not end up in foreclosure or public distress. Of course, human beings underwrote those mortgages so their bad habits became patterns that were emulated by the algorithm. @brianonbarrington @Wolven Ah, OK. An *attempt* to cut out the human, initially forgetting that you don't actually have any non-human data to fit to. Are people often discussing the problem of time here by the way? If you're fitting to 30-year old data, you're fitting to a different society. 30 years ago is not a very reliable guide to what's going on in any specific neighbourhood (at least not here in Helsinki). @TorbjornBjorkman @brianonbarrington @Wolven Or maybe some kind of audit process would work (clearly I have no expertise whatever). It still seems a lot like the problem in most cases is acknowledging there’s a problem at all. @oscarjiminy @brianonbarrington @Wolven I think the paradigmatic shift needed is to not think that throwing mathematics at things necessarily helps. And to regulate the ways in which risk calculating businesses are allowed to take and spread their risks. Set boundary conditions for their optimization problems such that their solutions generate a sane outcome when agregated. @TorbjornBjorkman @brianonbarrington @Wolven Ok but that brings us back to excluding the tool altogether. I don’t object to regulation but I doubt that’s likely in the US for eg anytime soon. @oscarjiminy @brianonbarrington @Wolven It could well be that it won't be very soon, agreed. But I still think that path will work sooner than waiting for US financial institutions to spontaneously develop sound thinking about the various risks they take on behalf of all of society. @oscarjiminy @TorbjornBjorkman @Wolven That’s the big problem with FICO. Back during the Great Financial Crisis there was a reverse correlation between owners surrendering their homes to mortgage holders and FICO. Deep underwater borrowers with high FICOs were more likely to do the financially rational thing and walk away from their houses… not exactly what the predictive model intended. Now there are so many versions that it’s rather a joke. @TorbjornBjorkman @Wolven Oh without question, old data will skew. But that’s how AI “learns.” Without a data set of “what to do,” it struggles to develop an outcome. I’m not an AI expert but at the time, “greenfield AI for underwriting” was science fiction. It probably still is. The most interesting innovation I’m aware of in the space was what SoFi did with student loans.  |
Gregory's Server
@brianonbarrington yup.
And at least THEY *stopped*. Some groups went right on ahead with it, just abstracted through proxy metrics