Subject to approval

On 18 October Anthropic, the company behind Claude, announced a fund for the public good. The headline number is easy to remember: 5 per cent of the company's shares.

Here is what the announcement says, in its own words. "Subject to stockholder approval, 5% of Anthropic's shares will be committed today, with an additional 2% added at the board's discretion for every 50% increase in the company's value above a set baseline." The fund is a wholly owned company, "rather than a 501(c)(3) or other format", chosen "to give adjustability between different forms of giving". The equity, Anthropic says, is what tech companies usually hand their executives as escalating pay awards. It adds that roughly 20 per cent of its shares are now dedicated to the public good, counting the co-founders' pledge of 80 per cent of their own equity, an employee equity match that about half of staff take part in and the new fund.

What the fund might do is described in broad strokes: "large donations to traditional nonprofits or direct distributions to specific populations". Anthropic has already committed $200 million to the effect of AI on jobs, and says "bolder approaches may be needed" and that the fund "will have the versatility to pursue them". One limit is stated plainly: none of the fund's Anthropic equity will be used "for lobbying or political giving". The structure and governance, including the valuation milestones that add to the stake, are promised "in the coming months".

The claim this entry scores comes from chapters 7, 8 and 11. Proposals to share the gains of automation will keep surfacing, and most will fail the book's four tests: fund the stake from assets, not taxes; make it universal, not means-tested; protect it from raids; and build it before the gains are gone. The ledger scores each proposal by its shape when it surfaces. Since "Five per cent" in July it also records something the tests do not price: who is doing the offering.

Test by test, on paper.

Funded from assets: meets. This is equity, not a levy. It is the same instrument the ledger found in "Five per cent", when OpenAI was reported to be in talks about giving the US government a stake.

Universal: not stated. The examples are donations to nonprofits and payments to "specific populations". That is not an equal share for every citizen. A fund that can pay anyone can also be asked to pay only some.

Protected from raids: partly described. The ban on lobbying and political giving is a real limit on use. But the announcement does not say who decides how the money is spent, and no rule is described that would stop a future board from changing course. The 5 per cent itself is "subject to stockholder approval", so even the starting gift is waiting for a vote. And the part that grows with the company's success, the extra 2 per cent, is at "the board's discretion".

Built before the gains are gone: unclear. The stake is committed now and grows with the company, which is earlier than most proposals manage. But the baseline it grows from is not disclosed, and the recipient is a fund the company wholly owns, not a body with a claim of its own.

One point in the coverage needs care. TechCrunch's headline, "Anthropic says it's committing 5% of its shares to a 'public benefit fund'", follows Anthropic's own opening line. The condition sits one sentence later: stockholders still have to approve it. And the "20 per cent" is not one gift. About three quarters of it is earlier pledges by founders and employees.

Then the reaction. Elon Musk replied to Anthropic on X: "I recommend giving the shares directly to the public", and a few hours later, "This is what I will be doing". The next day he went further: "Next year, I will give one of my SpaceX shares to every kid in America. This will be the first of many." When a follower called it the beginning of universal basic income, he corrected it to "universal high income". No custodian, account or terms were described. Treat it as a pledge, not yet a programme. It is the criterion the book cares most about: shares in people's own names, not a fund that decides on their behalf. But a pledge is not yet a deed. Who holds a child's share, whether it can be sold and whether "many" will be equal are not yet said, and like Anthropic's extra 2 per cent, anything beyond the first share is at the giver's discretion. A deed needs a lock. Unless the claim is equal and cannot be sold, shares given "directly" can be gone by Monday.

So why complicates and not confirms? On the four tests alone, the fund clearly meets only one. But the claim expected proposals to be demanded of the owners. Twice this year an owner has moved first: OpenAI, reported in talks over a stake for a government fund, and now Anthropic, committing a stake to its own fund. An owner who gives is also an owner who sets the terms. The tests ask whether a stake will hold. They do not ask what the gift costs the giver, or what it buys.

What would settle it is readable in the next few months. If the governance names who decides, writes the distribution into a rule and pays the same to everyone, the book will have to admit that an owner can build what it said only politics could. If it keeps the discretion, the fund joins the long list of generous things that depend on the good mood of whoever holds the keys.

The book said the deed was the thing to ask for. Now an owner is offering a fund, with the terms still to be written. Who owns the machine?