“Wildly irresponsible”: DOT’s use of AI to draft safety rules sparks concerns

poltroon

Ars Tribunus Militum
2,011
Subscriptor
Writing regulations clearly and precisely is only important if you want them to be enforced. Otherwise nah.

The "word salad" is the difference between prose and code. It does mean something to the expert practitioners.

Now, if you're an ethical operator in the space, you probably do want clear and precise and well-formed regulations, for yourself as well as your competitors. But then, if you're an ethical operator you probably haven't paid bribes to Donald Trump.
 
Upvote
19 (19 / 0)

alxx

Ars Praefectus
5,006
Subscriptor++
So the use of AI in government is predicated on the idea that the national debt has been caused by there being too many technical writers in the DOT?
Its part of the death of experts. The stupid idea that you can use ai to make anyone capable of doing any job or replace experts with ai
 
Upvote
30 (30 / 0)

Kebba

Ars Scholae Palatinae
980
Subscriptor
If there's only one thing that companies are looking for in DOT rules, it's speed and changes and factual errors.
I have ranted about this in another AI comment section about writing reports, but I really wonder how many people are so ignorant about why creating these kinds of documents take so much time.

Are there really that many in "high" positions (and the general public) that think that "writing it" is the reason the process is slow? And not, you know, because writing good rules are really f-ing hard and require involvement by multiple stakeholder, assesment of risk/reward and secondary effects as well as just let citizens have their say before the loop is repeated with all the fixes that was found. Christ, writing a 100 page document modifying some old regulation takes what, a week for a competent writer once all info is there....
 
Upvote
35 (35 / 0)

DWofP

Smack-Fu Master, in training
67
Subscriptor
Most of the word salad is laying out history, connections... text that explains what and why.

Because when things go to court you need all that. It is hard enough to read that stuff now, let alone once it becomes auto generated.

Gov rules keep getting harder to write and harder to pass because of courtroom events... Legislative standard, etc.

Writing them automatically won't help any of that.
 
Upvote
18 (18 / 0)

Sconsi4nya

Smack-Fu Master, in training
31
DOT staffers were told that “most of what goes into the preambles of DOT regulatory documents is just ‘word salad,’” and “Gemini can do word salad.”

BOILERPLATE. The word you’re looking for is BOILERPLATE, you absolute tool. BOILERPLATE is a span of often boring but technically or legally important standardized text.

WORD SALAD is a psychiatric description of a specific symptom of dementia, which your handlers should be well acquainted with.
 
Upvote
48 (48 / 0)

SportivoA

Ars Tribunus Militum
1,769
How do these companies get away with only charging $1 or less for something that actually should cost millions of dollars? I'm a federal contractor and we have annual training that reminds us we must bill for every minute worked--we can't give away something of value to the government.
That was exactly my thought when the price came up. Over on the space discussions, there have been repeated assurances that SpaceX can be considered reasonable, if even overbidding (good for profit!), for their launch services because the government audits them to be assured materials and services aren't being bid below cost. I'm very confused as to how these cloud services were analyzed on bid because this sounds exactly like the "first dose is free, then get locked in" sort of situation government procurement is supposed to prevent. Also I'd definitely like to see what realistic bids, properly held to the cost + margin requirements, look like even if the rest of the auditing remains under seal (as it should).
 
Upvote
21 (21 / 0)

DrewW

Ars Tribunus Militum
2,042
Subscriptor++
How do these companies get away with only charging $1 or less for something that actually should cost millions of dollars? I'm a federal contractor and we have annual training that reminds us we must bill for every minute worked--we can't give away something of value to the government.
Because the data is what’s actually valuable here; no one in the trump admin is savvy enough to know if you aren’t paying, you’re the product.
 
Upvote
11 (11 / 0)
Technically, as long as the rules/regulations/etc. have a senior expert listed as the author (or the leading author), a person who puts their reputation on it and bears responsibility, I don’t really care about the process they use to get stuff done.

I mean, as long as it’s a product of a one-line prompt, not vetted by anyone who knows what they’re doing.

On the other hand, using a high-end LLM to brainstorm and work through the stuff and then use it to produce a raw draft is ok with me. In particular, when you take the said draft and feed it to a different LLM with instructions to be extremely critical of the thing. And when there is an actual expert that approves things as if they wrote it themselves. Yeah, I know, not in this universe…

I'm genuinely curious what you feel the benefit of using an LLM to then feed through another LLM, and then having it all reviewed by an expert is over just having the expert do it it.
 
Upvote
17 (17 / 0)

Fatesrider

Ars Legatus Legionis
25,408
Subscriptor
But the DOT’s top lawyer, Gregory Zerzan, isn’t worried about that, December meeting notes revealed, because the point isn’t for AI to be perfect. It’s for AI to help speed up the rulemaking process, so that rules that take weeks or months to draft can instead be written within 30 days. According to Zerzan, DOT’s preferred tool, Google Gemini, can draft rules in under 30 minutes.
This guy is a lawyer. He has no specialties in transportation, and in fact has never before been part of that side of government. He's worked mostly to draft rules, and nothing else.

So, his interest isn't in making the DOT better. It's just cheaper.

The fallacy he cited indicates to me that he's just a fuckwit mid-level bureaucrat with extreme delusions of adequacy.

Proof?
“We don’t need the perfect rule on XYZ,” Zerzan told DOT staffers at the meeting. “We don’t even need a very good rule on XYZ. We want good enough.”
There is no such thing. You're ALWAYS moving the goalposts in a world that actually has time passing. What worked before, becomes untenable. And this twit is aiming for a very low bar.
 
Upvote
19 (19 / 0)

Mechjaz

Ars Praefectus
3,380
Subscriptor++
My main experience with DOT is that it's a safety standard that lets US motorcyclists ride in novelty hats.
My novelty hat was $500 and manufactured by Shoei and independently certified by Snell, at least.

But yeah, crossing over into South Carolina (no helmet law) makes my jaw drop every time. I just.. it doesn't compute for me.
 
Upvote
7 (7 / 0)

RZetopan

Ars Tribunus Angusticlavius
8,281
It's too complicated having all these different kids of liquids. Could we get rid of water since it is tasteless anyway? We could replace it with a high energy drink. One plants would crave.
“But, what are electrolytes? It's what plants crave, dummy!” The AI explanation coming to you.
 
Upvote
4 (4 / 0)

RZetopan

Ars Tribunus Angusticlavius
8,281
I'm genuinely curious what you feel the benefit of using an LLM to then feed through another LLM, and then having it all reviewed by an expert is over just having the expert do it it.
Because it has been shown that trying to understand code that some else (especially AI) wrote and then fixing the errors in it, is so much less efficient that doing it all yourself. This is the LLM AI model in a nutshell. Create something that will pass a casual non-critical examination, but fails badly when tried in practice, so an expert is required to find the errors and fix them. Any other method leaves Altman, Google, etc. out of the loop and that is really horribly bad! Bad dog*!! Bad dog*!!!

*Purposely misspelled DOGE. /S
 
Upvote
4 (4 / 0)

RZetopan

Ars Tribunus Angusticlavius
8,281
Note that Musk is very much against the NHTSA rules that he must report all Tesla FSD accidents, so the obvious fix is to remove those innovation stifling rules so that he can start breaking things much faster. As a “jenius” once alluded to: “If you don't report bad things happening, the bad things quit happening”. It is completely obvious that obeying stop lights and signs should be entirely optional, and likewise for attempting to drive on train tracks, hitting emergency vehicles, stationary objects on the highways, and even in parking lots, etc. Tesla has thoroughly “proved” this, so the world needs to catch up and follow their fantastic lead. /S
 
Upvote
7 (7 / 0)

Nick31

Seniorius Lurkius
21
It’s for AI to help speed up the rulemaking process, so that rules that take weeks or months to draft can instead be written within 30 days.
I'm not sure I've ever read a more horrifying sentence. The last thing our over-regulated country needs is for its thoroughly captured bureaucracy to have the ability to create more rules in shorter time.
DOT is probably not the best place to test AI either. Can AI do word salad? Sure. But the part of the rules that correspond with real world activities and objects might be beyond the comprehension of AI, leading to dangerous mistakes.
 
Upvote
7 (8 / -1)

AdrianS

Ars Praefectus
3,911
Subscriptor
I'm not sure I've ever read a more horrifying sentence. The last thing our over-regulated country needs is for its thoroughly captured bureaucracy to have the ability to create more rules in shorter time.
DOT is probably not the best place to test AI either. Can AI do word salad? Sure. But the part of the rules that correspond with real world activities and objects might be beyond the comprehension of AI, leading to dangerous mistakes.

Paging mr Jorg X McKie.
https://en.wikipedia.org/wiki/The_Tactful_Saboteur
(He worked for a govt. bureau dedicated to slowing down govt. regulation by sabotage of various kinds)
 
Upvote
3 (3 / 0)
Note that nowhere is 'safety' addressed, which should be DOT's purpose

I’m all for the use of AI in rule making IF the vendors of the AI tools used can guarantee that it will no less error prone than a human drafted law and take full responsibility for any deviations.

What’s that? You’re saying that no AI vendors would accept such a liability? Well, then there’s your answer…
 
Upvote
5 (5 / 0)

PermanentDonut

Smack-Fu Master, in training
99
This is very on brand for a Trump business. Of course he'd rapidly adopt a buzz word technology (AI) and force it to be used even if its implementation is shoddy and not the right tool. Only to try and get attention that he's doing it "better" and "faster" than anyone else to feel like a winner. Of course, we know in the end it'll blow up in his face and this latest business venture will fail. Except this latest business venture is the US government.
 
Upvote
1 (1 / 0)
Rather than having the DOT serve as the guinea pigs for this idea, why not use a department that is perhaps less likely to kill someone if (when) things go bad? I have one in mind: the Internal Revenue Service. It's highly complicated and affects everyone in the US (so streamlining operations should[*] make things better for everyone) and it's not like anyone would take advantage of any massive loopholes AI may introduce (any more than they take advantage of loopholes that have already been caused by Natural Intelligence.) What could go wrong?

I mean, if the Federal Government should go financially bankrupt, that's a situation where Donald Trump has loads of experience.

[*] Definition of should: probably won't.
 
Upvote
-4 (0 / -4)

Jakelshark

Ars Tribunus Militum
1,712
I work in local government. Some volunteers used AI to re-write the city's complete street policy. It's fucking garbage and a pain in my ass because it's full of rhetoric that sounds great to the layperson (and problematic to the person actually in charge of doing the policy). They also can't actually explain what was wrong with the existing policy or why it needed to be re-written...

Leave ordinance and policy to the people who actually work with it day to day. AI can generate a jumping off point, but it is not saving that much time in the grand scheme of things.
 
Upvote
8 (8 / 0)

jezra

Ars Tribunus Militum
2,150
The US Department of Transportation apparently thinks...
No, it does not. People at the department 'think'... and in this case, it would appear that the 'thinking" is being done by lawyers representing the department; Daniel Cohen and Gregory Zerzan. Lawyers are good at ensuring no one responsible will be held liable for the impending deaths caused by AI created DoT policies; which like most gov policies will focus on increasing the profits of corporate sponsors first, and the safety of the constituents last.
 
Upvote
2 (2 / 0)

KrookedRooster

Ars Praetorian
554
Subscriptor
It's only the higher ups who probably don't write these things that want this. They need to "meet the numbers" and if the numbers are "regulations must be done fast" and also "do it with less headcount" they will find a way that is "good enough".

The actual people whose job it is to write and proofread these things hate it. And not because it "took 'r jobs" but because they are actually providing a service that should be better than "good enough" it is THE STANDARD!

Everything should really be done above standard and if the standard is.... oh.... I see now....
 
Upvote
3 (3 / 0)