The End of the Expert Pipeline?

Something strange is happening to Britain's employment courts. A worker gets fired, types their situation into ChatGPT, and within minutes receives a 200-page legal claim complete with citations to obscure provisions they never knew existed. The filing is free. The judge who has to read it? Also free — and there aren't enough of them.

The result is a queue stretching to 2030.

That's not hyperbole. A case filed today in a British employment tribunal may not be heard for four years, per The Economist's August 6 report. The backlog jumped 55% to 64,000 unresolved cases in a single year. Claims rose 39% through March 2026 alone. Interim relief applications — an emergency measure that used to get about 20 filings nationally each year — have gone roughly a hundredfold. The Economist, 2026-08-06

The proximate cause isn't more layoffs, more disputes, or a sudden surge in worker anger. It's ChatGPT.

But the Economist got only the headline right. The deeper story isn't about AI flooding a court system. It's about AI exposing a structural asymmetry that every professional system has been sitting on for decades — and about how the very people who are supposed to manage that asymmetry are the ones the asymmetry is destroying.

The mechanism: generation got cheap, review didn't

Interim relief is a real, narrow emergency measure in UK employment law. A judge can order a company to reinstate a fired employee or keep paying their wages while a full case is pending. Historically, it was invoked for whistleblowers or trade-union officials facing retaliation. The kind of provision that, before 2026, effectively required a specialist employment lawyer to know it existed at all. Obscurity was doing a lot of the gatekeeping work.

That's exactly the failure mode LLMs are good at closing. A worker who's been fired doesn't need to have read employment law; they need to describe their situation to a chatbot and get pointed at the right provision. That's a genuine, non-trivial access-to-justice unlock. The Economist notes that if AI "fulfils its promise, it could before long give every worker the equivalent of a top-flight lawyer in their pocket." The Economist, 2026-08-06

But here is where the commons comes in. The court system was sized for a world where filing a claim required enough friction — time, money, or specialist knowledge — that volume stayed within institutional capacity. AI removed that friction on the input side without adding capacity on the output side. Every claimant benefits individually. Every claimant pays nothing for filing. The cost is absorbed by the queue, and the queue has no owner.

This is what Garrett Hardin meant in 1968 when he named the problem: when everyone has free access to a finite resource and no one bears the full cost of their use, the resource gets consumed faster than anyone planned. Hardin, Science, 1968

Except Hardin was writing about physical commons. The AI version is stranger, because the "herd" is multiplying itself. Each new case spawns a response, which spawns a reply, which spawns a motion. The herd is multiplying. Each cow you add breeds another before you've even turned around.

The deeper layer: expertise itself is the commons

Here's where the story deepens. The Economist framed this as a queue problem: too many claims, too few judges. But a paper published in Human Resource Development Review in July 2026 offers a deeper reading. Nolan Lovett's "The Tragedy of the Cognitive Commons" argues that the real commons isn't the queue — it's the deep domain knowledge that lets judges, arbitrators, and expert reviewers do their jobs well. Lovett, HRDR, 2026 arXiv

Lovett introduces two forms of expertise: Internalized Mastery — deep domain knowledge built through sustained practice — and Distributed Mastery — fluency in orchestrating human-AI systems. The two are not the same. More importantly, effective Distributed Mastery depends on Internalized Mastery as a foundation: you can't reliably judge whether an AI's output is right unless you understand the domain deeply enough to recognize when it's wrong. Lovett calls this the "Validation Tether." Lovett, HRDR, 2026

The tragedy of the Cognitive Commons is that every organization's rational move — adopt AI, eliminate entry-level positions — depletes the shared expertise pool that all organizations depend on. An organization that stops hiring junior lawyers to learn employment law can still hire experienced ones from the shrinking pool. But the deeper the pool shrinks, the weaker everyone's validation capability becomes. The people who are supposed to manage the queue are the ones the queue is destroying.

This isn't just about law. Lovett's framework applies wherever AI substitutes for the cognitive labor that historically trained novices: financial analysis, software engineering, clinical medicine. Early evidence is already visible. Workers ages 22 to 25 in AI-exposed occupations experienced a 16% relative decline in employment between October 2022 and September 2025, even after controlling for firm-level shocks, while workers aged 35-49 in those same occupations grew by over 8%. Brynjolfsson et al., Stanford DEL, 2025

The pattern: AI eliminated the positions that used to be the training ground for expertise. The people who remain are more productive on routine tasks but shallower on the ones that matter when things go wrong.

Is "tragedy of the commons" even the right label?

The Economist's framing was useful but incomplete. The HN thread had a real fight about whether the concept was even applicable. A user named jmyeet claimed the concept was "debunked almost 20 years ago" by Elinor Ostrom, who won the 2009 Nobel Prize for showing that shared resources are often successfully managed by communities. Ostrom, Governing the Commons, 1990

TeMPOraL pushed back: "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."

kian clarified: tragedies of the commons are often resolved through central authority, but that's also just someone being granted ownership rights over the commons. So the term hasn't meaning-shifted; the contextual implications are just bimodal with respect to privatization.

The resolution? Ostrom didn't debunk the tragedy of the commons. She debunked the idea that it's inevitable. Communities can self-manage commons. Governments can manage them. Markets can manage them. The question isn't whether the tragedy exists — it's what institutional arrangement actually prevents it.

Applied to the AI courts: the queue is the commons. The question isn't whether it will overflow. It's whether we'll fix it with fees, with arbitrators, with AI-assisted judges, or with something we haven't named yet.

What gets the job done

The most structurally useful proposal from the HN thread came from a user named dozerly: maybe we need a better-scaling legal system that does not take years to resolve simple disputes. HN thread

In other words: the problem isn't AI. The problem is that we built a system to handle a few thousand cases a year and then told ourselves it was fine because no one filed more. Then AI showed us what "more" actually looks like.

There was also majormajor's proposal: move from adversarial each-party-has-lawyers-presenting-as-extreme-a-case-as-possibly-can-be-made-for-their-side litigation toward expert arbitrators and independent court fact-finders. The standard objection — that experts will favor the powerful incumbents who appear before them constantly — got acknowledged directly, with the counter that the current system is already doing that and costing way more anyway.

The sharpest technical insight came from a user called arionhardison, who has been building systems that autoformalize government procedures. Each "program" — file a discrimination charge, apply for a benefit, open a gun store — is an ordered chain of steps, each with a declared actor, typed inputs and outputs, and a citation to the provision that authorizes it. Agents walk each party through the chain.

The part that speaks to the AI-tragedy problem: "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."

Most hallucination becomes a build error instead of a plausible sentence. That's the whole trick. Not a smarter model; a narrower artifact.

It's not just Britain

The Economist cites parallel cases: Dutch municipal-tax appeals, the Canadian privacy regulator, city parking-ticket tribunals in every major city. Each one was sized for a world where filing a claim required friction. AI removed the friction.

The US faces a similar crisis. One federal judge called AI-generated lawsuits "an existential threat to the federal courts."

The Pakistan study showed it can go the other way: judges with AI tools and training processed more cases, faster. It cost $1 and returned $38.50. Pakistan study

The pattern is the same everywhere: AI made generating a claim cheap. It didn't make reviewing one cheap. Someone still absorbs that cost.

Every system has this problem

Every tech team has seen this. AI made writing a pull request cheap. It didn't make reviewing one cheap. Someone still reads every line. AI made writing support tickets cheap. It didn't make answering them cheap. Someone still reads every ticket. AI makes writing legal claims cheap. It doesn't make adjudicating them cheap. Same equation.

The commons in each case is the attention of the person on the other end of the queue. And attention is finite. Every system that generates work faster than it consumes it will eventually back up.

This isn't an AI problem. It's a generation-vs-review problem that AI just accelerated by a factor of ten.

The fix isn't to slow generation. It's to speed review — or to make the queue deeper, flatter, and more resilient to volume. The arionhardison systems do this by formalizing procedure instead of semantics. The Pakistan judges do it by using AI to triage and summarize. The UK tribunals might do it with cost-shifting, with expert arbitrators, or with something else entirely.

What we know for sure: every system built for analog volume and powered by digital generation is about to learn whether its review capacity can scale. Most won't. Some will. The ones that don't will be the ones we call tragedies of the commons when they're already overflowing.

The Real Tragedy

That worker who types their situation into ChatGPT and gets a 200-page claim back? They're not the problem. They're the signal. They found a provision that used to require a specialist lawyer, filed it for free, and moved closer to justice. The tragedy isn't that they filed. The tragedy is that everyone else has to wait four years to find out whether they were right.

The pasture wasn't overgrazed because the cows were greedy. It was overgrazed because the fence fell down and nobody rebuilt it.

The UK's employment tribunal system processes cases using a system called ET1 for claims and ET3 for responses. As of 2026, there are roughly 170 employment tribunal judges across Britain, each handling between 600 and 1,200 cases per year depending on the region.