I agree in general, but how do you resolve the problem that people will take the output of the chatbot as authoritative even when it could possibly have errors or miss important nuance? People being lead astray with information from the government could lead to very serious consequences, up to and including imprisonment.
Honest question, I had the same issue when my work wanted me to stick a chatbot in front of our HR portal and I never resolved it to my satisfaction.
> but how do you resolve the problem that people will take the output of the chatbot as authoritative even when it could possibly have errors or miss important nuance
You have to measure that problem against the status quo, which is that current government websites are very hard to understand and leave people with very little idea of what’s going on.
Giving people a big list of life events isn’t nearly as easy as allowing them to just type in what’s going on and having an LLM figure out what’s relevant to them.
It isn’t nearly as easy but it’s more reliable whereas using an llm is prone to unknown hallucinations (also ignoring the obvious ethical considerations of llms).
15 years ago I was talking to my HR rep about covering my child on healthcare. Her other parent and I were not married. “No you can’t cover her!” “Why not, I’m one of the two parents?” “Because you’re not married!” “So what?”
Cool, but the information on government websites generally meets a higher standard than just ‘making shit up all the time’ so it’s not really a relevant comparison.
We were discussing zoning in another thread, I think your comment here is very appropriate. People make up all sorts of reasons to oppose land uses, which is part of the reason municipal land control is so dangerous.
Yes, but when a human makes shit up, they're still legally liable for what they said, and depending on context can be compelled to fix the issue or lose their job.
If an LLM makes shit up, your one and only recourse is "computer says no".
Kind of like saying something happens "all the time" and sharing a single example from 15 years ago. Not to mention that a private company's HR team is not the government.
I encounter it daily in the LLM outputs my coworkers produce.
I see it every day in the text, images, and video posted by AI proponents on social media. Full of errors and wtfs and the people posting the tripe don't seem to notice it.
Maybe you've just stopped paying attention so you don't notice as much anymore?
Would your co-workers not make sh!t up if they weren't using LLMs? Pretty sure they would. And MUCH more frequently than the very rare hallucinations one runs into with decent LLMs these days.
My coworkers can be wrong, sure. But no, they weren't making stuff up pre-LLMs. But now making stuff up is automated and easy, so we've gone from simply mistaken to piles of "what the hell is this even trying to say?"
Maybe some things just are not easy. Maybe they should not be. Maybe you should be able to have basic citizenship skills to survive in a modern democ… Erm, whatever it is you guys are doing now.
Honestly just doing a Google search about what you want to do makes a page from this website appear, and it is usually very detailed while having a small "quiz" at the start to filter the information based on your situation. It's not just for life events, but for basically any public French administrative work. You'll have pages about your car, about your job, or if you want to immigrate, and so on.
At the bottom you also have other relevant links (usually to the website of the specific administrations of what you want to do) including reference to the current law that the page is based on (https://legifrance.gouv.fr is also great)
For what I used it for, it was never ambiguous in determining your situation
As far as I can tell, the america.gov LLM is mostly providing links to the landing pages you would have found through classical web search. Perhaps with a little bit of clarifying information when there are multiple choices.
Honestly, using an LLM the way america.gov does it, is pretty much the same thing as using search engines, except the result you wanted is almost always in the first result, not 2/3 of the way down the 4th page. Or 42nd on a list of things you are not interested in at all. So much more efficient!
reminds me of Minitel. France had online services for a lot of daily life back in 1982. Britain's equivalents weren't as good, but still miles ahead of America for about a decade, until the late 90s Web boom.
We had Prestel and Micronet (a form of viewdata; Micronet was hosted on Prestel, but was frequently the reason people signed up for Prestel in the first place), and there were also individual BBSes available.
Of course, we also paid for local calls so things never got as popular as they were in the US.
Speak for yourself (your country). In the EU I have interacted with several government websites which are better, faster, clearer than most other websites. Not always perfect (what is) but not frustrating either. People love to complain, but year over year I have seen steady improvements in usability. Phone support has always been better than any commercial entity, too (or it was, I haven’t needed it in years).
To enhance the status quo you don’t need LLMs, you need people who care.
As an American, I've never had issues navigating government websites. That said, I'm also not part of the apparently huge portion of our population who can't read above 6th grade, which I suspect is more of the issue than anything said here thus far.
Well we can’t exactly hang those people out to dry after failing them in our educational system. I refuse to believe that many Americans are simply incapable of reading beyond that level. I believe it is a failure of our culture.
I didn't say anything about them being incapable. That said I'm curious what a chatbot solves for people you concede have been failed by our culture to a degree where they can't read?
I know you didn’t say that, I said that. I said it because I was adding emphasis to the idea that these people are, in a sense, victims.
It’s not that they can’t read anything at all, it’s that they require more basic words. A chatbot is specifically the perfect thing to keep explaining something in more and more specific and basic ways as necessary.
> I know you didn’t say that, I said that. I said it because I was adding emphasis to the idea that these people are, in a sense, victims.
And I agree but your comment, IMO, reads accusatory. If that's not the case no worries.
> It’s not that they can’t read anything at all, it’s that they require more basic words. A chatbot is specifically the perfect thing to keep explaining something in more and more specific and basic ways as necessary.
I mean, sure, if it's accurate. Given these things' propensity to just make shit up, as stated, that feels like a way to make the problem worse, if anything.
Edit: And again it feels like a missing necessary component to this is: if a user of this LLM is told by that LLM that it's legal to, I dunno, dump waste oil in a drainage ditch, then that statement needs to be treated as authoritative and it's no longer appropriate to ticket the individual when they get caught doing that. It's similar in my mind to car dealers deploying these stupid things and some people managing to get them to agree to sell them a Toyota Tundra for $200. If you're going to put them in a place where people can reasonably assume "this looks like a proper avenue of communication" then what it says needs to be binding.
Nope! Not being accusatory, I’m just a very bad writer lol.
I think the major concern I’m picking up is that the potential for misinformation is high, though maybe we can agree it won’t necessarily be common. One person might get told to pour it in a ditch, but statistically most will be told to dispose of it properly. I also don’t think “the LLM made me do it” will work, or at least not for long, and at least not for companies. Maybe an elderly owner could get away with a fine and a reminder not to do it again, but if you’re big enough for lawyers (and thus big enough to really do damage), you’ll be expected to know better. Judges ARE still humans!
You can talk to the chat bot, or the input prompt and then in turn read out the results by screen reader. Does not solve the comprehension aspect though
I would be careful not to confuse navigating a government built website with finding information and records from government related entities. Those two are not the same issue and shouldn’t be painted as such regardless of one’s reading or comprehension (IMHO)
Throughout school my reading comprehension surpassed most of my peers. For most of my early career I was a writer. Yet I also have ADHD, and one of the ways that manifests is that I really struggle with the dense, bureaucratic language on so many government websites (same with insurance sites). I have to actively force my brain to refocus on the text multiple times per paragraph. I often just find myself skimming and then simply proceeding through a process of trial and error, fixing issues if I missed a caveat somewhere, and otherwise just crossing my fingers on form submission and hoping I did it right.
When you need to work with the government, the last thing you want is a program limited by a specific set of rules and interfaces. You want a person who understands your (sometimes unique) issue, who knows how things actually work, and can work around the restrictions imposed upon programs. Someone who can pick up a phone, help you out of bureaucratic corners.
U.S. government websites are usually pretty good post-Obama. But any structured website is too complicated for a lot of the population to navigate. They don’t even know what agency does the thing they need.
no. humans can be held accountable and learn from their mistakes.
they can be taught where they went wrong.
if a pattern emerges they can be moved to a role more fitting for them. or removed from a project entirely.
we can judge their effectiveness from past performance.
we can put them in less important roles and gauge whether or not they should be moved up.
so, no, pretending that the correct route forward is to “remove humans from the equation” for a bot that doesn’t learn, is never held accountable, and not held to the same standard as humans is silly tier thinking.
Some kid straight out of highschool being paid minimum wage to answer calls with little or no real training, and performance metrics that reward getting people off the line as quickly as possible whether they've given the right answer or not; or an LLM that by many very reasonable measures provides answers as good as those provided by post-graduate students for difficult problems, and spookily accurate answers for questions that are not... I'll take the LLM please and thank you.
depending on which model one is using, and the error the employee needs to be retrained on, i strongly disagree that you can retrain an llm as easily as you can a person.
particularly most of the commercial models.
again, i’m sure we can all come up with a thousand “hypotheticals”, and tbh, the hypothetical parade is not something i’m interested in having. but i’ll state again, “entirely removing humans” from the situation is silly-tier thinking.
particularly as even the ceos of sota frontier producing models will each and every single one tell you to never trust their model.
“our model is smartest thing in the world…”
…next breath..
“wait, you trusted our model? that was silly of you. always double check it”
If I instruct an employee to prioritize my profit in a way that opens both of us to criminal prosecution, there is a non zero chance of them whistle blowing or even cooperating with the authorities to prosecute me.
We shouldn't use an LLM because...LLMs never improve?
This seems like a great source for domain-specific RL, but even without that, we can expect there to be higher-quality LLMs that can be trivially swapped in within a year, if not a week.
Okay, but I don't understand what you're getting at with them not learning? They acquire more information, get better at providing accurate and relevant information, and acquire new skills. What is the learning they're not doing?
Maybe a human should be available as a backup or something, but not because LLMs don't "learn" -- the system seems to learn in all relevant senses.
(I'd also object to the no accountability, but another subthread is already on that.)
I have read the opening chapters of the book, and honestly I had to put it down after reading 20 percent of the book. I just don't see any new information that I had already known (rightwing effect, propaganda, Fox/Breitbait nonsense, bla bla).
I used to believe Trump winning the 2016 election was a product of rightwing propaganda and fringe theory, but with the 2024 election, I had to revise my belief. What if this is the mainstream mindset of Americans all along, and Trump simply tried to capitalize on that?
Yes of course, the book sets it out that they are covering it in a sort of historical manner, so the first few chapters are covering ground you may well know.
The later chapters are where they go into the more recent details. There is a video of one of the authors (not either Robert Faris or Hal Roberts) is giving a talk at a university panel, and he gives a shorter synopsis of the main mechanism, Its about 11 to 15 minutes in.
The set up of the book puts the evidence for the case they make later, namely that the left and center media are trying to follow journalistic standards, while the machine on the right is entirely different.
I'm pulling from memmory here, bu as an example - a fringe theory will start circulating on the fringes of the internet, the most engaging narrative will be brought up on the podcast circuit. At some point a talking head (Guiliani back in the day) will reference the theory on the news. Which will in turn be referenced by the government who will say this is a topic being discussed on the news.
The other distinction they make is that people on the right who say the narrative is wrong, simply don't get platformed.
They go through this in exhaustive detail. If you know most of the history you can see if skipping to later sections works for you. Its the best description of the actual media mechanics at play.
The 2024 elections were the ones that forced me to really re-examine my priors, and I came to a similar conclusion that Benkler and co did. Their book simply brings all the receipts.
From which equation? Because we're talking about humans who need to interact with the government, which is... pretty much everyone. And, yeah, I guess omnicide would solve the immediate problem.
The status quo is people googling shit, finding either legalese they can't understand (but maybe think that they do) or information that was legitimate ten years ago (and is dangerously inaccurate today). If I ask a question and get an answer that's been inaccurate for years I'm going to have as much or more false confidence as an AI bot that misunderstands the source. The difference is the AI bot gets better with time instead of worse. It may not be perfect but it's a step in the right direction: Let them cook.
No, someone who gets a bad answer from a random source is not going to have as much false confidence as someone who gets bad advice from a .gov domain.
The parent was obviously talking about the US government. But the UK still has plenty of legacy websites and even on the new websites there are sometimes complicated or hard to understand rules one must follow. I think a bigger thing in the UK though is that, for the most part, legislation (and legal contracts) tend to be drafted in less arcane language while still being reasonably precise, compared to the US
In fairness, the UK government is masterfully good at bureaucracy. Clear directions; spacious, beautifully designed forms that are easy to read, and that you don't need a team of lawyers to fill out. And jaw-droppingly beautify typography. And an implicit understanding that every piece of UK bureaucracy should be easy to navigate for anyone. And their websites reflect that.
The US government, on the other hand ranks close to the very bottom of 1st world nations for forms culture. Forms that are difficult to read, and even more difficult to fill out accurately, often requiring supplementary information from multiple sources. So much so that people are advised to consult a lawyer just to fill out an application for Tax Identification Number! (My personal worst experience with American forms culture).
So I don't think it's entirely fair to compare UK government websites to American government websites. Navigating a bureaucracy that's fundamentally broken is a significantly different kind of problem.
fwiw, I'm Canadian. Canadian forms culture falls somewhere in between. There aren't a lot of forms that are terribly difficult to fill out; it's rare to find a form that requires more than about a dozen pages of supplementary information (all of which comes with the form when you download it). But still not anywhere as easy to use as UK forms are.
The new status quo is a site that doesn’t mention an insurgency.
If America was going to follow China’s footsteps, then what was all the hullabaloo about freedom and democracy all about.
Even if we grant that there is a difference between facts about history and … recent history, the status quo is improved by building better websites.
A website is also a document of regulations and a source of evidence. If a government LLM hallucinates a new feature or rule, and someone acts on it, a fresh legal hell has been created.
America is also unique, in that one part has part of its strategy to make the government as ineffective as possible, since it benefits Republican political goals.
> You have to measure that problem against the status quo, which is that current government websites are very hard to understand and leave people with very little idea of what’s going on.
The solution to that is to make the sites easier to use.
Isn't that kind of a good thing at the scale of a government? Each task should be hard, use a lot of energy, to ensure it's important enough to do. The DMV should take 1/2 day of work. Otherwise people's egos start think they have it all figured out and then people start imagining improvements and become dissatisfied. Wouldn't this just lead to unrest and resentment?
Painfully this will fall to the court decide and create the precedence here. If this is an official government app, then the onus probably is on the government to provide correct information? Who knows, it could go either way and the courts could say "you're on the hook".
I don't think we have a good answer for liability here, like for Tesla FSD, Waymo, or even the current frontier models.
I don't know for the gov. For HR, the best approach I've seen is to have source and links extremely prominent and have the bot speak like a mascot ("bip bip boop I'm the HR bot" style)
People don't read disclaimers nor warnings about the content, but usually won't come to the HR to complain a mascot fed them weird info they didn't bother to check.
> how do you resolve the problem that people will take the output of the chatbot as authoritative even when it could possibly have errors or miss important nuance?
Compare this with the benchmark.
Earlier this year I called ABC with a question about my liquor license and spoke to three different agents. Each one gave me a different answer. All of them were wrong.
There's a ton of hearsay and rumor about how the US Government works, even among educated/sophisticated people who are dextrous with bureaucracy. The problem gets much worse as you go down the educational ladder, which is most of the US population.
The Perfect, please meet The Good. You are not enemies.
If you're really curious, I want to get a Type-86 tasting license for my store. The ABC site[1] says my kind of store (<5000sqft) is "generally" not qualified, which makes it sound like there are exceptions (spoiler: there are).
The first agent told me that the license is not an add-on, meaning I'd have to complete a very long and expensive process to obtain the license with the risk of getting denied. This was wrong. It's an add-on.
The second agent told me that I was not eligible because I have less than 5000sqft of space. Not true (see below).
The third told my wife that 75% of the items in our store need to be alcoholic beverages to qualify. Wrong again.
I finally asked my legal LLM, which pointed me to the statute, which says, "Type-86 ... can be issued to businesses which also hold off-sale retail licenses [that] earn at least 75 percent of their total gross sales from the sale of alcoholic beverages." Ultimately, it comes down to how ABC interprets and enforces this law, so I really expected to get some clarity from them rather than complete falsities.
> I agree in general, but how do you resolve the problem that people will take the output of the chatbot as authoritative even when it could possibly have errors or miss important nuance?
The bar for something like this should not be 100% accuracy. It should be 100% accountability and transparency, followed by being at least as accurate as google search.
Similar to autonomous driving: it doesn't need to be perfect, just less likely than humans are at causing an accident.
I strongly disagree. The preferred accuracy of an official government source of information should be higher than a google search. The bar should be as close to 100% accuracy as possible as well as 100% accountability and transparency.
"Sometimes computers just make shit up now but that's OK because humans do to and if they do we can just sue them" should not be acceptable.
You can disagree but the fact is people will take the path of least resistance. I'm not saying we shoudldn't strive for 100% accuracy but we also need to be realistic about what an LLM is. I'd rather we not pretend there's some magical combination of weights out there that will make it completely perfect.
I'd rather we not feel obligated to use LLMs for purposes they aren't suited to rather than just "being realistic" about the consequences of using them everywhere for everything.
Let's bring the subject back into focus: This is an example of an LLM being used to search a massive database of scattered text and files. Are you saying LLM's are not suited for searching through text?
>This is an example of an LLM being used to search a massive database of scattered text and files.
That isn't how LLMs work. LLMs are statistical language models, not search engines[0,1]. They can be prompted to call external software to search databases, but they themselves are not capable of doing so, and more often than not they generate responses based on their own model, which may not be accurate. We've had technology that was capable of searching databases for decades without the quirk of not being capable of presenting that information accurately.
>Are you saying LLM's are not suited for searching through text?
I am saying that first and foremost they don't do that and furthermore that they are less suited as a substitute for that than what we had before. An obvious example of this is the AI feature of Google Search, which I've seen hallucinate results numerous times. But you can also look up the numerous times AI has fabricated citations when used in scientific research.
no, not at all. in addition, the funding should be spent on organizing that database in a more easily searchable format and system instead of introducing a stochastic prediction engine.
The problem is that this isn’t possible without making it impenetrable to ordinary people. It’s a usability problem with tradeoffs.
My local county government has a pretty good website, but even then it’s completely overwhelming if you try to find something off the happy path of the most common services/tasks. I can’t imagine the nightmare of trying to organize a federal portal manually and keep it up to date. Plain search isn’t sufficient because there are so many overlapping functions that are just slightly different.
I hate to say it, but this is one area that an LLM assistant actually makes sense. Maybe it needs a second validation pass with routing to a human assistant if it can’t figure out how to give an accurate answer to a query.
(Personally I’ve found that “search assistant” LLMs are the only task where I’ve found LLMs to be occasionally helpful to me.)
Edit: the big problem with this sort of thing, even if not AI powered, is that government itself isn’t well-organized and the legally “accurate” answer may not be the correct one, as per this comment: https://news.ycombinator.com/item?id=49895741
What happens now if someone uses traditional search and comes across an outdated wiki or Confluence page with outdated/wrong/bad instructions, or interprets something incorrectly?
I’m responding to your point that it would be “unacceptable” for a regular government website to provide wrong information. The case I linked to says that the government isn’t responsible for mistakes in what the government tells you about the law. If someone in the social security office says the law is X, but it’s actually Y, then the citizen is still required to follow Y.
The case isn’t about whether it’s “illegal” for the government website to be wrong. It says you as the citizen are held responsible for the actual law, even if you followed incorrect government advice. So the current rule already acknowledges that advice provided by the government may be mistaken and says that’s no excuse.
So the question is is the chat bot better than people trying to understand complicated government websites by themselves?
It has always been the person's problem to solve, even after this website.
No government employee is your lawyer. If the government gives you incomplete or wrong information, you are still liable for not doing the correct thing.
Having a website that does a best guess at what steps a person must take after changing a name or getting married is better than nothing, but the status quo isn't "nothing".
It needs to be kept up-to-the-minute updated and it can't skip any nuance / details. Does anyone here really trust the lackeys who spectacularly failed at DOGE to do boring and highly detailed work?
My team experimented with this a couple of years ago with a complex government regulatory body of knowledge.
Google and Microsoft have APIs that will ground with and reference the specific policy documents, which helps. Even then, it was more useful as a job aid for people already knowledgeable of the data than for a layman.
With all LLMs, the quality of the question correlates with the quality of the answer.
> people will take the output of the chatbot as authoritative even when it could possibly have errors or miss important nuance?
Meh, the kind of person who would accept bot output without even reading its sources are already trusting the world's dumbest model, Google's "AI Overview" by searching and reading that instead of clicking any results. If we can get even some of them to start out at america.gov instead of google.com for those questions, it's likely going to be far higher quality due to using a better model and having been trained, I assume, to only answer using knowledge from a .gov primary source rather than guessing based on vibes, or on jokes once seen on Reddit, as AI Overview tends to.
Considering that informational tools should be accessible and reliable, I think it's very worth not writing off people who may not have the time, inclination, or understanding that it's necessary to double-check generative responses
They can take it as authoritative, but that won't change the consequences if they're wrong. This is already a thing between two people; "my cop friend told me it was okay" isn't a legal defense. Changing the 2nd person to an AI doesn't change much, both are still agents of the government.
Doesn't really have any bearing on a similar system for HR. The government is sort of unique in a bunch of ways, but not really being responsible for what their low level agents say is one of them.
I'm not sure how to fix that, to be honest. I'm also not sure how much worse it is than trying to Google it? Google is full of stuff that's either wrong, or won't apply for a reason that takes some reading comprehension to grasp (e.g. state-specific rules/programs/etc).
Honest question, I had the same issue when my work wanted me to stick a chatbot in front of our HR portal and I never resolved it to my satisfaction.