[lotsofpulp]

That is not how the US has chosen to do it. The ACA prohibits using anything other than age, location, and tobacco use for premium pricing, and the ACA prohibits denying coverage (resulting in a wealth transfer from healthy to sick).

Even the age rating factor is capped at 3, so there are also massive wealth transfers from young to old.

Mathematically, health insurance premiums in the US are more tax than insurance premium.

[trollbridge]

The ACA doesn't prevent some magical fairy-dust AI from pricing premiums, though, which is currently all the rage amongst insurers. (Not because AI will be accurate or anything, but because it offers a completely opaque pricing process.)

[avalys]

Why not? If the only factors allowed to be used in settings premiums are age, location and smoking status, then those are the only parameters that could be input to an AI model, no?

[kyboren]

Warning: I am not in this industry and the below is speculation:

AIUI the idea is to predict the "correct" price for an individual premium, Y, which is restricted to being the result of some insurance model function, f(), that is itself restricted to the domain of age (A), location (L), and smoking status (S):

  Y = f(A, L, S)

My impression was that the idea was that this would handicap insurers' natural desire to price premiums individually and have a smoothing effect on prices over the population.

But why is location useful for insurers to price premiums? I assume because healthcare has different costs in different locations, and different utilization rates: People living in coal mining towns or near oil refineries may be expected to use more healthcare and thus cost more to insure.

Thus, you can imagine insurers building a price map (like a heat map) overlay for the state/country, plotting each applicant on it, and checking the "color" there as part of their model function. So they are effectively embedding out-of-band information (prices and utilization rates for locations) into the model function f() itself.

What "AI", or large-parameter Deep Neural Networks, fundamentally change here is:

   - They can approximate any effectively computable function, the same class of functions that a programmer can write and execute on a computer[0].

   - They can be scaled up to an arbitrarily large number of parameters, i.e. an arbitrarily precise approximation limited only by the number of parameters and the amount and quality of model training data.

   - Absolutely critically: They are programmed implicitly, through brute-force training on input-output pairs, rather than explicitly, with some programmer writing an explicit series of instructions to compute the function.

This last point is probably the most important.

Large insurers previously had sophisticated models for f() hand-built by math whizzes, but they were limited in the amount of out-of-band information they could encode into that function by the limited cognitive and programmatic capacity of a human team.

But now with DNNs you can scalably encode unlimited out-of-band information into the function f(), while also totally obscuring how you're computing that location-based price adjustment.

The result, in extremis, is that f() is not some fancy algorithm cooked up by a few dozen math whizzes. Instead f() becomes a fancy database, allowing the tuple (A, L, S) to act merely as an index to an individualized premium Y, which defeats the entire purpose of restricting the domain of their model function.

[0]: https://en.wikipedia.org/wiki/Universal_approximation_theore...

[avalys]

That’s not how courts and laws work. If you cheat and use other illegal factors to compute a premium for each person, and then create an AI model that effectively looks up your illegally calculated premium of each person by their location, they’re going to reach the obvious conclusion - you are calculating a premium using illegal factors.

[kyboren]

Like I said: It's only useful to have location in your model's domain if you have side-channel information embedded in the model function itself about what those location data mean for the correct premium price. What we're talking about here is just a way to embed much more information in your model function than a human reasonably could.

Given the magnificent ability of DNN models to serve as obfuscatory black boxes and the general techno-ignorance of legislators and regulators, I suspect that "AI laundering" your violations actually a very effective way to juke all sorts of laws and regulations.

But both of us are just speculating. If you have insider industry knowledge or can point to regulatory guidance and/or enforcement actions in this area that corrects or confirms my understanding, I would love to read about it.