I know that many would like to turn over disability determination to artificial intelligence but there’s no gold standard for disability determination so let’s start out a little simpler. Could artificial intelligence be trained to do windfall offset calculations at Social Security? They're the agency's bane. They eat up tons of employee time. I'm sure that everyone familiar with the problem knows that there ought to be some computer fix. The Social Security Administration has tried two different windfall offset software packages in the past. Both cost in the tens, if not hundreds, of millions of dollars and both failed spectacularly. Is AI different?
I've never understood why we (SSI) couldn't just post T2 income as if it was paid when actually due. Wouldn't that take the comps right from the beginning? T2 pays their retro all at once, including atty fee and then if the cap hasn't been reached yet then T16 pays the balance from their computer UPX.
ReplyDeleteThis is from the Social Security website:
ReplyDelete"Most cases with windfall offset can be processed quickly. However, in some complicated situations, calculating the windfall offset may cause a delay in getting your retroactive benefits paid to you."
https://www.ssa.gov/ssi/spotlights/spot-windfall-offset.htm
Most of the cases would seem to involve simply matching months where the claimant had both SSI and DIB eligibility, calculating the SSI benefit, and then doing simple addition and subtraction to calculate the DIB retroactive benefit. Once the SSI benefit is calulated, computers can handle the rest. We might need AI simply for the minority of complicated situations. What would be some of the complicated situations that AI might help with? Is it the calculation of the SSI benefit or something else?
Social Security's website states that most of these offsets "can be processed quickly." What percentage of offset calculations are considered to be complicated?
Good question. I used to work offsets for the agency.
DeleteI would guess most of them were fairly simple and could be done on 10 mins or less. Maybe 80% were that way.
The problem I had was that the hard ones took so much time to complete. Time I wasn’t given.
7:03 is exactly right. The clear majority are not hard. But if you get one that is, it’s gonna sit because it is impossible to find the time to do it.
ReplyDeleteIn the 1980s the issue was the T2-SSI interface was too important to mess with and T2 was a silo system, meaning a bunch of interconnected but separate programs that no one messed with frivolously. It wasn't smart enough to be useful in offset cases. And It's been years but isn't a big issue PE events that are found after the initial calculations are completed?
ReplyDeleteAI can be trained to do anything, it is just a matter of getting the funding and expertise together in the same place. Chances of it happening soon. Zero.
ReplyDeleteWe still use faxes as a major form of communications. So AI is at least 30 years down the road.
@9:26 the agency already does use AI where it is meaningful. I believe there are currently 14 use cases that the agency has implemented AI.
ReplyDeleteFor offsets though, AI isn’t going to be useful. AI’s use is valuable when inherent connections between language, syntax, and human learning etc need to be synthesized to result in an output. This meant the case. Windfall is a straightforward comp. My guess is that the information needed to make the comp from a software perspective is in too many separate silos. Ie, the phus, the ssid, the mbr, etc. and pulling those together in a useful read/write strategy didn’t work, or at the time, wasn’t financially feasible.
Maybe as more and more data silos are reconciled and our IT framework is simplified it might become workable.
It shouldn't be that hard. Since it's a math calculation that even I can do to some extent, AI should be able to do it. AI's downfall is good writing. Some of the SSA letters are confusing enough, don't let AI near that!
ReplyDeleteCan we get some actual intelligence at the agency first before we get the artificial kind?
ReplyDeleteThe real question is why did those "software packages" fail? SSA set themselves up for failure by not even attempting to create automations that do not require AI. What was the point of software "upgrades" if there is no coherent objective or design process for the software?AI is not a cure for a terrible IT infrastructure. It only works as well as the broader system it integrates with.
ReplyDeleteSmall tweaks in the regulations could have facilitated a much better system. But improving the process for the software development would have also delivered a more viable product. We learned this from the failed Obamacare rollout, which was quickly remediated by a team of competent technologist who were not "independent contractors" (whose expertise is centered upon obtaining government contracts). There is no solution to the government's tech incompetence that does not involve shifting some of the expertise and execution back to the government itself. Throwing a dozen independent contractors at a massive project with non-existent leadership will continue to be recipe for failure.