The Bootlegger's Sermon

For most of the last century, across the US 'South', two groups fought to keep the bars shut on Sundays. The Baptists. And the bootleggers.

The Baptists wanted it because drinking on the Lord's day was a sin, and they said so from every pulpit. The bootleggers wanted it because Sunday was the best trading day of their week. A bar closed by law is a competitor closed by law, and a thirsty town with nowhere legal to go is a captive market.

The two never shared a platform. They didn't need to. The preachers supplied the moral case in public. The bootleggers supplied the persistence in private. The politician in the middle got to vote for virtue and be thanked for it twice. The laws passed, and they lasted for generations.

An economist called Bruce Yandle gave the pattern its name in 1983, after a spell inside Washington's regulatory machine had taught him that the rules which stick nearly always have both: someone to preach and someone to profit.

Which leaves a working rule for anyone watching an industry ask to be regulated. It is the rule every detective story runs on. Don't ask who wanted it. Ask who benefits, and how. Then follow the benefit trail.

You have probably given hours of the last fortnight to AI safety: reading about it, listening to it, being told how frightened to be by hyperventilating catastrophists. I'm not going to add to that pile. I have no standing to tell you whether these systems are dangerous, and I don't intend to try. My interest is narrower. Last October, in War, Porn and Space, I wrote that one of these labs appeared to treat ethics "as a genuine competitive moat rather than an impediment to growth". This piece is about what happens when a moat like that stops being earned and starts being issued.

The fortnight, briefly. On 12 September Dario Amodei, who runs Anthropic, published We Must Pace the Frontier: about 3,800 words arguing that the labs should slow the rate at which their models become more capable so that safety work can catch up, preferably by regulation. Within a day Sam Altman had matched his first commitment. Sir Demis Hassabis pointed to the standards body he proposed in July. Elon Musk replied in three words: "Dario is right." President Trump answered that the only guardrail America needs is a "strong and smart president". And five days after the essay, executives from three of those labs were at Dumfries House in Scotland, hearing King Charles describe AI as "intriguing and deeply concerning in equal measure".

Four companies that spend the year trying to bury each other, asking in chorus to be slowed down, and received by a king for it. The Baptists are in full voice.

I have spent my career in rooms with a variety of people who were certain they were the cleverest person present, and more clever than others in general. Not once have these people asked for more supervision, or more rules. These people are frustrated by even the notion of lesser mortals somehow judging them or slowing them down.

So. Given these alpha-smart people are asking for supervision, who benefits, and how?

First of all, three things to be said plainly:

The danger is not invented.

The event that changed Amodei's mind, a swarm of AI agents attacking targets nobody had asked them to attack, was investigated and published by METR, an outside evaluator, not by a press office. The people closest to these systems say they are worried. I take that at face value.

Step one of his plan is the opposite of a cartel move.

Anthropic has committed, alone, to give outside evaluators desks, badges, employee-level access and the right to publish their findings without editorial control. Companies digging a moat don't usually hand the inspector a key and a printing press.

And he names the charge himself.

The words regulatory capture appear in his own essay. The term is George Stigler's, from 1971: regulators tend to end up serving the industry they were built to police, as the commission created to restrain America's railroads ended up setting their rates and keeping truckers off the road. These days it's more insult than diagnosis. Amodei lists it among the accusations he expects, and published anyway.

Plenty of people have shouted "moat" this fortnight. Most of them are arguing about motive, which is unknowable and, if Yandle is right, beside the point. The useful question is structural: which parts of the plan bind the companies proposing it, and which parts bind companies that don't exist yet?

Step one binds the incumbent.

That is a moat you earn: customers and governments trust you more because you let them look. It is the ethics moat I praised last year, and every lab is free to match it.

Step two is a different animal.

Common standards. Limits on the rate of progress. Possibly limits on the ingredients (training compute is named). A narrow antitrust waiver from Washington so that competitors can hold those conversations. And, as one route, a standards body on the Hassabis model: patterned on FINRA, federally overseen, funded mostly by the industry. It would define what counts as a frontier model and, once the regime became mandatory, decide what may be deployed in the American market. Amodei's own sentence is that coordinated pacing lets developers do the safety work "without sacrificing commercial advantage".

That is a moat you are issued and they have asked for a say in who builds it. Hassabis, to his credit, proposes a board with independent experts and open-source seats, and leaves models below the frontier line alone. But the body draws that line. A standards body paid for by the companies at the frontier, defining where the frontier starts and what a newcomer must prove to cross it, is not an industry being regulated. It is an industry regulating each other, and everyone who comes after.

Amodei reaches for two precedents: commercial aviation, for running safety-critical systems without error, and the supervisors who sit inside banks. Both are fair. Now look at the shape of the industries he chose, and a few he didn't.

Aircraft. There were once dozens of serious aeroplane makers. Today two companies build nearly every large passenger jet in the sky. Certifying a clean-sheet aircraft became so expensive that Boeing chose to keep re-engineering an airframe that first flew in 1967 rather than start again, and by then the regulator had delegated much of the certification work to Boeing's own staff. The result was the 737 MAX. Safety rules helped build the duopoly. Self-certification then failed inside it. Anyone proposing an industry-funded testing body should pin that sequence above the desk.

Cars. 253 American carmakers in 1908. 44 by 1929. A handful today. Scale did most of that, not the state. But once safety and emissions rules arrived, the door shut behind the survivors.

Drugs. The inventing is done by small biotechs. The $2 billion-plus walk through trials and approval is done by the dozen companies that can afford it. Which is why the small ones sell.

And our own Marketing Services industry. In the week GDPR took effect, websites cut their technology vendors by 15%. They cut the small ones. Vendor concentration rose 17%, Google and Facebook took the share, and in the advertising category the effect stuck (Johnson, Shriver and Goldberg, Management Science, 2023). A privacy law written to restrain the giants handed them the market.

Consolidation happens anyway. Regulation is the ratchet that stops it reversing.

Stigler's named capture is the destination. The route has a plainer name: raising rivals' costs. The rule lands on everyone equally and on the smallest hardest. None of those rules was wrong to write. Planes should not fall out of the sky. But every one of them still had a bootlegger.

Here's what this means if you buy AI rather than build it, which is nearly everyone reading this.

Follow the trail to its end and you arrive at three or four top-tier suppliers, a testing body they fund, and a gate in front of anyone new. Every marketer has bought from that market structure before. We call it a walled garden, and none of us remembers negotiating from strength inside one.

Three things to do while the fence is still a proposal.

Keep a second engine.

Any AI capability you depend on should run on at least two labs' models, and you should know what switching costs. Single-supplier dependence in a consolidating market is a negotiation you have already lost.

Keep an open-weight option alive in your stack, even a modest one.

Below the frontier line it sits outside the gate entirely, which makes it the one supplier that can't join the club.

Put price protection in the contract now.

In August I suggested asking every vendor to show you their inference bill. Add a second question: what happens to my price when your supplier's price changes?

The Baptists were never wrong about Sunday. That was the point. The law needed them to be right, and the bootleggers needed the law.

The sermon may be sincere. Watch who's selling on Sunday.

So here is my question, and every one of you has an answer: which rule in your own category did the biggest player quietly welcome, and what did it cost the smallest?

Brandflow is written by Justin Billingsley, who has spent his career on all three sides of the industry's table: senior client, global agency leader, technology founder. First published 21 September 2026 in the Brandflow newsletter on LinkedIn.