From Nextgov/FCW:
The Social Security Administration is introducing a new chatbot for employees powered by generative artificial intelligence, according to an internal email obtained by Nextgov/FCW.
“This initiative aligns with our commitment to leverage innovative technology to improve efficiency, support our mission and provide a secured way to use GenAI,” read a Wednesday email to staff announcing the new “Agency Support Companion.” …
SSA told staff that the new chatbot is meant to help with content creation and summarization, as well as research and coding assistance. Staff have to watch a four-minute video with guidelines and best practices to access it.
The new tool uses an OpenAI model and wasn’t trained or fine-tuned on any SSA data, according to an agency FAQ document obtained by Nextgov/FCW. No testing was done internally to assess the application against SSA-specific questions, it said.
The chatbot was trained on data up to October 2023, according to the FAQ, which also warned that employees should validate its responses and not feed it personally identifiable information. …
Sounds a little vague on how this is to be used. Could it turn out to be a solution in search of a problem?
34 comments:
Complete waste of time and money. Doesn't help frontline staff in any way, and doesn't do anything that ChatGPT can't do for free already. This is right up there with the CCE "no to all button" for pointless achievements by SSA executives.
I imagine the long-term goal will be to get one that does have access to SSA systems, that they'll try to train on staff to eventually replace as many of them as they can. All while lying through their teeth about it being "helpful".
You aren’t allowed to enter PII, which makes it useless for 99% of the real work being done. Oh, and they forgot to mention that in the training video, guaranteeing this is going to be a source of PII leaks. Another epic failure by Leland Dumb*** and Elon’s band of merry fascist bigots.
No internal testing! 4 whole minutes training!
Holy flirking spaceballs!
This is going to be interesting.
Doesn’t help with policy, so it’s an absolute turd. Help with PACs? Sure.
Seems like AI could be used to write ALJ decisions
The sabotage continues unabated
The amount of denial within SSA is astounding. This is a first step towards helping employees do their jobs faster and better. Obviously it’s not going to revolutionize things overnight.. AI is just what’s happening in the world - there is no evil cabal in charge of undermining SSA by introducing it. The choice is between moving toward using the best technology available to serve the public, or not.
To those who think this is a stupid idea - don’t just snipe and snark, actually make the case for ignoring AI. What exactly is your position: we should stick with old technology and human based methods as long as we possibly can, regardless of whether it harms public service in the near future?
@12:55 Probably a net improvement, assuming they haven't already.
My position: It does a poor job of doing anything beyond uncritically summarizing information, with no ability to distinguish fact from fiction, and creates a massive risk of inadvertent disclosure of extremely private information.
Not a stupid idea. It is just not real good right now. A too early roll might be as bad or worse than a defective product.
This.
You have an extremely inflated sense of what even the most advanced AI is capable of doing.
Unlikely. The training video used an example of asking the question "what is a disabilty," and the answer used the ADA definition, which is not the standard for SSA. I asked it the requirements for SSA disabilty and it used the SGA level from 2023. Definitely not ready for prime time.
If you take your first step and fall flat on your face, does that still count as forward progress? People are literally dying because of problems at SSA, and this program fixes exactly zero of those problems. Instead, it's time and money spent on a shiny toy for executives to play with while the people who do the actual work lose their jobs or have mental and physical breakdowns.
How about instead of us making the case for ignoring AI, you go ahead and make the case for using it. Not what it might be someday, but what it is right now, then explain in detail how many claimants you're going to help with that. Not someday, but today.
Ask the ASC Chatbot who the president of the United States is. It doesn't reply Donald Trump LOL
@5:13 - Time will tell, and not really all that much time. Check back in a year. As for making the case for what it can do, do a little digging about what is taking place in the field of AI right now. Listen to what experts say, across the spectrum. Nobody was saying that stuff before AI arrived. Are they all crazy? It’s not a close call. Subscribe to a premium AI service. See what features and functionality it offers, which will sooner or later be available to the masses, including internally within SSA. The only way SSA will not be utterly transformed by AI is if it simply does not try to implement and integrate it
There is no way of winning this argument in the abstract, but if you want a use case, here’s one: you upload the medical record into an AI agent, with sufficient safeguards for PII. (Behind firewalls in other words, like everything else we do.) The AI agent takes 30 seconds or less and generates a highly organized summary of what medical evidence is in the file, and where. You then can also ask it specific questions about the evidence. Comparable functionality is being made available in other fields right now, so it is absolutely realistic to believe that will be an option in the not distant future. AI goes far beyond cutesy-pie chatbot conversations. And it is progressing at an astonishing pace In terms of the cost and its capabilities. It practically obsolete itself every six months because of progress.
Change is absolutely coming, and, being in the business of helping others, there really is no principled argument that we should not utilize the best software we can get.
My position is that we shouldn't inefficiently waste time and resources on what amounts to a beta test.
@5:42 ask It who the Coss is? Answer it gave was someone zinhabe never heard of and I am a 20+ year employee.
There wasn’t even a desktop icon to access. I was going to try it out.
And all of that is what MIGHT be available SOMEDAY. For now, right now, it's a toy. Claimants are waiting and dying right now. Workers are burning out and quitting right now. Maybe we'll commute to crystal high-rises in flying cars and use AI to do all of our work for us someday but that's not today. The problem with living in the future is that everyone else lives in the present.
Your point of view is that of someone who spends their time talking to computers, not people.
Go talk to some people. Go talk to a dying person who can't get their benefits because the website is down. Talk to the person who's had to go through 8 rounds of robo-phone hell trying just to talk to a human. Talk to an SSA employee who has to try to help those people with broken hardware, broken software, insufficient training, staffing, and management. Tell that worker how great AI is going to be, maybe, someday.
Dreams are a luxury you have because you can afford to not live in reality. The millions of people who rely on social security and the people trying desperately to help them don't have that luxury.
I’m an SSA employee of 15+ years. The announcement came out. We all immediately said “ok we will take help for our work!,” we watched the 4.5 minute required training video, read the user guide, and tried it out.
The bot has no idea what even the most common acronyms mean. It’s a glorified version of Clippy.
@7:18 What you say might make a reasonable argument for testing the AI product and assessing its utility and accuracy BEFORE deciding whether to deploy it. The thought process behind choosing and deploying this product resembles that of a kid with an impulse control disorder in a candy store.
For instance, Bisignano and others critique the fact that SSA erroneously pays less than 1% of the benefits that SSA administers. They think a less than 1% error rate is too high. Well, what percentage of this AI product's advice and information will be erroneous? A recent study showed that AI on average had a 60% error rate in accuracy even for the relatively simple task of reporting news articles. (https://www.techspot.com/news/107101-new-study-finds-ai-search-tools-60-percent.html). Will SSA now rely on a product with a 60% error rate? Since peoples livelihoods are at stake with retirement and disability benefits, it is grossly negligent for SSA to implement these products without first doing careful planning, testing and evaluation of the results.
“Not real good right now” is an understatement. It’s the worst interactive AI tool I’ve seen. It’s generations behind the more powerful and capable AI described below.
The ASC FAQs state the data the bot uses is up to October 2023. It's a gross understatement to say much has changed since then. The "advanced policy research" "feature" is worthless at best and actively harmful at its worst.
I have been using ASC for months and I have liked using the tool for my analyst job. It has been good for revising my emails, talking points and meeting notes.
A good tool if you know how to use it correctly. I even run my five weekly accomplishments each Monday.
You do have to be careful as ASC can change the context of your text .
@ 11:51 Your humanistic arguments are exactly why AI tools need to be developed ASAP. You can't claim the moral high ground just because your workflow has more humans in it. You are the one insisting on doubling down on processes which are demonstrably too slow to serve the public adequately, leaving us always begging (unsuccessfully) for more employees as the only solution to get the job done. If the first baby step SSA takes toward utilizing AI isn't to your liking, just stick with your methods. There will come a time, soon, where you will not be able to compete with co-workers who take advantage of better and better AI assistive tools. There is nothing pie-in-the-sky about what I'm saying, no crystal palaces. It's what's openly taking place outside the sanctuary of SSA's bureaucracy right now. If SSA's adoption and implementation of AI tools to date has been clumsy and plodding, the solution isn't to retreat to 20th century methods, it's to prioritize it and get it right as fast as possible.
Has anyone followed the money on this? What company did SSA contract with? How much are they paid? What’s happening to the data this product collects?
Ok Elon. Enjoy Mars.
You are using a tool with data that only goes up to October 2023 to check your work. That says a lot more about you and the quality of your work than you think, and none of it is good.
I'd like to see information regarding the awarding of this contract.
@11:51 Many disagree with your position, because they recognize the need for AI to prove it can "get it right" as you say, before its deployment. SSA is a program to benefit retirees and people with disabilities. Its beneficiaries are not a pool of test subjects for private companies that want to develop their deficient AI products at taxpayer expense. Let aspiring contractors pay to bring their own AI product's accuracy to acceptable levels before the taxpayer gives them a dime. Otherwise, you just end up getting crappy service while contributing to government waste.
There is significant misinformation out there about the (Agency Support Companion) ASC Chatbot.
https://www.govexec.com/technology/2025/04/ssa-rolling-out-new-chatbot-employees/404721/
I tested it for a few months before I left SSA.
"The tool “is not configured to learn from user interactions” and isn’t connected to other agency applications, but all content will be logged and viewable by the app’s administrators, the FAQ said."
I was actually able to teach it the various COSS and ACOSS progression after October 2023 and it was able to tell me who it was if I prompted it with the current date.
Imagine what else this thing will learn about you if you treat it as an intelligent resource instead of a spybot for the current administration.
So, we’re supposed to be impressed that you were able to “teach“ the chatbot to tell you things you apparently already knew? What is the utility of a tool that crude?
No, you're supposed to be alarmed that this thing was able to "learn" new information beyond it's October 2023 training date. We could be at a Jurassic Park moment for AI.
Post a Comment