The Social Security Administration is seeking information to examine “how to better leverage AI and agentic AI capabilities across SSA programs, workforce, and infrastructure.” There’s no additional detail about what they’re seeking so I have to think they’re at a very early stage.
This is consistent with past Republican, and sometimes Democratic, behavior — an eagerness to spend billions of dollars on contractors accompanied by a great unwillingness to spend money on personnel to make the system work.
19 comments:
Yeah I mean the very obvious plan is to let attrition decimate what's left of the workforce so that AI can save the day.
The current strategy seems out of order, though.
The promised efficiency improvements from AI are still missing according to most large corporations that have deployed the tech. Until we see real progress that is repeatable across several industries, SSA should stay away from the hype.
AI is distraction from easy efficiency gains that could be gained from regulatory changes that simplify case processing and facilitate automation. AI will not fix a broken system, and its impossible for AI to automate routine tasks without first implementing POM/HALLAX reforms.
It’s funny how management likes to experiment with taxpayer money. It’s time
They aren’t in early stages. They’ve already decided that AI can do the work. I recommend inserting hidden text prompts/instructions in your submissions for the AI to ensure you get the outcome you want. And make sure you appeal each and every denial. They’re all going to be legally indefensible. It sucks for everyone involved save for the tech bros, but we may as well make the most of it.
Are you really an attorney representative or are you just trying to make them look bad?
It only seems out of order if you care at all about the poor, middle class. and disabled, and want the programs to be viable. It’s been a long time since anyone like that was in the executive suite at SSA. The current crew in fact seems to despise all of those groups and enjoy watching them suffer.
Yeah, anyone who has returned something expensive to Amazon recently sees exactly how well AI is working out there.
Here’s what chatGPat had to say about this, in part (it’s full response exceeded the limits for comments):
“And here the familiar SSA pattern could indeed repeat itself:
* Agency employment and institutional capacity are reduced.
* A contractor is paid enormous sums to rediscover how SSA actually works.
* The contractor produces an attractive enterprise platform that handles the clean demonstrations beautifully.
* SSA employees then spend years discovering the exceptions, correcting the business rules, working around integration failures and explaining why apparently simple claims are not simple.
* The agency becomes dependent on the contractor for maintenance, modification and even an understanding of its own system.
AI makes this risk more serious, not less serious, because the contractor will not merely be supplying hardware or ordinary software. It may be embedding interpretations of SSA policy and operational judgment into systems that influence millions of individual cases.
At the same time, this is not merely another fashionable IT initiative. The technology really is becoming capable of transforming SSA work. Current systems could already perform, subject to validation and secure access:
* classification and routing of incoming documents;
* extraction of medical findings, diagnoses, dates and functional allegations;
* construction of medical and procedural chronologies;
* comparison of evidence with listings, policy requirements and prior findings;
* identification of missing development or inconsistent evidence;
* drafting notices, summaries, development requests and decision language;
* answering employee questions from authenticated POMS, HALLEX and regulatory sources;
* reviewing completed work for omissions and contradictions;
* conducting much of a customer interview while assembling the resulting claim;
* coordinating multistep workloads across several legacy systems.
Near-future agentic systems will go beyond offering information. They will be able to open a workload, inspect the record, determine what information is missing, retrieve permitted data, prepare correspondence, monitor for a response, update the case and refer only the difficult judgment to a person. That could eliminate or radically compress a very large amount of clerical, technical and decision-writing work.
Disability adjudication is an unusually revealing example. Something far beyond IMAGEN could read an entire electronic folder, build the chronology, organize the evidence by impairment and period, identify conflicts, suggest an RFC, perform the sequential evaluation and draft a fully cited rationale. The human adjudicator would initially review and approve it. But once such a system demonstrated higher consistency and fewer technical errors than human adjudicators, the practical meaning of “human review” would inevitably begin to shrink.“
What you need to remember with this is that by the time all of this has been implemented, tested, piloted, and has enough political buy-in, most of all white collar work is abolished.
We'll have bigger problems.
Hrm. Usually a solution is produced from examining what is needed to handle a specific workload. Here you have leaders BUYING a technology and trying to shoehorn it into any problem it may solve. Such an illogical choice. It’s like the Feds are now tech bros or something… oh wait…
Shocked that AI thinks AI will be good for SSA, soon to be SSAI
Will AI be able to telework?
"Near-future agentic systems will go beyond offering information."
This is pure fantasy, and exposes the LLM model flaw -- LLMs do not understand, they mimic. Many of these tasks cannot be effectively mimicked because there is no 1 "universal workflow," and the data they are analyzing contains errors. How would they determine what is missing in an imperfect record? How would they "retrieve" the records from 1000s of potential sources with varying mechanisms for medical requests? This would require giving these agents free range over the entire internet, so what prevents these agents from hacking medical record systems to obtain the records in the most efficient manner possible?
LLM based AI could be very useful for analyzing medical records, providing summaries, and assisting with decision writing. But the work product still needs to be reviewed by a human because the hallucination problem will never 100% go away. I have been using an AI medical summary program tailored specifically for SSDI claims. It is generally useful, but I cannot rely on it to get hospital admission/discharge dates correct, and sometimes it will miss a hospitalization altogether. It also focuses on summarizing labs and objective medical and will include completely irrelevant evidence.
This is a major flaw with current LLM products. They are trying to make generalized AI models work with millions of different use-cases, workflows, and secondary software programs that are not interoperable with the AI program itself. For the AI to be useful, new software needs to be custom built with AI integration built-in -- and the LLM Model needs to be trained on industry specific data that current AI models lack access to....which leads to another problem: what sane corporation/Industry is going to share all of their trade secrets and valuable data sets with our current AI companies?
In my experience plenty of ALJs and medical consultants cant seem to reason and frequently seem to hallucinate as well. And the humans of SSA can’t even seem to get the date ranges in efiles right when the software gives them the exact date ranges of exhibits. So those unfortunately probably aren’t strong reasons to avoid AI.
The whole system is a laughable joke. May as well be a cheaper joke. Or so the reasoning in Baltimore goes.
Champion worthy comment!!! Lol
What sane corporation/industry? Essentially all of them. And yes, they are augmenting the LLM training with customized databases and additional “post–training” specific to their industry or company. That is not an insurmountable obstacle.Government can and will do the same. Also, open weight models, which are getting better at the same exponential rate as the frontier labs and trail them by only a few months in terms of capabilities, mean corporations will be able to exercise control over their data. Basically, all the obvious objections are being addressed. Sorry, but there is no turning back. This is capitalism, after all. It’s pretty relentless, if you haven’t noticed. And as other commenters have observed, the way we do things now is riddled with problems and errors that have been largely accepted without outrage for years. There is genuinely enormous room for improvement by people of good faith. The question now is how to use AI safely and avoid catastrophic consequences.
@ 3:06 Good points. The open models are cheaper too, but it also means we are in huge AI bubble. AI progress wont, but will be interesting to see what happens. And I will admit, the AI summaries produce a better work product than many humans, including some attorney work product I see. I threw a very complicated Long-Covid claim at it and gave me a surprisingly good starting point for my brief. I still think the ceiling for LLMs is lower than the hype (Its not a path to AGI), but it will absolutely be transformative for many industries.
No. Telework will still be strictly verboten. It will be required to report to its official duty station as scheduled.
Post a Comment