Maybe we need a better-scaling legal system that does not take years to resolve simple disputes. I am not convinced AI can replace lawyers, but more people having access to the law without extremely costly fees does sound good to me.
Moving from an adversarial each-party-has-lawyers-presenting-as-extreme-a-case-as-possibly-can-be-made-for-their-side system to something more like expert arbitrators and independent/court fact-finders would avoid the 'bury them in paperwork and filings' lawyer-cost explosions.
Usually the concerns I've seen are around the lines of "the experts are gonna favor the powerful incumbents who they see a lot" buttttt if our system is already doing that AND costing way more anyway, barring entry entirely to many, is that so much worse? Even assuming we can't try to regulate that?
Yes, although they don't really obviate the issues raised by GP (costs, length, complexity).
If anything, some of the largest civil law countries in Europe, like Germany and Italy, can arguably be considered worse than common law countries in all of these respects.
Whether that's incidental or due to civil law, I cannot say.
Check out the amount of lawyers per capita. From experience in Berlin, one of the most dysfunctional German states, the length of lawsuits is a big problem, complexity I don't know - but costs are not crazy, and loser pays (a fixed amount depending on properties of the case). AFAIU, a side can pay more than the fixed amount for their lawyer, but the loser will not pay more. All that discourages outspending someone with frivolous litigation.
How is that different from the current system? Wouldn’t you just bury those fact-finders in paperwork? Or are you trying to shift the cost and responsibility of the discovery phase to the court instead? Would the court pass it through to the plaintiffs/defendants like the lawyers do or is it just covered by the state?
We have quite a few tribunals, ombudspeople, and other non court dispute resolution services in Canada.
Typically, the arbitrator is very well versed in the rules/area they are enforcing, and can see right through a snow job. In my province there is a tenancy tribunal that has to make a ruling before a court gets involved, and the courts rarely disagree with a decision made by the arbitrator in the 1% of cases that do get appealed. Because the rules are so clear, you can shovel piles of paperwork at them, and they will ignore it because they know what pieces of paper they actually need to make their decision.
It isn't for complex big money cases, but it means that courts have more time for those cases because they aren't dealing with ticky tacky minor disputes over damage deposits and fender benders.
Tenancy arbitration is just one of them, BTW. Just about every industry or agency has an outside ombudsperson that is the first step of dispute resolution before a court will be interested.
With frontier models, I have found so many self contradicting points in Indian tax law that I would rather welcome a good AI helping to reduce discretionary powers of officers via highly subjective interpretations. This power asymmetry is basically the driver for a lot of developing world corruption.
I've been thinking about autoformalizing local laws using agents into TLA+ or something, but it's sufficiently past enough my actual skillset that I'm pretty sure I'd just end up wrestling with slop like a pig in the mud. It's a shame though, I consider law to be just kind of a shitty codebase, with natural language being tortured into cooperating, so it's a really natural fit.
I'll probably yield to my temptations eventually and proceed anyways. Lord help me from all the creative but completely detached interpretations I'll land on.
I do, though I'm not entirely sure what am I looking at on those links. Could you start by explaining that? They look like training courses or something.
I saw a sequence diagram browsing around, seemed to be specific to a sample scenario?
Each time I try and explain, it flags comment and says its ai slop.
Each "program" here is a government program, agents orchestrate everything including the collaboration between all parties required.
High points: I have been able to help over 100 people get housing with no HITL on my side.
Note: Each host/subdomain is a project, they all inherit policy from each other and that drives the program generation and orchestration layer. Policies can be managed for the diff agencies at rnc/dnc.dev
tl;dr a "program" here is a government program (get an FFL, file a discrimination charge, apply for a benefit), codified so that every step has an actor, typed inputs and outputs, and a citation to the provision that authorizes it. Agents then walk each party through it. And note it points the opposite way from ChatGPT-drafts-your-tribunal-claim in TFA: that dynamic broke because AI made filing free while adjudicating stayed expensive, so the queue explodes. Codifying the procedure attacks the other side; what's actually required, where it actually goes, and whether you have it; before it becomes a hearing in 2030.
Fair question, and the "training course" read is not an accident; it's the same shape underneath. A program is an ordered chain of modules, each with a declared actor and typed inputs/outputs. Courses are also that. So it renders with the same components. The sequence diagram you found isn't a sample scenario, it's the deal template's actual step graph; the thing an instance runs on.
Three authored files per domain:
- an ontology: the domain's vocabulary, its regulatory frameworks with real citations, the O*NET occupations that staff it, the systems of record it touches
- intents: what a person actually shows up wanting ("open a gun store, need an FFL"), with typed parameters
- deal templates, one per intent: ordered pipeline_steps, each with an actor, inputs, outputs, and a policy_check
The page you clicked is generated from the last two deterministically. No model in that path.
The part that speaks to your TLA+ instinct: I deliberately don't formalize what the law means. I formalize the procedure, and bind each step to the provision that authorizes it. Formalizing semantics is exactly where you get the creative, detached interpretations you're worried about, because every gap gets filled by the model's guess. Formalizing procedure asks the model to transcribe and cite, which is checkable:
- every step input is a ref; param:x, step:3.some_output, system:NICS.event; and it has to resolve. A step: ref must name an earlier step's declared output, so the dataflow is a DAG with referential integrity.
- every step's policy_check must name a framework declared in the ontology. A step that no provision authorizes fails validation.
So most hallucination becomes a build error instead of a plausible sentence. That's the whole trick. Not a smarter model; a narrower artifact.
Concretely, since you're right to expect slop: my first pass at four new agencies came back with 100% of step inputs referencing parameters that didn't exist, and prompts that literally said "Subject?". The validator refused all forty programs. That's the mechanism working; I'd have merged them on a read-through.
Limits, since you'll ask. It decides nothing; no adjudication, and consequential steps are human-gated. It's also not a formal method: the invariants are referential integrity and citation binding, not model checking. The genuinely temporal parts are the deadlines, and those do bite; the NLRB's six-month charge window runs from filing and service, with service being the filer's own duty, so a filing-date-only clock computes the wrong date on a deadline that destroys the claim if you miss it.
Re: the sibling comment about discretion; that's the actual pitch. Discretion hides in the gap between the written rule and the practiced procedure. Writing the practiced procedure down, with a citation per step, is what makes the gap visible.
AI doesn't give you a lawyer service without fees though. It makes you feel like you're getting a lawyer service, then you show up in court and say what the AI tells you to say and lose ten times your net worth because what the AI told you to say was legally nonsense.
This is a thing, and in the context of the article too. In UK, if an employment tribunal finds you’ve wasted the tribunals time they can and do order costs paid for both the tribunal and other party. Odd they don’t mention it.
As with most such things, though, what is needed are escalating penalties. If you're a Joe Schmo who files a complaint that turns out to be invalid, no penalty for you. If you're a rich jerk who files a bunch of such complaints, a few million maybe. If you're a giant company that is using the legal system to push people around, $50 billion.
That deters honest complaints from poor plaintiffs. The Economist may approve, but justice shouldn't depend on the relative finances and/or power and status of the opposing parties.
Of course it does, often. But that's a pathology, not a feature to be proud of.
It's true that some complaints are irrational, vindictive, or vexatious, but the system seems quite good at identifying those already.
> That deters honest complaints from poor plaintiffs.
If the complaint has a high probability of success then it has a low probability of making the plaintiff pay costs. If it has a low probability of success then isn't that what we're trying to deter?
I think there's a distinction to be made between "a nuisance complaint" and "a legitimate complaint whose presumptive odds of succeeding in court are less than guaranteed"
When the odds of success are 99% but not 100%, that's not going to deter a lot of people.
When the odds of success are 50/50, in this context that's a nuisance complaint because making it "free" to turn all of those into real adjudications would flood the system with exactly the cases where the outcome is unclear which are correspondingly the hardest to efficiently resolve. The only practical way to address that is to clarify and simplify the rules to begin with so that the outcomes are predictable rather than ambiguous, and then we're back to not really deterring anyone because it's obvious who is going to win.
And in the meantime plugging up the system with unlimited questionable complaints isn't very prudent.
In Germany, the losing party pays 3x the court costs. The only issue is that everyone gets paid - the lawyer, the court etc. The only person that doesn’t get paid for their time investment is the person who is suing.
But as far as I understand in many lawsuits both sides lose at least for some of their claims/counterclaims, at least in business law. So court sees it as both sides lost, who should pay 3x?
what about a "rich loser pays" rule? if the losing side can afford to pay the other sides fees without getting close to bankruptcy they have to do it, otherwise it reverts to each side paying for itself.
that would encourage lawsuits against the state and big corporations, and make it less worth it to sue small businesses or your own employees. thats common sense because bigger more powerful organizations tend to be more evil (as in breaking the law because they can afford to pay fines) and money is a good measure of power in capitalism.
Lawyers normally don't have any liability risks, no? And reputational risks are very limited, a lawyer can always point to a judge, to the other lawyer, to circumstances, and away from their own (or the AI's output) blunder.
"Free, AI-powered legal advice should be good news for workers. Instead, it is proving to be a tragedy of the commons. For workers with genuine grievances, the surge in demand means longer waits for justice. For employers, it means bigger legal bills to respond to claims, both well-founded or fantastical. In the age of AI, a system intended to provide access to justice suffers from, if anything, too much access."
That’s what a “tragedy of the commons” is though. The degradation of a shared resource (state capacity) by people acting in their own best interest.
The state using AI “to make bureaucracy instantaneous” is a tragedy of the commons. The experience is degraded because before you trusted your grievance will be heard, but now you might get auto rejected by an AI. It’s pretty much the definition of the phrase.
> The state needs to use AI to make bureaucracy instantaneous
The stuff of many sci fi tropes - it seldom ends well.
A bit OT but there's a character in a couple of frank Herbert books - Jorj X Mckie saboteur extraordinary in the Bureau of Sabotage (BuSab). BuSab is a government agency responsible for conducting dirty tricks "in lieu of red tape" to help slow down and regulate the vast galaxy-spanning bureaucracy.
In principle you can exclude someone from being able to file a plausible complaint after they've been wronged, but if we take it as a premise that we don't want to do that then it's not very excludable anymore.
This is supposed to be a joke but that's pretty obviously what would happen if you actually tried to do this.
You apply for your permit or whatever and the AI instantly denies you, so you start over and pull the lever again and this time the AI's random number generator comes up 777 and you get your permit.
Then the house can tune this for either reducing red tape (e.g. pulls are free) or favoring the rich (you have to pay a fee for each attempt so only the poor have to take the L once they run out of pocket money).
I am not surprised. Contrary to popular belief, free markets nor capitalism do not resolve tragedy of the commons; they just hide it in (negative) externalities.
The biggest example is probably the CO2 emissions and global warming, but also the social costs of economic outcasting - making people jobless, homeless or otherwise stressed due to poor distribution of resources.
TFA ends on this note though, which seems pro-labor:
> But it would present employers with a different problem. If AI fulfils its promise, it could before long give every worker the equivalent of a top-flight lawyer in their pocket, able to file precisely constructed cases against their bosses at will. A deluge of slop claims could give way to a wave of winning ones. Labour said its act would shift power from employers to workers. With AI, power will move faster and further than the politicians imagined.
Even the “pre-case” legal stuff is out of control. In my Australian organisation we’ve noticed a deluge of what I can only call “privacy demands” that talk the talk… but have the jurisdiction or subject legislation wrong.
But AI told them they could demand these things (even if it would be illegal for us to comply! Think KYC stuff.) and they don’t have the skill to criticise the AI or know when to stop.
Yep, and the cost is borne by the receiver of the letter. The person sending shitty LawyerGPT messages will just keep doing it until there’s consequences, but there aren’t any.
So I have to wonder if this article written about clogging up employment courts with AI-generated cases is itself generated because the premise (in the title and the article) refers to the "tragedy of the commons", which was debunked almost 20 years ago.
For a bit of history, the term was the title of a 1968 essay [1] based in part on ideas that originated earlier. It becamse popular as a driver for privatization of government services in many Western nations based on the (flawed) idea that privatization would increase "efficiency". All it did was transfer wealth from customers with inelastic demand to newly-minted private owners.
But it never seemed to match reality. The opposite seemed to be true: private resources became over-exploited whereas shared or public resources were successfully managed and shared by communities across the world in some cases going back centuries. This culminated in Elinor Ostrom [2] winning the 2009 Nobel Prize in Economics for officially debunking the idea.
Now I understand how ordinary people still quote this debunked idea. It did become popular. But journalists, particularly those for a publication called The Economist, should know better.
It's the first time I hear something like this; if that's true, then this must be one major case of meaning shift, because since forever I've known and used and seen used this term to refer to the flaws of privatization. The solution to tragedy of the commons is literally a central authority. It's, like, the opposite of "driver of privatization".
Tragedies of the commons are often aimed to be 'resolved' through central authority, but another way of stating what's happening there is that they are being resolved by someone being granted the ownership right over the commons (in the cases you're mentioning, government). Therefor, it hasn't exactly meaning-shifted, but 'contextual' implications of the term for the reader are definitely bimodal with respect to privatization.
Tragedy of the Commons was specifically the idea that cooperation would fail due to selfish competition, so some force was needed to administer it.
Economists used it to argue both for privatization and against, but always for top-down, authoritative control.
In the US, it was almost always used as an anti-"socialist" argument for privatization, by arguing that the free market would in fact be an analogue for the "natural" competition, because as we all know the Free Market always finds the most equitable and consumer-desired outcome.
I'd imagine in Europe it was more common to argue for government management.
You can have a tragedy or not, Ostrom didn’t show it’s always well managed by the public. One can say ‘tragedy of the commons’ while still knowing about solutions where it’s not a tragedy anymore
Almost nothing you said is true. The point of tragedy of the commons is that regulation is necessary, as opposed to the libertarian "self-regulating market". It's very real, and not in any way "debunked". Many fish species are close to extinction because of it.
> Interim relief is a case study of how AI, like a heat-seeking missile, can lock on to the most obscure provisions of the law—and create carnage. The impact on Britain’s employment tribunals (courts that resolve disputes between employers and workers) illustrates a phenomenon emerging everywhere. AI-induced demand is overwhelming bureaucracies built for the analogue age—from Dutch municipal-tax appeals to the Canadian privacy regulator to parking-ticket tribunals in every major city. In Britain workers now ask large language models, rather than human lawyers, to help them sue their bosses quickly and cheaply. Claims have surged and backlogs grown. A case filed today may not be heard until 2030.
>
> Free, AI-powered legal advice should be good news for workers. Instead, it is proving to be a tragedy of the commons. For workers with genuine grievances, the surge in demand means longer waits for justice. For employers, it means bigger legal bills to respond to claims, both well-founded or fantastical. In the age of AI, a system intended to provide access to justice suffers from, if anything, too much access.
I think this is another case of "we've been getting away with murder for a long time. How dare they use a floodlight?" syndrome. Or, floodlight syndrome for short.
There are a lot of laws that exist on the corporate and individual level solely for the purpose of selective enforcement to throw "the book" at the unpopular; the insurgents; or the under-resourced. It's an implicit component of the legal system.
For example, a fossil fuel utility, Entergy, stopped an insurgent wind farm / project by arguing that the startup making HVDC lines, Clean Line Energy, couldn't make power lines, because only utilities could make power lines. And to be a utility you need to have power lines. From the paper, https://cdn.vanderbilt.edu/vu-wordpress-0/wp-content/uploads...
> Entergy pointed out that only public utilities can build transmission lines in Arkansas, and that Arkansas law defines “public utility” as a company that “own[s] or operat[es] in [Arkansas] equipment or facilities for...transmitting...power to or for the public for compensation.”152 The Arkansas law creates a catch22. Because Clean Line did not own or operate any transmission lines in Arkansas, it was not a public utility. And because it was not a public utility, it was not authorized to build transmission lines.
And that's not the only such case, as the saying goes, many such cases,
> In 2011, a fossil fuel utility convinced the Arkansas Public Service Commission to deny certification because the wind company had no existing transmission infrastructure, and so did not fit the legal definition of a “utility.” In 2017, the Illinois Supreme Court denied certification for the same reason. The Missouri Public Service Commission claimed that certification was not in the public interest because “harm” to landowners “outweighed any in-state benefits.” Projections for the wind project, however, suggested that it would create over 1,500 jobs and reduce electricity prices for Missourians by over $10 million annually.
> In 2016, authorities in Oklahoma prosecuted bartender Colin Grizzle for serving vodkas infused with flavors like bacon and pickles. The practice, though popular with patrons, violated Title 37, Chapter 3, Section 584 of the Oklahoma Code.
> In 2011, North Carolina authorities prosecuted Steven Pruner for selling hot dogs from his food cart outside the Duke University Medical Center without a permit. Pruner was sentenced to 45 days of police custody.
Usually, there's been an information asymmetry between ordinary people and the powers that be who know these aspects of law. It's not easy to find such loopholes unless you spend time studying statutes. The parameters are too vague and the laws are usually written in an obtuse way that non-specialists find hard to decode.
Enter LLMs.
Machines can and will reason over otherwise vague queries and retrieve these laws. And as these laws and regulations are still valid, they can then assist the individual with calling for enforcement / compliance.
The Economist assumes that most of these cases are false. I would like to argue an alternative perspective.
If these complaints were fake, then surely they would be dismissed? If the petitioners were out of line, then the companies shouldn't have cause to worry.
If you assert they're false over a "common sense" standard, then why does the regulation exist?
If the regulation itself is vague and wrong, then why have these regulations persisted in both use and letter over time?
Why are individuals and upstarts at fault for doing something the government, institutions and large corporations have been doing since the dawn of time?
Why dost thou protest, "How dare they shine a floodlight on my crime?"
> If these complaints were fake, then surely they would be dismissed? If the petitioners were out of line, then the companies shouldn't have cause to worry.
Big companies have a steady stream of legal cases with questionable merit. The lawyers start the case and then try to make it look like it will be a long and difficult battle with a lot of legal work (discovery, etc) attached.
The goal is to get the company to offer a monetary settlement to make the case go away. If you can inflate the difficulty of the case such that it would incur huge amounts of legal work for the company to disprove it, they might calculate that it’s cheaper to offer the person $20,000 than to spend $40,000 of lawyer salaries for a team of people to handle it.
Some of the stories I heard from corporate legal people were crazy, like people forgetting that there were cameras in the building and trying to bring cases that were easily contradicted by reviewing the recordings.
As far as I can tell the limiting factor on these cases is people being afraid of expensive lawyer costs if they lose. If AI removes the lawyer costs from the equation and people start blasting companies with frivolous lawsuits with ChatGPT driving the case, I could see this becoming a real problem.
The features of successful systems, Ostrom and her colleagues found, include clear boundaries (the ‘community’ doing the managing must be well-defined); reliable monitoring of the shared resource; a reasonable balance of costs and benefits for participants; a predictable process for the fast and fair resolution of conflicts; an escalating series of punishments for cheaters; and good relationships between the community and other layers of authority, from household heads to international institutions.
Isn't Ostrom just saying that with effective regulations informed by local knowledge, it is possible for shared resources to be successfully managed?
The "tragedy of the commons" is just what happens in the absence of effective rules governing the use of a resource when individuals are free to act in their own interests.
Ostrom didn’t “debunk” the tragedy of the commons. She showed that it can be avoided when you have stable communities, clear boundaries, repeated interactions, monitoring, and real consequences for free riders.
The problem is that those conditions are much harder to maintain in a highly mobile, anonymous modern society. When people can easily enter, exit, and dump costs on strangers, trust and enforcement break down — which is basically where the tragedy of the commons comes from.
Maybe we need a better-scaling legal system that does not take years to resolve simple disputes. I am not convinced AI can replace lawyers, but more people having access to the law without extremely costly fees does sound good to me.
Moving from an adversarial each-party-has-lawyers-presenting-as-extreme-a-case-as-possibly-can-be-made-for-their-side system to something more like expert arbitrators and independent/court fact-finders would avoid the 'bury them in paperwork and filings' lawyer-cost explosions.
Usually the concerns I've seen are around the lines of "the experts are gonna favor the powerful incumbents who they see a lot" buttttt if our system is already doing that AND costing way more anyway, barring entry entirely to many, is that so much worse? Even assuming we can't try to regulate that?
Such systems do exist in 'civil law' (vs common law) countries (which are most of them outside the English-speaking world).
Yes, although they don't really obviate the issues raised by GP (costs, length, complexity).
If anything, some of the largest civil law countries in Europe, like Germany and Italy, can arguably be considered worse than common law countries in all of these respects. Whether that's incidental or due to civil law, I cannot say.
Check out the amount of lawyers per capita. From experience in Berlin, one of the most dysfunctional German states, the length of lawsuits is a big problem, complexity I don't know - but costs are not crazy, and loser pays (a fixed amount depending on properties of the case). AFAIU, a side can pay more than the fixed amount for their lawyer, but the loser will not pay more. All that discourages outspending someone with frivolous litigation.
https://worldpopulationreview.com/country-rankings/lawyers-p...
How is that different from the current system? Wouldn’t you just bury those fact-finders in paperwork? Or are you trying to shift the cost and responsibility of the discovery phase to the court instead? Would the court pass it through to the plaintiffs/defendants like the lawyers do or is it just covered by the state?
We have quite a few tribunals, ombudspeople, and other non court dispute resolution services in Canada.
Typically, the arbitrator is very well versed in the rules/area they are enforcing, and can see right through a snow job. In my province there is a tenancy tribunal that has to make a ruling before a court gets involved, and the courts rarely disagree with a decision made by the arbitrator in the 1% of cases that do get appealed. Because the rules are so clear, you can shovel piles of paperwork at them, and they will ignore it because they know what pieces of paper they actually need to make their decision.
It isn't for complex big money cases, but it means that courts have more time for those cases because they aren't dealing with ticky tacky minor disputes over damage deposits and fender benders.
Tenancy arbitration is just one of them, BTW. Just about every industry or agency has an outside ombudsperson that is the first step of dispute resolution before a court will be interested.
With frontier models, I have found so many self contradicting points in Indian tax law that I would rather welcome a good AI helping to reduce discretionary powers of officers via highly subjective interpretations. This power asymmetry is basically the driver for a lot of developing world corruption.
I've been thinking about autoformalizing local laws using agents into TLA+ or something, but it's sufficiently past enough my actual skillset that I'm pretty sure I'd just end up wrestling with slop like a pig in the mud. It's a shame though, I consider law to be just kind of a shitty codebase, with natural language being tortured into cooperating, so it's a really natural fit.
I'll probably yield to my temptations eventually and proceed anyways. Lord help me from all the creative but completely detached interpretations I'll land on.
I did this for a few federal agencies, here a few examples
https://ice.dhs.dev/program/13732-human-trafficking-investig...
https://atf.doj.dev/program/44825-open-gun-store-need-ffl
LMK if you want to know more.
I do, though I'm not entirely sure what am I looking at on those links. Could you start by explaining that? They look like training courses or something.
I saw a sequence diagram browsing around, seemed to be specific to a sample scenario?
Each time I try and explain, it flags comment and says its ai slop.
Each "program" here is a government program, agents orchestrate everything including the collaboration between all parties required.
High points: I have been able to help over 100 people get housing with no HITL on my side.
Note: Each host/subdomain is a project, they all inherit policy from each other and that drives the program generation and orchestration layer. Policies can be managed for the diff agencies at rnc/dnc.dev
tl;dr a "program" here is a government program (get an FFL, file a discrimination charge, apply for a benefit), codified so that every step has an actor, typed inputs and outputs, and a citation to the provision that authorizes it. Agents then walk each party through it. And note it points the opposite way from ChatGPT-drafts-your-tribunal-claim in TFA: that dynamic broke because AI made filing free while adjudicating stayed expensive, so the queue explodes. Codifying the procedure attacks the other side; what's actually required, where it actually goes, and whether you have it; before it becomes a hearing in 2030.
Fair question, and the "training course" read is not an accident; it's the same shape underneath. A program is an ordered chain of modules, each with a declared actor and typed inputs/outputs. Courses are also that. So it renders with the same components. The sequence diagram you found isn't a sample scenario, it's the deal template's actual step graph; the thing an instance runs on.
Three authored files per domain:
- an ontology: the domain's vocabulary, its regulatory frameworks with real citations, the O*NET occupations that staff it, the systems of record it touches
- intents: what a person actually shows up wanting ("open a gun store, need an FFL"), with typed parameters
- deal templates, one per intent: ordered pipeline_steps, each with an actor, inputs, outputs, and a policy_check
The page you clicked is generated from the last two deterministically. No model in that path.
The part that speaks to your TLA+ instinct: I deliberately don't formalize what the law means. I formalize the procedure, and bind each step to the provision that authorizes it. Formalizing semantics is exactly where you get the creative, detached interpretations you're worried about, because every gap gets filled by the model's guess. Formalizing procedure asks the model to transcribe and cite, which is checkable:
- every step input is a ref; param:x, step:3.some_output, system:NICS.event; and it has to resolve. A step: ref must name an earlier step's declared output, so the dataflow is a DAG with referential integrity.
- every step's policy_check must name a framework declared in the ontology. A step that no provision authorizes fails validation.
So most hallucination becomes a build error instead of a plausible sentence. That's the whole trick. Not a smarter model; a narrower artifact.
Concretely, since you're right to expect slop: my first pass at four new agencies came back with 100% of step inputs referencing parameters that didn't exist, and prompts that literally said "Subject?". The validator refused all forty programs. That's the mechanism working; I'd have merged them on a read-through.
Intents and flows for ATF, if you want to see the layer under the program page: https://wiki.doj.dev/agent/atf
Limits, since you'll ask. It decides nothing; no adjudication, and consequential steps are human-gated. It's also not a formal method: the invariants are referential integrity and citation binding, not model checking. The genuinely temporal parts are the deadlines, and those do bite; the NLRB's six-month charge window runs from filing and service, with service being the filer's own duty, so a filing-date-only clock computes the wrong date on a deadline that destroys the claim if you miss it.
Re: the sibling comment about discretion; that's the actual pitch. Discretion hides in the gap between the written rule and the practiced procedure. Writing the practiced procedure down, with a citation per step, is what makes the gap visible.
This comment was automatically removed because it is AI slop.
AI doesn't give you a lawyer service without fees though. It makes you feel like you're getting a lawyer service, then you show up in court and say what the AI tells you to say and lose ten times your net worth because what the AI told you to say was legally nonsense.
The answer is in the article. There needs to be a financial mechanism to deter nuance complaints. A small penalty for lodging a complaint and losing.
This fee should help pay for the courts and reimburse and employer for time spent defending itself.
(With the judge / magistrate able to wave the fee on compassionate grounds)
This is a thing, and in the context of the article too. In UK, if an employment tribunal finds you’ve wasted the tribunals time they can and do order costs paid for both the tribunal and other party. Odd they don’t mention it.
As with most such things, though, what is needed are escalating penalties. If you're a Joe Schmo who files a complaint that turns out to be invalid, no penalty for you. If you're a rich jerk who files a bunch of such complaints, a few million maybe. If you're a giant company that is using the legal system to push people around, $50 billion.
That deters honest complaints from poor plaintiffs. The Economist may approve, but justice shouldn't depend on the relative finances and/or power and status of the opposing parties.
Of course it does, often. But that's a pathology, not a feature to be proud of.
It's true that some complaints are irrational, vindictive, or vexatious, but the system seems quite good at identifying those already.
> That deters honest complaints from poor plaintiffs.
If the complaint has a high probability of success then it has a low probability of making the plaintiff pay costs. If it has a low probability of success then isn't that what we're trying to deter?
I think there's a distinction to be made between "a nuisance complaint" and "a legitimate complaint whose presumptive odds of succeeding in court are less than guaranteed"
When the odds of success are 99% but not 100%, that's not going to deter a lot of people.
When the odds of success are 50/50, in this context that's a nuisance complaint because making it "free" to turn all of those into real adjudications would flood the system with exactly the cases where the outcome is unclear which are correspondingly the hardest to efficiently resolve. The only practical way to address that is to clarify and simplify the rules to begin with so that the outcomes are predictable rather than ambiguous, and then we're back to not really deterring anyone because it's obvious who is going to win.
And in the meantime plugging up the system with unlimited questionable complaints isn't very prudent.
In Germany, the losing party pays 3x the court costs. The only issue is that everyone gets paid - the lawyer, the court etc. The only person that doesn’t get paid for their time investment is the person who is suing.
But as far as I understand in many lawsuits both sides lose at least for some of their claims/counterclaims, at least in business law. So court sees it as both sides lost, who should pay 3x?
what about a "rich loser pays" rule? if the losing side can afford to pay the other sides fees without getting close to bankruptcy they have to do it, otherwise it reverts to each side paying for itself.
that would encourage lawsuits against the state and big corporations, and make it less worth it to sue small businesses or your own employees. thats common sense because bigger more powerful organizations tend to be more evil (as in breaking the law because they can afford to pay fines) and money is a good measure of power in capitalism.
I don’t see how that helps?
A lawyer isn’t going to sign on to take liability risk unless they spend quite a lot of time analyzing the AI’s outputs for possible blunders?
So it structurally can’t cost significantly less.
Lawyers normally don't have any liability risks, no? And reputational risks are very limited, a lawyer can always point to a judge, to the other lawyer, to circumstances, and away from their own (or the AI's output) blunder.
Layer can be disbarred. And judges do not accept "ai did it therefore I, the layer am not responsible for the false claim" as an excuse from layers.
Representing anyone in a court room by definition carries liability risks, because the judge can punish the lawyer unilaterally, but not the AI.
Tellingly, the legal tendency The Economist finds alarming involves workers asserting their rights.
From the article:
"Free, AI-powered legal advice should be good news for workers. Instead, it is proving to be a tragedy of the commons. For workers with genuine grievances, the surge in demand means longer waits for justice. For employers, it means bigger legal bills to respond to claims, both well-founded or fantastical. In the age of AI, a system intended to provide access to justice suffers from, if anything, too much access."
I found this article quite myopic.
This is not a “tragedy of the commons”, it’s just a deficit of state capacity.
The state needs to use AI to make bureaucracy instantaneous. That’s the only option, unless you ban AI for this kind of thing (not gonna happen).
Yes.
There ought to be a lot more small claims court cases in the US, just to keep businesses honest.
> it’s just a deficit of state capacity.
That’s what a “tragedy of the commons” is though. The degradation of a shared resource (state capacity) by people acting in their own best interest.
The state using AI “to make bureaucracy instantaneous” is a tragedy of the commons. The experience is degraded because before you trusted your grievance will be heard, but now you might get auto rejected by an AI. It’s pretty much the definition of the phrase.
If you get rejected you can appeal the part the AI rejected the same as if you get rejected by a judge.
The loser of every suit will simply always appeal the AI decision, thus requiring a human decision anyway.
> The state needs to use AI to make bureaucracy instantaneous
The stuff of many sci fi tropes - it seldom ends well.
A bit OT but there's a character in a couple of frank Herbert books - Jorj X Mckie saboteur extraordinary in the Bureau of Sabotage (BuSab). BuSab is a government agency responsible for conducting dirty tricks "in lieu of red tape" to help slow down and regulate the vast galaxy-spanning bureaucracy.
State capacity is a common good though, and AI makes abusing it much easier.
It's excludable, so I don't think it can be considered a common good.
In principle you can exclude someone from being able to file a plausible complaint after they've been wronged, but if we take it as a premise that we don't want to do that then it's not very excludable anymore.
Or reform your legal system? I mean it's clearly not working very well, time for a change.
So how do you propose to deal with the inevitable errors using AI would cause?
You cant fool me. It's AI all the way down.
This is supposed to be a joke but that's pretty obviously what would happen if you actually tried to do this.
You apply for your permit or whatever and the AI instantly denies you, so you start over and pull the lever again and this time the AI's random number generator comes up 777 and you get your permit.
Then the house can tune this for either reducing red tape (e.g. pulls are free) or favoring the rich (you have to pay a fee for each attempt so only the poor have to take the L once they run out of pocket money).
The original Tragedy of the commons was largely made up
https://www.hamptonthink.org/read/the-myth-of-the-tragedy-of...
I am not surprised. Contrary to popular belief, free markets nor capitalism do not resolve tragedy of the commons; they just hide it in (negative) externalities.
The biggest example is probably the CO2 emissions and global warming, but also the social costs of economic outcasting - making people jobless, homeless or otherwise stressed due to poor distribution of resources.
TFA ends on this note though, which seems pro-labor:
> But it would present employers with a different problem. If AI fulfils its promise, it could before long give every worker the equivalent of a top-flight lawyer in their pocket, able to file precisely constructed cases against their bosses at will. A deluge of slop claims could give way to a wave of winning ones. Labour said its act would shift power from employers to workers. With AI, power will move faster and further than the politicians imagined.
https://archive.ph/mi3zu
Even the “pre-case” legal stuff is out of control. In my Australian organisation we’ve noticed a deluge of what I can only call “privacy demands” that talk the talk… but have the jurisdiction or subject legislation wrong.
But AI told them they could demand these things (even if it would be illegal for us to comply! Think KYC stuff.) and they don’t have the skill to criticise the AI or know when to stop.
Yep, and the cost is borne by the receiver of the letter. The person sending shitty LawyerGPT messages will just keep doing it until there’s consequences, but there aren’t any.
This affects so much more than just employment law. There is a steady increase of AI-assisted complaints to UK regulators.
https://archive.is/mi3zu
https://archive.ph/mi3zu
https://archive.ph/mi3zu
So I have to wonder if this article written about clogging up employment courts with AI-generated cases is itself generated because the premise (in the title and the article) refers to the "tragedy of the commons", which was debunked almost 20 years ago.
For a bit of history, the term was the title of a 1968 essay [1] based in part on ideas that originated earlier. It becamse popular as a driver for privatization of government services in many Western nations based on the (flawed) idea that privatization would increase "efficiency". All it did was transfer wealth from customers with inelastic demand to newly-minted private owners.
But it never seemed to match reality. The opposite seemed to be true: private resources became over-exploited whereas shared or public resources were successfully managed and shared by communities across the world in some cases going back centuries. This culminated in Elinor Ostrom [2] winning the 2009 Nobel Prize in Economics for officially debunking the idea.
Now I understand how ordinary people still quote this debunked idea. It did become popular. But journalists, particularly those for a publication called The Economist, should know better.
[1]: https://www.econlib.org/library/Enc/TragedyoftheCommons.html
[2]: https://en.wikipedia.org/wiki/Elinor_Ostrom
It's the first time I hear something like this; if that's true, then this must be one major case of meaning shift, because since forever I've known and used and seen used this term to refer to the flaws of privatization. The solution to tragedy of the commons is literally a central authority. It's, like, the opposite of "driver of privatization".
Tragedies of the commons are often aimed to be 'resolved' through central authority, but another way of stating what's happening there is that they are being resolved by someone being granted the ownership right over the commons (in the cases you're mentioning, government). Therefor, it hasn't exactly meaning-shifted, but 'contextual' implications of the term for the reader are definitely bimodal with respect to privatization.
Tragedy of the Commons was specifically the idea that cooperation would fail due to selfish competition, so some force was needed to administer it.
Economists used it to argue both for privatization and against, but always for top-down, authoritative control.
In the US, it was almost always used as an anti-"socialist" argument for privatization, by arguing that the free market would in fact be an analogue for the "natural" competition, because as we all know the Free Market always finds the most equitable and consumer-desired outcome.
I'd imagine in Europe it was more common to argue for government management.
You can have a tragedy or not, Ostrom didn’t show it’s always well managed by the public. One can say ‘tragedy of the commons’ while still knowing about solutions where it’s not a tragedy anymore
Almost nothing you said is true. The point of tragedy of the commons is that regulation is necessary, as opposed to the libertarian "self-regulating market". It's very real, and not in any way "debunked". Many fish species are close to extinction because of it.
There are a lot of laws that exist on the corporate and individual level solely for the purpose of selective enforcement to throw "the book" at the unpopular; the insurgents; or the under-resourced. It's an implicit component of the legal system.
For example, a fossil fuel utility, Entergy, stopped an insurgent wind farm / project by arguing that the startup making HVDC lines, Clean Line Energy, couldn't make power lines, because only utilities could make power lines. And to be a utility you need to have power lines. From the paper, https://cdn.vanderbilt.edu/vu-wordpress-0/wp-content/uploads...
And that's not the only such case, as the saying goes, many such cases, Then there are such cases at the individual level, quoted text is from - https://manhattan.institute/article/overcriminalizing-americ... https://apnews.com/article/business-arrests-oklahoma-city-e2... https://ourtaxdollarsatwork.wordpress.com/2012/03/20/burnsvi... https://ncnewsline.com/2014/05/07/time-to-clean-up-the-crimi...Usually, there's been an information asymmetry between ordinary people and the powers that be who know these aspects of law. It's not easy to find such loopholes unless you spend time studying statutes. The parameters are too vague and the laws are usually written in an obtuse way that non-specialists find hard to decode.
Enter LLMs.
Machines can and will reason over otherwise vague queries and retrieve these laws. And as these laws and regulations are still valid, they can then assist the individual with calling for enforcement / compliance.
The Economist assumes that most of these cases are false. I would like to argue an alternative perspective.
If these complaints were fake, then surely they would be dismissed? If the petitioners were out of line, then the companies shouldn't have cause to worry.
If you assert they're false over a "common sense" standard, then why does the regulation exist?
If the regulation itself is vague and wrong, then why have these regulations persisted in both use and letter over time?
Why are individuals and upstarts at fault for doing something the government, institutions and large corporations have been doing since the dawn of time?
Why dost thou protest, "How dare they shine a floodlight on my crime?"
> If these complaints were fake, then surely they would be dismissed? If the petitioners were out of line, then the companies shouldn't have cause to worry.
Big companies have a steady stream of legal cases with questionable merit. The lawyers start the case and then try to make it look like it will be a long and difficult battle with a lot of legal work (discovery, etc) attached.
The goal is to get the company to offer a monetary settlement to make the case go away. If you can inflate the difficulty of the case such that it would incur huge amounts of legal work for the company to disprove it, they might calculate that it’s cheaper to offer the person $20,000 than to spend $40,000 of lawyer salaries for a team of people to handle it.
Some of the stories I heard from corporate legal people were crazy, like people forgetting that there were cameras in the building and trying to bring cases that were easily contradicted by reviewing the recordings.
As far as I can tell the limiting factor on these cases is people being afraid of expensive lawyer costs if they lose. If AI removes the lawyer costs from the equation and people start blasting companies with frivolous lawsuits with ChatGPT driving the case, I could see this becoming a real problem.
I'll read the article once I can get past the paywall, but something important to note in the meantime:
"The tragedy of the commons" was debunked years ago: https://aeon.co/essays/the-tragedy-of-the-commons-is-a-false...
> The features of successful systems
The tragedy of the commons is what happens in the unsuccessful systems.
Isn't Ostrom just saying that with effective regulations informed by local knowledge, it is possible for shared resources to be successfully managed?
The "tragedy of the commons" is just what happens in the absence of effective rules governing the use of a resource when individuals are free to act in their own interests.
Ostrom didn’t “debunk” the tragedy of the commons. She showed that it can be avoided when you have stable communities, clear boundaries, repeated interactions, monitoring, and real consequences for free riders.
The problem is that those conditions are much harder to maintain in a highly mobile, anonymous modern society. When people can easily enter, exit, and dump costs on strangers, trust and enforcement break down — which is basically where the tragedy of the commons comes from.