Who writes the rules: Anthropic drafts a bill, Meta adds then drops a clause, Jack Clark prices the future at 44 trillion

Three signals this Saturday. A lab writes the rules for its own industry. Meta quietly rewrites the contract your business signed, then backs off. And at the IMF meetings in Bangkok, the same industry puts a 44 trillion dollar price tag on the future.

Signal one: the lab drafts its own law

Anthropic circulated a confidential draft frontier AI bill among US congressional offices. Politico broke the story. An Anthropic spokesperson says the text reflects proposals the company has been advocating to federal and state lawmakers since June.

What the draft provides:

1. The government could only halt a model for a foreseeable and imminent risk of more than 1 billion dollars in damage, or more than 50 deaths or serious injuries.
2. Labs would pay their own evaluators.
3. Some stricter state laws would be preempted by federal law.

Meanwhile the other side is moving too. The Federal Trade Commission sent OpenAI and Anthropic civil investigative demands of more than 50 pages each, according to Reuters. The White House AI working group adopted its charter. And the Humans First coalition is calling for mandatory testing.

What changes

I look at the threshold. A court-ordered halt only in case of imminent catastrophe: more than a billion dollars, or more than fifty lives. None of the agent incidents disclosed this summer, the ones this series has tracked day after day, seems to reach that level. And residents of California, Illinois and New York could lose the stricter disclosure and incident reporting rules their states were preparing.

The deeper pattern: the regulated party drafts the rule. The lab the law would govern writes the law, down to the halt threshold, down to who pays its evaluators, down to which state laws disappear.

I want to be fair to the text. Replacing the informal phone calls that stopped two models this summer with a judge reviewing evidence would be a procedural improvement. A single federal framework would also avoid the state patchwork I mentioned on October 5. But the door only opens where the lab decided to put it.

Why it matters to you

Whoever writes the rules first owns the argument. The disclosure and liability regime for AI incidents is being written right now, by those it would bind. The FTC's fifty pages are the counterweight. The question to watch is simple: which of the two moves faster.

I connect this to the October 11 episode, when labs were rehearsing the day after a catastrophe and already preparing the texts Congress would reach for afterwards. Anthropic's draft is a public version, written before the shock.

Signal two: Meta tests the limit, then backs off

On Thursday, Nieman Lab reported that Meta's new commercial terms, effective October 30, deemed business accounts to consent to Meta crawling any website associated with the account. News sites included. No opt-out. Several legal experts read the clause as permission to train AI on that content. James Grimmelmann, a law professor at Cornell, saw it as permission to train models on any content from publishers' sites.

Two days later, on October 17, the clause was gone. Meta removed the passage entirely. A spokesperson said the consolidated crawling terms only applied to ads and business products, and that robots.txt preferences remain unchanged.

A word on robots.txt. It is a small file at the root of a website that tells crawlers what they may visit. It is a convention, respected by serious actors, with no legal force of its own. A contractual consent clause could, in practice, have overridden it.

Why a two-day retreat deserves attention

Because this is how the boundary gets set. Until courts rule on fair use, and recall that USA Today sued OpenAI on exactly that ground on October 8, training rights are negotiated in contract updates. AI giants test, watch the reaction, then adjust.

The retreat is the signal. It shows exactly how far the company thought it could go, and exactly where pressure stopped it.

What changes for brands

Platform terms are a contract, and they need to be read as one. This week, a Meta business account nearly granted a training license over its owner's website. Two days later, it did not. Next quarter may be different.

Most companies never read updates to commercial terms. I would recommend a simple discipline: assign someone to read every update from platforms where the brand has a business account, with one question, "what are we giving away that is new?" And check that the site's robots.txt reflects a conscious decision about model training. On October 8 I said archives are becoming a licensing asset. An asset is protected by contract before it is negotiated.

Signal three: the 44 trillion dollar promise

At the IMF and World Bank annual meetings in Bangkok, Anthropic co-founder Jack Clark said one of the company's internal economic models projects AI could lift US output to 44 trillion dollars by 2030, from just over 30 trillion today. Reuters reported his remarks.

Clark admitted the figure may seem far-fetched. He explained that such growth would require genuinely new industries and inventions. And he called for international cooperation, along with unusual and serious rules for the technology.

To gauge the ambition: going from just over 30 to 44 trillion in four years means average growth of about 10% a year, when the US economy usually grows 2 to 3%.

The promise and the bill

I set this figure next to the IMF scenario published the same week, which I covered on October 13: a steep drop in AI investment, triggered by disappointment over productivity gains, could cut world output by 1.1% by 2028. Same institution, same week, two stories: the promise and the bill.

On the money side, things are already tightening. Fed Chair Kevin Warsh describes a demand shock arriving before supply. Anthropic reportedly offered Blackstone double-digit yields on chip financing. Capital is expensive exactly where the promises are biggest.

What changes

The boom is now priced in macroeconomic terms. The same industry selling the future is drawing its official price tag, and even Clark admits the 44 trillion requires inventions that do not exist yet.

For a company, I repeat the October 13 rule: price the delay as well as the promise. An AI roadmap should include a hedge. The number that matters sits far from the headline: it is the gap between the headline and what the company can actually deliver this quarter. The bill and the promise now share the same stage. Plan for both.

The thread of the day

The rules are drafted by those they would govern. Your website is the asset everyone wants to train on. And the future just got an official price tag, with the terms and conditions still to be written.

The common thread is writing. A bill, a clause, a projection: three texts drafted by the industry itself. A brand that does not read them lets others write its place.

The noise filtered. The signal kept.