Both of these men describe the kindly guardian better than almost anyone has, and that's the trouble with describing a job that well.

If neither name means much to you, here's the short version. Alexis de Tocqueville (1805– 1859) was a French aristocrat and magistrate. In 1831 he and his friend Gustave de Beaumont sailed for the United States, officially to study its prisons; their report came out in 1833 (On the Penitentiary System in the United States, translated by Francis Lieber). He also came back with Democracy in America (two volumes, 1835 and 1840), still the book politicians quote about America and, I suspect, haven't always read. Later he sat in France's Chamber of Deputies, wrote its reports on abolishing slavery in the colonies and on governing Algeria, and served briefly as foreign minister; his own Recollections list the 1849 cabinet as "Foreign Affairs Tocqueville" (Project Gutenberg). He went to study the prisons and came back with a theory of the entire country, which is a remarkable return on a site visit.

Dario Amodei is the co-founder and chief executive of Anthropic, an artificial intelligence company. He worked at OpenAI before that: in 2023 a senator asked him, under oath, "Let's talk about OpenAI, where you used to work. Right?", and he said "Yes" (S. Hrg. 118-108). Anthropic makes the Claude models. He writes long public essays about where the technology is going and what it might do to us: Machines of Loving Grace (October 2024),The Adolescence of Technology (January 2026), and, in September 2026, a short one, "Wemust pace the frontier". In October 2026 Anthropic filed its S-1, the document a company files to sell shares to the public.

Why put them side by side? This is the second essay in Captain to Captain, a series that pairs a thinker from history with someone building things now, and holds each to the other's standard. Vol. 1 was about a method: how you decide a discovery deserves to be believed. This one is about people: who gets trusted with power, how they're chosen, and whether the power ever gets handed back. Tocqueville wrote what I think is the best description we have of a power that is kind, competent and permanent. Amodei is building a powerful thing on purpose, with a stated commitment to care, and has written a great deal about how to keep it, and the people building it, in check. The question underneath matters whether or not you've heard of either man: when a power arrives that can't be stopped, who makes sure it leaves people more able to govern themselves, not less?

Two disclosures, and then I'll get out of the way.

The first is mine, and I wrote it in September before any of this research existed: "I think Dario Amodei is mostly right and asking mostly the right questions." Saying so is what buys me the room to be hard on him, and on Tocqueville, for the next twenty pages. When I am, it's the design I'm pressing on, not the man.

The second is the machinery's. This essay was researched and drafted with help from Claude Opus 5.5, a model made by Anthropic: the company that is half the subject of this essay. The conflict isn't hypothetical; Claude Opus 5.5's own release appears below as evidence. So the house rules for this one are stricter than usual. Every quotation here was checked by code, word for word, against a saved copy of its source. The critical review is led by models from two other companies, OpenAI and Google. The Anthropic model on the panel can argue but can't break a tie on any claim about Anthropic. None of that proves anything; it describes how I worked. Matching a quote to its source checks the words, not what I made of them, so the argument has to stand on its own links, and you're invited to check them. If a reader still thinks the machine went easy on its maker, that's a counterexample, and this series exists to be told about those.

Here's the belief this essay is built on. It's mine, I hold it firmly, and I hold it from lived experience, not from a study. That puts it in the faith bucket, and I'm telling you so up front:

When power is granted, power accumulates. It rarely stops and thinks if it should really do something, or if it really needs that power.

Mostly true, in my experience. A belief I can't lose is a belief I shouldn't publish, so here's how this one could lose. For each man I'll ask three separate questions, and grade them separately:

1. The diagnosis. Is what he says about power supported by the record?
2. The practice. Does what he, or his institution, does match what he says?
3. The door. Is there a correction that someone outside the power can start, and make binding?

My spine predicts something narrower about the third: without an outside correction that can bind it, granted power will usually keep the authority it could surrender. It may pause, refuse a demand, or act against a short-term interest; those matter, but they aren't the same as giving someone else the last word. My belief says "rarely", so one case won't kill it; a pattern would. It loses force if institutions, repeatedly and without being compelled, hand real authority to outsiders who can enforce limits the institution can't later undo alone. That's the only kind of case that counts against it. Restraint isn't surrender. Pauses and costly refusals are evidence about restraint and judgment, and I'll credit them, but separately: they answer whether careful people stay careful, not whether anyone else now holds the power to decide next time. And a power stopped from outside doesn't count against me; that just shows the door works. Whether a case was forced is my judgment, made in the open, case by case, and you can check it. A single clear case of unforced surrender counts as a dent, and I'll report it as one.

This series has a through line: expertise needs the right tools (the Moot, a panel of models built to argue with me), the right mental model (Bacon, in Vol. 1), and the right people and scaffolding, which is where these two take over. Vol. 1 ended on "The counterexample is a person. It should be allowed to be anyone." This one asks the next question. Who decides who's allowed?

Start with the description, because nobody has bettered it.

Near the end of Democracy in America, Tocqueville asks what kind of tyranny a democracy should fear. Not a Caesar, he decides. Something new: "it would be more extensive and more mild; it would degrade men without tormenting them" (vol. 2, book 4, ch. 6, Henry Reeve's translation, Wikisource). Then he draws it:

"Above this race of men stands an immense and tutelary power, which takes upon itself alone to secure their gratifications, and to watch over their fate. That power is absolute, minute, regular, provident, and mild."

Tutelary is the word to hold onto: the power of a guardian, a tutor, someone appointed to look after a person who can't yet look after themselves. Tocqueville says what would make it acceptable, and what makes it not:

"It would be like the authority of a parent, if, like that authority, its object was to prepare men for manhood; but it seeks on the contrary to keep them in perpetual childhood"

Reeve's "perpetual" is fair. Tocqueville's French leans a little harder: « il ne cherche, au contraire, qu'à les fixer irrévocablement dans l'enfance » (De la démocratie en Amérique, 1848 edition; the 1840 first edition reads the same), to fix them irrevocably in childhood. The guardian's kindness was never the danger. The missing exit is.

Then comes the line this whole essay hangs on. People in a democracy, he says, want both to be led and to stay free, so "they console themselves for being in tutelage by the reflection that they have chosen their own guardians" (Reeve; in the French, « Ils se consolent d'être en tutelle, en songeant qu'ils ont eux-mêmes choisi leurs tuteurs »). So elections don't fix it, though he's careful to say a constitution with an elected guardian "appears to me to be infinitely preferable" to one run by "an irresponsible person or body of persons": not enough, but not nothing. You can vote for the tutor and still be the child.

What did he think the cure was? Practice. Not wisdom handed down, but self-government done badly, often, in public, until people got good at it. "Town-meetings are to liberty what primary schools are to science; they bring it within the people's reach, they teach men how to use and how to enjoy it" (vol. 1, ch. 5, Reeve). "Political associations may therefore be considered as large free schools, where all the members of the community go to learn the general theory of association" (vol. 2, book 2, ch. 7, Reeve). And the warning, the other way around: "A people amongst which individuals should lose the power of achieving great things single-handed, without acquiring the means of producing them by united exertions, would soon relapse into barbarism" (vol. 2, book 2, ch. 5, Reeve).

Here's that, stated plainly enough to be wrong, which is the only kind of statement this series trusts. I'll do the same for everyone in the essay.

Tocqueville's hypothesis, plainly (my condensation, not his sentence): people learn to govern themselves only by governing themselves. A power that does it for them, however well, takes away the practice, and with it the ability. Nothing hurts, so people accept the trade, and console themselves that they chose the guardian.

What has to be true for that to hold? That self-government is a skill and not a trait, that skills fade without use, and that comfort makes the trade easy to accept. All three are testable. Remember the third one.

(One thing you won't find in this essay: "America is great because she is good." It's on plaques and in speeches, attributed to him. "Nowhere do they appear in Democracy in America, or anywhere else in Tocqueville," as Snopes puts it, after looking. It's this volume's "knowledge is power". We have the real one.)

Amodei uses a related metaphor and moves the clock forward. His January 2026 essay is called The Adolescence of Technology, and the frame is in the title: we are growing up, and the question is whether we grow up in time.

He is not naive about where that ends. "The triumph of liberal democracy and political stability is not guaranteed, perhaps not even likely, and will require great sacrifice and commitment on all of our parts," he wrote in 2024 (Machines of Loving Grace). His answer is strategic. He wants "democracies [to] build and wield the technology before autocracies do" (The Urgency of Interpretability, 2025; the bracketed word is mine, to fit his phrase into my sentence), and in 2026: "A coalition of the US and its democratic allies, if it achieved predominance in powerful AI, would be in a position to not only defend itself against autocracies, but contain them and limit their AI totalitarian abuses" (Adolescence).

He is also not naive about his own industry. In September 2026: "since roughly this summer, AI has been advancing drastically faster, driven primarily by AI's growing ability to build the next generation of AI. This dynamic is called recursive self-improvement, and it is starting to happen across the industry, including at Anthropic, as we and others have described. Left unchecked, it could outrun our ability to understand and control these systems, and so must be pursued very carefully, if at all" ("We must pace the frontier").

Including at Anthropic. If at all. I'd like that on the record, because it's to his credit. And the essay doesn't stop at the sentence. It proposes three steps: "Embedded Evaluators", third parties with "ongoing, employee-like access" who "verify adherence to safety practices and commitments, report incidents"; "Democratic Coordination" on "common safety standards as well as limits on the rate of unchecked AI progress"; and "Global Coordination", "to the extent this is possible". Of the first: "Anthropic is unilaterally committing to this step now." That's a mechanism, not a mood. It isn't a pause, and the page says so: "pacing does not mean halting model training or technical progress" ("We must pace the frontier"). No. 02 found that "A pause is almost never in the grant." This is the next thing down: a slowdown with someone outside watching. The question the rest of this essay keeps asking is narrower: can the evaluators stop anything, or only see it and say so?

And here's the needle he's trying to thread, from the same essay (written in Anthropic's "we", on his own site):

"Not building the technology deprives humanity of benefits or simply places AI in the hands of authoritarian powers, while building it too fast is reckless. We have sought a middle way: to show that it's possible to build carefully and succeed commercially, and to make safety something on which AI companies compete. In other words, to create a race to the top."

My own picture of that: he is trying to thread a microscopic needle with a rope the size of a ship's hawser, while running a race. Tough. Not impossible.

Amodei's hypothesis, as I reconstruct it: the powerful thing is coming either way. Better that careful people build it, inside democracies, under rules they help write, than that careless people or autocracies build it without any.

What has to be true for that to hold? Four things. That it really is coming either way; that careful people stay careful once they win; that the rules they help write bind them, not only their rivals; and that someone outside can tell, and do something, if they don't. The first is an argument about the world. The other three are about Anthropic, and Anthropic has a public record. So let's read it, and come back to all four at the end.

First, though, the metaphor deserves a hearing, because "adolescence" smuggles in a promise: that adulthood comes next, naturally, the way it does for people. Two writers have described what that adulthood might look like, and they don't agree.

Ray Kurzweil, inventor and author, follows the curve: "we'll witness on the order of twenty thousand years of progress — at today's rate of progress, that is" ("The Law of AcceleratingReturns", 2001), with "the date for the Singularity" set "as 2045" (Q&A, his site). A "year of progress" wouldn't survive a laboratory inspection, but the point does: change may outrun the institutions meant to govern it, and in Kurzweil's telling the adult is us, enhanced.

Cory Doctorow, novelist and digital-rights activist with the Electronic Frontier Foundation, has watched platforms grow up, and has a word for how it goes. "Here is how platforms die: first, they are good to their users; then they abuse their users to make things better for their business customers; finally, they abuse those business customers to claw back all the value for themselves. Then, they die. I call this enshittification" (Pluralistic, 2023). In Doctorow's adulthood the adolescent grows up into a landlord.

And Doctorow has the sentence the rest of this essay leans on. Platforms used to be better, he argued in his 2024 McLuhan lecture, and here's why:

"The pre-enshittification era wasn't a time of better leadership. The executives weren't better. They were constrained."

Constrained by four things, he says: "First: competition." "Second: regulation." "Third: self- help" (what users can undo themselves). "Fourth and finally: workers." (Pluralistic, 2024.) And then: "One by one, each of these constraints was eroded until it dissolved, leaving the enshittificatory impulse unchecked."

Doctorow's hypothesis, plainly: the quality of the people at the top matters much less than the strength of what can stop them. (My gloss, one step past him: good behavior is what constraint looks like from outside.)

That's close to my spine, said better by someone who has watched it happen longer; whether it holds for this company is what the record below is for. It's where my third question, the door, comes from.

I have a name for the thing that worries me, and it's borrowed. Watchmen (Alan Moore and Dave Gibbons, DC Comics, 1986–87) runs on an old Latin question, quis custodiet ipsos custodes: who watches the watchmen? Its own answer, as I read it, is a well-meaning man who decides for everyone and isn't stopped. That's the problem I mean: a guardian with good intentions and no one able to check it drifts, over time, toward control, and the good intentions are the alibi. The question itself is Juvenal's, from his sixth Satire, where the watchmen are guards hired to keep a wife faithful; the fuller version, in lines that survive in a single manuscript, says they now keep quiet « hac mercede », for this fee (Satire VI; my translation). Juvenal's watchmen, if those lines are his, were bought.

The question is structural, and it's the same for both men: does the design have a watcher with real power, who isn't chosen or controlled by the watched? Does the grant ever come back?

One objection deserves an answer first. Tocqueville wrote about governments; Anthropic is a company, and nobody votes on a drug maker's research agenda either. But notice what we built for drugs: an outside regulator with binding power over the decisions that carry the risk. Nobody votes on the lab bench; somebody outside can stop the drug. The guardian question applies whenever a private decision's risks land on people who never agreed to them and have little recourse, and Amodei's own case is that this one could settle "predominance in powerful AI" and "could outrun our ability to understand and control these systems" (Adolescence; "We must pace the frontier"). Anthropic seems to agree: it built a trust to "balance the interests of the public alongside the interests of Anthropic's stockholders" (Anthropic, 2023). It isn't unconstrained from outside: it's subject to California's SB 53, which it endorsed (Anthropic), and its filing warns of possible "government investigations or penalties" under export rules (S-1). So the question is the door, my third one, which for a company means two things: an independent authority that can stop a release, and a way for the people carrying the risk to challenge the decision through institutions answerable to them. The second is stricter than drug regulation asks; I think the stakes justify it, and that's my view, not a rule. A company can have testers, enforcement and capability and still fail both.

A voice, an answer and a veto are different things, so I'll grade them separately. Voice: an affected person can bring an objection and have it heard. Remedy: that person can put the objection before someone independent who has the authority, if the objection holds, to change what happens. Accountability: that decision-maker answers to the people carrying the risk. The first two make a door. The third makes it theirs. And a veto with no rules of its own is just another tutor at a higher desk.

Tocqueville first, because he's had longer to answer.

The grades, up front: the diagnosis, sharply stated, with its causal prediction still untested; the practice, at odds with it; the door, none the governed could open themselves. The evidence follows.

Tocqueville warned against the guardian. Yet in a prison study he approved one, and in a private working paper and two reports for the Chamber he helped design others, for other people. The record is in documents he wrote or co-wrote, and it's worse and more interesting than the usual summary. One distinction runs through it: two of these documents are reports he wrote as rapporteur for a commission, so they're the commission's recommendations in his words, not simply his private views. I'll say which is which.

The prisons, 1833. In their report, he and Beaumont approve of the American system, with eyes open: "it must be acknowleged [sic] that the penitentiary system in America is severe. Whilst society in the United States gives the example of the most extended liberty, the prisons of the same country offer the spectacle of the most complete despotism. The citizens subject to the law are protected by it; they only cease to be free when they become wicked" (Lieber translation, p. 47; checked against the page image). Their translator, Lieber, footnoted despotism to object that the system is "truly merciful if compared to the former systems of prisons." In the same book, in the authors' own foreword (« Il n'est pas une institution philanthropique, dont l'abus ne touche à l'usage »), is the sentence I'd put over the door of every benevolent institution, including the one I'm about to examine: "There is not one philanthropic institution, the abuse of which does not border closely on its usefulness" (p. xlvi).

Slavery, 1839. Tocqueville wrote the Chamber of Deputies' commission report on abolishing slavery in the French colonies, and his opposition to slavery is explicit: the government "cannot long resist the pressure of public sentiment upon this subject," and "after a very short delay, it will be compelled, directly or indirectly, to abolish slavery" (English translation, 1840). On principle it is plain: "Man has never had the right of possessing man, and the possession itself has always been and still is unlawful" (p. 21). Then the commission designs what comes after:

"The State should become the only guardian of the enfranchised population; it should grant, on its own conditions, the services of the blacks to the planters, retaining in its own hands all exercise of discipline."

(p. 48 of the 1840 translation, checked against the page image.) It's meant to be "an intermediate and transitory state between slavery and liberty", but read what the transition is for: "the most favorable moment to found that empire over the minds and the habits of the black population, and acquire that salutary influence, which will be necessary to direct them when they shall be entirely free" (p. 47). Until then, the government should act on the freed slave "by a wise and prudent legislation, which shall first introduce, and then, if necessary, compel the laborious and manly habits of liberty" (pp. 27–28). And read who pays. The planters have "strictly no right to an indemnity," the report says, unless freedom brings a refusal to work or wages above what forced labor cost them, but should get help anyway, and the same page 48 explains how: with "moderate wages for the work of these one hundred and sixty-six thousand laborers, the State can both cover the interest of the indemnification, and create a sinking fund to pay off the whole amount." The freed would pay for their owners' compensation, out of wages the guardian set.

The next year, 1840, he published his description of a guardian that "seeks… to keep them in perpetual childhood." In that report from a year before, the guardianship is declared temporary, and the influence is designed to outlast it. The grant ends; the reach stays.

And then history declined the design. When France abolished slavery in 1848, the decree said slavery "sera entièrement aboli dans toutes les colonies et possessions françaises, deux mois après la promulgation du présent décret dans chacune d'elles" (Décret du 27 avril1848): entirely abolished, two months after the decree was promulgated in each colony. No State as only guardian. (It wasn't hands-off: an account written that year says the decree came "accompagné d'une série de mesures ayant pour objet certaines précautions contre les désordres", with a series of precautionary measures against disorder (Galos, Revue desDeux Mondes, 1848).) But the state's interest in the freed people's labor didn't end with the decree. Another, signed the same day, punished vagrancy and begging in the newly free colonies: anyone found "sans aveu" could be placed "à la disposition du Gouvernement" for "trois à six mois" and put to public works "dans des ateliers de discipline", on the principle that "le travail est la première garantie de la morale et de l'ordre dans la liberté" (Martiniquepromulgation, 1848). Tocqueville's design lost, but the new law still gave the administration coercive power over some freed people's work. How often it used that power is a different question, and these documents don't answer it. The room was different in 1848, too. The commission that drafted abolition included François-Auguste Perrinon, an officer born in Martinique (Cochin, 1861) and a man of colour (Outremers360). So what I'm grading is a design, not a government. The guardianship was proposed and not granted. One thing the 1839 report had asked for did become law, though compensating owners was common practice by then (Britain had done it): Article 5, « L'Assemblée nationale réglera la quotité de l'indemnité qui devra être accordée aux colons », the Assembly would set the indemnity owed to the colonists. The planters' compensation made it into law. The freed people's guardian didn't, and neither did the plan to have their wages pay for it.

Algeria, 1841 and 1847. In a working paper written after his visit, his own and not a commission's, he takes on the critics of French tactics directly. In Jennifer Pitts's translation: "I have often heard men whom I respect, but with whom I do not agree, find it wrong that we burn harvests, that we empty silos, and finally that we seize unarmed men, women, and children", and "These, in my view, are unfortunate necessities, but ones to which any people who want to wage war on the Arabs are obliged to submit" (Pitts, 1999; the French is « des nécessités fâcheuses », Travail sur l'Algérie, 1841). Pitts calls this essay Tocqueville "at his most ruthless." And: "le droit de la guerre nous autorise à ravager le pays", the law of war authorizes us to lay the country waste. On the same page, to be fair to him, he argues France must not reproduce "ce qui en eux méritait la détestation du monde," what in the Turks had earned the world's loathing. (Pitts's translation where cited; the other renderings are mine, and her edition, Writings on Empire and Slavery (2001), is the standard English text.)

By 1847, writing the commission's report to the Chamber, he describes the thing from the other side, as a warning. Civilized peoples, he writes, often oppress others "par leur seul contact, sans le vouloir": by mere contact, without meaning to. The rules that look to a European like guarantees of liberty look to the governed like intolerable oppression, and then the line that closes the circle:

"The forms we call tutelary, he calls tyrannical" (Rapport sur l'Algérie, 1847, extracts). Tutélaires. The same word as the immense and tutelary power.

And this report does more than warn. It proposes a real restraint. Wherever land could change hands between "le propriétaire barbare et l'Européen civilisé", the barbarian owner and the civilized European (his categories, not mine), without control, it says, the land passed "rapidement, et à vil prix", quickly and dirt cheap, into European hands, until the native population "cessait d'avoir ses racines dans le sol." So: « il faut que nulle part les transactions de cette espèce ne soient entièrement libres », nowhere may such sales be entirely free. He warns against turning Algeria into "un champ clos, une arène murée, où les deux peuples devraient combattre sans merci, et où l'un des deux devrait mourir," a walled arena where one people must die: "Ne recommençons pas, en plein XIXe siècle, l'histoire de la conquête de l'Amérique." Let us not begin the conquest of America over again. I take him to mean it. And the report's own subject heading is "domination et gouvernement des indigènes," domination and government of the natives. The domination is the premise. The restraint is on how it's done.

The same report records the commission trying something bolder. It told the government that a majority of its members found a planned expedition into Kabylia "impolitique, dangereuse"; the Minister of War replied that the question fell "exclusivement dans les attributions de la prérogative royale" and "refusait de recevoir la communication", refused even to receive it (Œuvres complètes, 1866). A deliberative body tried to interrupt the government, and the argument became whether it was allowed to.

And the commission wrote a second report, dated 2 June 1847, on a government bill for "camps agricoles", army-run farming settlements (same edition). It opens by taking as proven that settling Europeans in Africa "serait le moyen le plus efficace d'y asseoir et d'y garantir notre domination", the surest way to secure French domination there. Then it records a dissent. "Deux membres seulement", only two members, doubted it was humane and wise. The country, they said, "est déjà occupé, possédé et cultivé"; bringing in a new population meant "y éterniser la guerre et y préparer la destruction inévitable des races indigènes", making the war endless and preparing the inevitable destruction of the native peoples. The report's answer: "Le pays est occupé, il est vrai, mais il n'est ni rempli, ni même, à vrai dire, possédé", occupied but not, truly speaking, possessed; and land held only in common is easier to settle than land where every inch is defended by a private right. The commission did reject the bill, unanimously, but "dans la forme que le gouvernement lui avait donnée": on how settlement would be organized and paid for, not on whether it should happen. These are the commission's conclusions in his words, and the report doesn't say which lines were his alone. Here the objection got into the room, and into the record. The commission kept the hinges.

Tocqueville, graded.

1. The diagnosis: clearly stated and well documented. His description of the tutelary power is, in my view, the best we have, and in 1847 he warned that forms Europeans call protective can look tyrannical to the governed, and that France must take care it didn't repeat North America in Algeria. Whether civic skill really decays the way he predicted is an empirical question this essay can't settle; Anthropic's own survey, the Public Record (below), suggests people fear something like it, which isn't the same as showing it.
2. The practice: doesn't match the principle. He warned, in his own words, that his side's guardianship could oppress "par leur seul contact, sans le vouloir", and his recommendations didn't resolve it. In 1839 the commission he wrote for designed a guardian for freed people; in 1841 he called harvest-burning an unfortunate necessity; in 1847 the commission's reports, in his words, proposed real limits, on land sales, inside a domination it didn't question, and answered the two members who warned of "la destruction inévitable des races indigènes" by redefining possession. Noticing and proceeding is harder to excuse than not noticing.
3. The door: a door in time, not a voice. The 1839 guardianship was meant to end, "transitory", but the freed people weren't in the room where its terms were designed, and neither were the Algerians of 1847; the nearest thing to their voice in the record is two unnamed commissioners, outvoted. The restraints were the commission's to propose, in his words. In the three terms: no voice for the governed, no remedy they could start, no accountability to them.

Against my spine: the 1839 design is no evidence either way, because it was never granted. The Algerian record fits it: the power he served didn't pause, and the restraint he offered came from inside it. What would have broken the spine here: a door for the governed, written into his own proposals. I looked for one in these documents and didn't find it. That's a finding about these proposals and the people they governed, not about his political thought, which is full of checks for people he counted as citizens.

The grades, up front: the diagnosis, clearly stated, his own company included, with the big forecasts still open; the practice, matching it more often than not, with one consequential rewrite of the company's own safety rules; the door, no binding release remedy I can find that an outsider can start. Against my spine, one qualified dent. The evidence follows.

Same three questions, same standard: what he says, against what the record shows Anthropic does, verified row by row.

Regulation. Under oath, in 2023: "However, private action is not enough. This risk, and many others like it, requires a systemic policy response" (S. Hrg. 118-108). In 2026 he went further than many CEOs go: "I would support civil liberties-focused legislation (or maybe even a constitutional amendment) that imposes stronger guardrails against AI-powered abuses" (Adolescence). And what Anthropic did: in September 2025 it endorsed California's SB 53, a state AI-safety law, while saying "frontier AI safety is best addressed at the federal level instead of a patchwork of state regulations, powerful AI advancements won't wait for consensus in Washington" (Anthropic), and then published a framework for complying with it (Anthropic). Says and does: match. This is Doctorow's second constraint, and Amodei asked for it to be pointed at himself.

Outside testing. In 2023 (S. Hrg. 118-108, prepared statement): "New AI models should have to pass a rigorous battery of safety tests before they can be released to the public at all, including tests by third parties and national security experts in Government." Anthropic's own view is that self-testing is "insufficient as it relies on self-governance decisions made by single, private sector actors" (Anthropic). What it did: Claude Opus 5.5 "was tested before release by external evaluators, including Frontier Design and METR" (Anthropic). Third-party testing: match. The system card also says Anthropic "collaborated with the U.S. Center for AI Standards and Innovation (CAISI)" at NIST "on measurement of cyber and biological capabilities and safeguards" (system card), but not whether release waited on it. Government testing: partial match. What none of these documents says is whether an outside tester can stop a release. An outside tester can see some of the risk-relevant work outsiders normally can't, and say when it looks wrong. I haven't found a public right to carry that objection to someone who can stop the release.

Pacing. On 12 September 2026 (the page carries no date; the date is from contemporarycoverage), in "We must pace the frontier": "We must slow the pace at which we improve the capabilities of AI models. Progress will still seem fast, and we must make wise use of the time we gain." So what does that promise? "Slow the pace at which we improve the capabilities" is about the rate of improvement, not the number of launches, and it's a "we must" aimed at the industry; the step Anthropic took on itself is the embedded evaluators: "Anthropic is unilaterally committing to this step now." So launches aren't the test. On 22 September Anthropic released Claude Opus 5.5, which it called, on its release page, "our first release since we called for pacing the frontier," and on 7 October Claude Haiku 5.5, "the cheapest, fastest, and most capable small model we've ever released" (Anthropic). Timing cuts both ways, too: "Claude Opus 5.5's knowledge cutoff date is June 2026", so its training data predates the essay; when its training ended, the card doesn't say.

So there are two things to watch, kept apart. The first is an observation: capability against trend. Anthropic published that arithmetic itself, in the Opus 5.5 system card: a trend through "the eight releases from Claude Opus 3 to Claude Opus 4.6," rising "14.75 AECI points per year" on its own index, best explained by one jump at an earlier model, Claude Mythos Preview. Opus 5.5 "scores about 6.5 points above the historical trend line and within a point of the shifted trend line that starts at Claude Mythos Preview"; in the card's words, "we do not read this as a further slope change." Haiku 5.5's card puts it "much lower" (167.11 against Opus 5.5's 174.56; Haiku 5.5 system card), so it isn't a frontier step. Two cautions: it's Anthropic's own index, and it moves when refit; Opus 5.5 scored 169.36 in its own card and 174.56 in Haiku's. Says and does: a slowdown is not yet shown, and neither is a further speed-up. One more data point arrived the day this was revised. Anthropic's amended S-1, filed 2 November, tells future shareholders that "our internal estimates and our expectations regarding future models, rather than our current models, indicate that we may imminently meet our Responsible Scaling Policy's definition of 'dramatically accelerated progress attributable to AI'", that "We have not designated any of our current models as exceeding this threshold," and that it is "considering updates to our Responsible Scaling Policy to keep humans in control of frontier model development at Anthropic" (S-1/A). The brake question may arrive soon. Who gets to write those updates is the question this essay keeps asking.

The second is the test, because it's the commitment Anthropic actually made: the embedded evaluators. In an update dated 15 October, Anthropic says it has "onboarded six METR employees", that "Anthropic can't veto publication, but we can require them to redact information if it falls into predefined categories", and that "Anthropic ultimately decides what information does and does not fall into these categories, although METR reserves the right to make certain statements about the level of access and information it received" (Anthropic). So, stated before the fact: does METR publish; does it say its access was adequate; and does anything it finds visibly change what Anthropic ships? The dates matter here. METR's 22 September report on Opus 5.5 was a separate, earlier pre-release assessment, not the work of the embedded arrangement described in October. It shows what outside evaluation can and can't establish; it doesn't yet show what the newer arrangement produces. METR's own summary of its Opus 5.5 work is careful in both directions: "the data we have is insufficient for distinguishing consistent, accelerating, or decelerating rates of improvement"; the work "was not meant to verify claims about compliance with any specific threshold from Anthropic's policies"; and "Anthropic had the opportunity to review and edit the text. We signed off on this final text" (METR). It also passes on, from a separate METR team with more access, a preliminary estimate of "~1.5X overall acceleration in capabilities due to AI (i.e. 1.5 years in 1 year)," and warns that because the period wasn't specified, "it is unclear whether this estimate applies to the development of Claude Opus 5.5 or another period." That's a different measurement from the chart's slope, and neither one says whether anyone outside can stop a release. The editing note is a disclosure, not a sign of capture, and it belongs next to the chart. Credit where it's due: a company that publishes the chart its critics need is doing something many don't.

Rewriting the rules. There's a harder test than writing a safety rule: watching the writer change one. On 24 February 2026 Anthropic published version 3.0 of its Responsible Scaling Policy and said why. "We found pre-set capability levels to be far more ambiguous than we anticipated," and "the robust mitigations we laid out in the prior RSP might prove outright impossible to implement without collective action." So it separated "the mitigations that we plan to pursue regardless of what others do" from what it recommends for the whole industry, added "external review of Risk Reports in certain circumstances," and put its new roadmap in a different category: "Rather than being hard commitments, these are public goals that we will openly grade our progress towards" (Anthropic). An outside assessor, SaferAI, compared the current policy with version 2 and found that Anthropic "Reduced involvement of external parties in evaluating risks and mitigations" and "Removed or significantly weakened" its "commitments to halt development"; the same assessment notes the policy still states a "willingness to pause development outside formally specified scenarios" (SaferAI; an analyst's reading, not a ruling).

Two readings start from the same fact. The kind one: an engineer revised a specification that had stopped describing reality, in public, with reasons, and a rule nobody can meet isn't safety. The unkind one: a company loosened a restraint it had set on itself. I don't have to pick one to see what it shows for this essay. The party that wrote the limit could rewrite it, alone. Says and does: published openly, with reasons; it binds the writer until the writer revises it.

Restraint under pressure. Three 2026 episodes test the practice harder than any essay can, and they point different ways.

The first is a refusal. On 26 February, Amodei published a statement on two limits Anthropic would not drop from its work with the Department of War: no "mass domestic surveillance", which he called "incompatible with democratic values", and no "fully autonomous weapons", because "today, frontier AI systems are simply not reliable enough to power fully autonomous weapons." The Department, he wrote, had threatened "to designate us a 'supply chain risk'—a label reserved for US adversaries, never before applied to an American company— and to invoke the Defense Production Act to force the safeguards' removal." His answer: "these threats do not change our position: we cannot in good conscience accede to their request" (Anthropic). It cost something. The designation letter came on 4 March (Anthropic). Anthropic sued, twice, because the government had used two different statutes. In August a federal district judge in California "vacated and permanently enjoined enforcement of the Section 3252 designation," in the Congressional Research Service's summary (CRS), and found the actions were "unlawful retaliation against Anthropic for constitutionally protected expressive activities" (order of final relief). In September the D.C. Circuit, which reviews the other statute, denied Anthropic's challenge, "upholding DOD's supply chain risk designation under Section 4713. As a result, DOD may continue to enforce its directive to remove all Anthropic products from DOD systems, barring additional court action" (CRS). One court didn't overrule the other. The D.C. Circuit said it had "no quarrel" with the California ruling on section 3252, and that "no such bad motive is required to support a designation under the much broader definition set forth in section 4713" (D.C. Cir., 25 September 2026). In October Anthropic asked the full court to rehear the case (Lawfare).

The Department had a case, and it belongs here. In the D.C. Circuit majority's words: "The Secretary raises the deeply sobering prospect of overly constrained AI models shutting down unexpectedly and thus causing important military operations to fail," after "a dispute arose over whether the contractual prohibitions barred the use of Claude in an ongoing overseas military operation." The Department's own memo, quoted in the California order, put it more sharply: Anthropic "seeks to grant itself an operational veto." That's this essay's question asked from the other side: who gets to stop whom. I can credit the refusal without pretending the government had no argument, and without needing either side to have won every point.

The second is a proposal. Anthropic's June 2026 Advanced AI Framework says "there should also be a way to block or deter deployment of models that pose significant catastrophic risks", and "Self-assessment is not enough" (Anthropic). That's a company asking for the outside lock this essay keeps looking for. It's a proposal, not a law.

The third is a lock that turned. On 12 June, per CRS, the Commerce Department "reportedly imposed export controls" on two Anthropic models; "Anthropic disabled access to both models for all customers"; the controls were lifted on 30 June. A government stopped access, for its own reasons.

By my own rule these count differently. The interruption proves an outside authority can stop access. It doesn't prove a person carrying the risk could start the challenge or demand a remedy, so it isn't the public's door, and it doesn't touch the spine. The framework is advocacy for such a door: credit, not surrender. The refusal is the one that matters: a self- imposed limit, kept under threat, with documented consequences (a designation, two lawsuits, removal from the Department's systems) and no outside door making it hold. That's costly restraint, and it's real. It isn't surrender: Anthropic kept the power to set its own policy and used it to say no. I'll grade it where it belongs, below.

Who governs the company. This is the row Tocqueville would have read first. Anthropic's structure is unusually explicit, and it publishes it, which is to its credit. In short, once the offering completes: the founders' company holds "50.1% of outstanding voting power" on ordinary matters; the Trust elects the Class T directors, "a majority of our board of directors"; and employees and public shareholders each hold a vote on some changes to the Trust. (The full table of who votes on what, in the S-1's own words, is in the appendix at the end.)

The S-1 says what that adds up to, more plainly than I would have dared. The Trust and the founders' company are expected "collectively" to "elect all members of our board of directors," and the founders' company to "determine the outcome of all matters submitted to a vote of our stockholders generally for the foreseeable future." Public buyers are told they "generally will not be able to affect the outcome" of those votes "regardless of how they vote" (unless the founders can't agree among themselves). The filing doesn't hide the arrangement; it's candid about whose decision it will remain. Who owns the founders' company? "Dario Amodei, our Chief Executive Officer and a member of our board of directors, and Daniela Amodei, our President and Chairperson of our board of directors, together with our five other co-founders, will initially be the sole members of the Founder LLC." And who picks the trustees? "New Trustees are selected by existing Trustees, in consultation with Anthropic, and have no financial stake in Anthropic" (Anthropic, 2026).

It isn't forever. The founders' share ends after "there have been fewer than three Continuing Founders… for 90 consecutive calendar days," which starts "a 450- day transition period," after which "all of our Class F common stock and Class E common stock then issued and outstanding will be redeemed by us for no consideration" (S-1). Note the second class in that sentence: the employees' vote ends with the founders' control. So the one internal key I credit below is coupled to the thing I'm worried about, and goes when it does.

So, once the offering completes, the guardians choose the guardians. Reeve's Tocqueville: "they console themselves for being in tutelage by the reflection that they have chosen their own guardians." Prospective public shareholders are being told before they buy. The rest of us, who may carry the wider consequences, haven't been given a comparable choice. There's nothing unusual about an IPO not seeking broad public consent, and the standard I'm holding this one to is stricter than the law's, because of what it says it's building. (That last part is belief, not record.)

None of this is secret, and founder control on its own is ordinary; the Trust is the unusual part, and it's exactly the part I'm crediting. It's an unusual check, with an override, and the S-1 spells the override out. The founders can't change the Trust alone. With the Trustees' agreement, a change takes supermajorities of both the board and the Trustees; without it, while the founders' share lasts, it takes the founders, the employees' class and the public shareholders' class at once; after the founders' share ends, a large supermajority of public shareholders can do it without the Trustees. (The thresholds, in the filing's words, are in the appendix.) Anthropic's 2023 announcement called these "failsafe" provisions, because the Trust is an experiment and experiments need a way to be fixed. Anthropic's noncompliance policy also "introduces a pathway for employees to make informal inquiries about potential RSP violations" (Anthropic). Against No. 02's four needs for a safe handover of power: provenance and audit, it scores. Employees have a limited but real governance key over specified changes to the Trust, not an independent release veto. For a customer or citizen wholly outside, I haven't found the voice-and-remedy combination I mean by a door. Memory of failure remains ungraded; the mechanism hasn't been tested that way.

Who pays the watchers? Anthropic does, and the S-1 says so: "We provide funding to the LTBT, which holds all of our Class T common stock, to support its operations and the compensation of its Trustees": "$2,716,750, $0, and $906,089" in 2023 to 2025, and

"$8,425,213" in the first nine months of 2026, with "annual cash compensation of $400,000" per trustee (the chair gets $25,000 more). It also says "Trustee compensation will not vary with or be conditioned upon the profits or any action or inaction of our company," and trustees may not hold Anthropic stock (S-1). Juvenal's watchmen were paid to keep quiet; these are paid to watch, which isn't the same thing. Independence turns less on who signs the check than on who can appoint, remove or overrule the watcher. On paper the Trustees pick their own successors; in practice, the S-1 says, "the Trustees and company executives have collaborated closely on sourcing, interviewing, and selecting Trustees," and the Trust is "accounted for as a consolidated variable interest entity" on Anthropic's books (S-1). Neither fact alone defeats its independence; both belong next to it. So does today's board, which the filing lays out: of "six directors", three "were elected as the designees nominated by the LTBT."

The amended filing discloses one more relationship that belongs here. In its words, "in August 2026, we made a $450,000 charitable donation to CNAS, of which Mr. Fontaine is the Chief Executive Officer" (S-1/A); Richard Fontaine is one of the Trustees, and the filing says the money supports CNAS research on "how AI is reshaping global competition, national security, international security, democratic governance" and related questions. It's disclosed openly. The filing doesn't say it influenced anything, and I have no evidence it did. Still, when the question is who funds the guardians, the institution one of them runs belongs on the page too.

One of those three Trust seats is held by Daniela Amodei, the company's president, a co- founder, and one of the seven members of the founders' company. Today, a seat the outside check fills is filled by an inside founder.

To be fair, the filing says that ends with the offering: the Trust's four seats go to "Mr. Hastings, Dr. Narasimhan, Mr. Liddell, and one vacancy" (Dr. Narasimhan is Vasant Narasimhan, a director "since April 2026", in the filing's biographies), and she moves to a seat the founders' and public shareholders' classes elect (S-1). And a trustee "may also be removed by order of the Court of Chancery of the State of Delaware upon petition by a person designated by the Trust Agreement", for "disqualifying conduct": an outside check on the watchers, though only someone the agreement names can ask for it.

Public input. This row started well and the record is richer than it first looked, so it needs four words kept apart: consultation (asking people), influence (what they said changes what's built), implementation (documented in a product) and authority (they can make it happen, or stop it).

In 2023, "Anthropic and the Collective Intelligence Project recently ran a public input process involving ~1,000 Americans to draft a constitution for an AI system" (Anthropic). A constitution, here, is the written set of principles a model is trained to follow. They published the method, the result, and the judgment calls, including what got cut: statements left out "due to either low overall agreement or a lack of consensus across opinion groups." That's Tocqueville's free school, run by an AI company, and I mean that as a compliment. And it reached a product. Anthropic's Claude 3 model card: "In response to our work on Collective Constitutional AI 17, we added an additional principle informed by our public input process, which instructs Claude to be understanding of and accessible to individuals with disabilities, resulting in lower model stereotype bias" (Claude 3 modelcard). That's implementation, documented: one principle.

What came after: in January 2026 Anthropic published a new constitution, and says who wrote it. "Amanda Askell is the primary author and wrote the majority of the text," and

"Many people at Anthropic and beyond contributed to the creation of this document, as did several Claude models" (Anthropic). In May 2026 it described "dialogues with groups whose work and traditions bear on the questions raised by AI", starting with "scholars, clergy, philosophers, and ethicists from more than 15 religious and cross-cultural groups" (Anthropic). And in June 2026 it published the first wave of a survey series, the Anthropic Public Record: "a nationally representative online survey in November and December of 2025 of 51,993 Americans, sourced from YouGov and weighted to US Census benchmarks" (Anthropic). Two of its findings belong in this essay. "The second most prominent fear was cognitive dependency (56%)", meaning AI leaving people "unable to think for themselves". That's a cousin of Tocqueville's fear, not the same one; his was about governing together. And Anthropic's own reading is that it's "a mostly anticipatory fear": of those worried, "only roughly 1/5 would feel significant disruption if AI became unavailable." Fear isn't evidence of lost agency. The second finding: "Only 15% of Americans said they trust AI companies to make decisions about how AI is developed and used." A survey is consultation, not consent, and Anthropic published a result that cuts against its own position, which is to its credit.

Now take that too far, on purpose. Taiwan's Join platform (more on it at the end) makes a minister answer once 5,000 people countersign a petition. Run Anthropic's 51,993 respondents through Taiwan's rule and they could force about ten answers; the 15% who trust AI companies, about 7,800 people, rate one and a half. That's an imaginary crowd at the survey's weighted percentages, not a headcount, and it breaks exactly where it should: a survey isn't a petition, an answer isn't a veto, and neither gives anyone control over a model release. That gap is the whole essay in four digits. So: consultation, yes, ongoing. Influence, documented once. Implementation, one principle. Authority, none that I can find: the public, and the clergy, are asked; the company decides. Grade: real and continuing consultation, without a door.

Amodei, graded. Back to the four things his hypothesis needs:

1. It's coming either way. An argument about the world, not about Anthropic, and I'm not settling it here.
2. Careful people stay careful. Barely tested, and no longer untested. The February refusal is the first observation, and it's positive: a limit kept under threat, at some real cost, though what it cost commercially nobody outside can see. The pacing chart is the second, and its verdict is "not yet shown."
3. The rules bind the writer. Partly, and February 2026 showed where "partly" ends. The scaling policy and the evaluator commitment are written by Anthropic, for Anthropic, enforced, so far, by Anthropic, and, in version 3.0, rewritten by Anthropic, which kept some commitments and moved others into a roadmap of goals rather than "hard commitments". The one exception is one it asked for: under SB 53, in Anthropic's own summary, developers must "Be publicly accountable for the commitments made in their frameworks or face monetary penalties."
4. Someone outside can tell, and act. They can tell, more than I expected going in: the system cards, the published constitution, the testers in the building. Acting is thinner. As far as I can find, there's no publicly enforceable release veto that a citizen can start. The Trust can't run the company ("neither the LTBT nor the Trustees have decision- making authority in the day-to-day management or strategy setting of the company"), but after the offering the directors it elects will be a board majority, and a board can stop a release; whether they ever would, against management, the record doesn't show. The shareholders mostly can't; the public is consulted; the employees will have a vote on some changes to the Trust, when the Trustees don't agree and while the founders' share lasts. In the three terms: voice, yes, by consultation; remedy, partly: independent constraints exist through law and board governance, but I haven't identified a public route to invoke a binding release remedy; accountability to the people carrying the risk, not yet. That's the finding.

In the three questions: the diagnosis, clearly stated and documented, including the uncomfortable part, "including at Anthropic"; whether its strongest predictions come true is open. The practice, matches the diagnosis more often than not: it points its constraints at itself, held a limit under threat in February, with documented consequences, and in the same month rewrote its own scaling policy, openly and with reasons, replacing some hard commitments with publicly graded goals while keeping others. The door, the one that's missing, for anyone wholly outside. Outsiders can see more than I expected: independent evaluators can report, and one public rule, SB 53, attaches penalties to a company's own commitments. I haven't found a general right for an affected person to put a release objection before an independent decision-maker who can require a remedy. That's the missing door: they can see, and they can't make it stop.

Against my spine: one dent, and I owe it to him to say so. It's the Trust: a power the founders gave away unforced, a commitment that after the offering a board majority will be elected by a body they can't override alone, even as the same filing keeps them 50.1% of the ordinary vote. What is it a dent in? Against the founders' power to choose the board, the Trust is a qualified surrender. Against the public's lack of binding recourse, it isn't yet a door: the founders gave up one kind of last word, and the rest of us still don't have one. It has its limits: the company funds the Trust, its executives helped pick the trustees, and today one of the Trust's seats is a founder's. An earlier draft of this essay counted the February refusal as a second dent. It isn't one. It's a different kind of evidence, costly restraint rather than surrender, and it belongs to his hypothesis, not against mine: the first good mark on condition 2. So the record shows two marks of different kinds, a dent in my spine and a pass on his. One dent isn't a pattern. A second and a third would start to make one; I'd be glad to read about them, and I'd owe this essay's first belief a rewrite.

The conclusion should rest on a real case, not a story. So: has anyone actually used these tools to make people better at governing themselves, and measured it?

There are real deployments; whether they make people better at it is, as far as I found, mostly unmeasured. Polis is "a platform that leverages machine intelligence to scale up deliberative processes," used, according to its developers and colleagues (three of the paper's nine authors were at Anthropic, and its pilots used Claude), "in thousands of conversations tied to policy making processes across dozens of countries and five continents" (Small et al., 2023). In Taiwan, a civic process called vTaiwan uses it, among other tools. vTaiwan "brings citizens and government together in online and offline spaces to deliberate and reach rough consensus on national issues," and one design rule is the whole point: "Participation by staffers at agencies with the domain to change policy based on the results of the consultation is mandatory" (Computational Democracy Project). That doesn't let participants compel a decision; the government still decides. It does put officials who can act in the discussion, which is not nothing: the counterexample can reach them. The binding remedy, the door I'm looking for, is one step further on.

It doesn't always work, and some public reports say so. Austria tried Polis with its 2022 national climate assembly, and the people who ran it wrote that "in many ways the Pol.is public engagement did not work out as we had anticipated", while calling it "a wonderful precedent" (KNOCA). And the machines help less than you'd think: even successful uses "require dozens of hours of (human) analysis per conversation" (Small et al.). Take that too far: if "dozens" means two dozen and "thousands" means two thousand, that's 48,000 hours of analysis, about 24 working years for one analyst. (The numbers are my assumption; the paper gives neither.) So the machine doesn't remove the work, it rations it. The software can widen the conversation; it can't quietly inherit the judgment of the people who have to make sense of it. If that work falls to a few analysts rather than the citizens, that's the failure mode to watch: the machine can lighten the analysts' work, and there is little evidence yet that it makes the citizens better at theirs.

The mechanism I'd test: people could become more capable individually while getting less practice acting together. An assistant good enough to do what once took an association is a liberation. It's also the removal of the occasion on which I'd have found a common interest, argued out a difference, or learned whom I could rely on. vTaiwan answers it one way: use the machine to find where people agree, and then make the people with power show up. Its record is mixed; Audrey Tang said in a 2024 interview "vTaiwan is now moving into a much more lab-like, research-like role" (Ministry of Digital Affairs). A system that collects everyone's preferences separately and hands back a settlement answers it the other way, skipping the practice, however good the settlement. I have a mechanism, not a measurement: nothing here shows the skill itself fading. That's the part someone should test, over months, in groups that decide real things with the machine and without it.

(How fast new things arrive, against how much schooling Americans have and how much they value it, is in a short companion piece, Speed and Schooling, with the chart and the data, to follow separately. It's context, not evidence, so it lives outside the argument.)

(Verse. The words are mine, not theirs.)

Sailed to count the cells and came home with the nation, / Drew the softest cage in the whole administration. / Called the tutor's sin a childhood with no key, / Then wrote the freed a tutor, and the land not wholly free.

Writes "if at all" in a post and means it, I believe, / Puts the testers in the building, lets them watch but not reprieve. / Seven founders hold the share, and the trustees choose their heirs, / And the public gets consulted, very kindly, on the stairs.

[What follows puts each man's yardstick against the other's record. The words are mine; the yardsticks are theirs. It's inference, labeled as such.]

Start with what they share, because neither man got to choose the world his argument starts in. Tocqueville treated the movement toward equality as "a providential fact" that "constantly eludes all human interference" (Introduction). Amodei treats the race to build powerful AI as a condition no careful builder can safely ignore: not building means leaving it to "authoritarian powers." Those are different claims about inevitability. Both force the next question: who governs the power once it's here? And Amodei also worries about the thing itself: AI systems acting on their own. Both, to their credit, also put the danger inside their own house: Tocqueville in democracy itself, including the one he expected France to become, Amodei "including at Anthropic."

Tocqueville's standard, applied to Anthropic. Does the guardian prepare people for adulthood, or fix them in childhood? The test isn't kindness; Tocqueville's guardian is kind. The test is the exit. Anthropic has built more constraints and openings than I expected: a public "if at all"; testers in the building; a Trust the founders can't simply overrule; a published constitution; one public principle carried into a shipped model. They deserve separate credit. None, on the record I've found, yet gives an affected outsider the independent, binding remedy this essay calls a door. What it hasn't built is the town meeting: a place where people outside the founders, the trustees and the staff can bring the counterexample and make it binding. It ran the closest thing to one in 2023, and kept a line of it. By Tocqueville's standard, that's credit, and it's also the same gap I graded as his failure: consulted is not admitted.

Amodei's practice, applied to Tocqueville. Amodei has a habit of naming his own company inside the danger, and then publishing the evidence a critic would need. Tocqueville did both. He saw the danger, wrote it into his own report, « les formes que nous appelons tutélaires, il les nomme tyranniques », published it to the Chamber, and even proposed limits. What he didn't do was let the people inside it into the room. By Amodei's standard, Tocqueville passes on candor, and only on candor: he also argued for the conquest itself, and for harsh means to win it. His failure wasn't blindness, and it wasn't secrecy. It was seeing clearly and keeping the pen, and on that one test, not the others, Anthropic hasn't passed yet either.

The payoff isn't a compromise, and it isn't tidy. Tocqueville brings the yardstick: does the power make people more able to govern themselves, measurably, or less? Amodei brings the builder's capacity, and a habit of writing the brakes into the blueprint and publishing the dashboard. Doctorow brings the warning that neither good men nor good blueprints last; rules a builder can rewrite fade, and constraints someone else enforces hold. And

Juvenal brings the oldest footnote: the watchmen can be bought. What none of them settles is the door. My answer, and it's mine: a guardian is safe only to the extent that the governed can make it stop. Not ask. Make. Not every person on every question, since emergencies and minority rights sometimes need guardians a majority can't simply overrule, but through people the governed choose, with recourse they can enforce. Tocqueville's associations came close, for a while. Taiwan's Join platform is a smaller thing: "5,000 people joined in electronic countersignature can interpellate, basically demanding an answer out of a minister" (Audrey Tang, Ministry of Digital Affairs). An answer, not a decision, and that's the honest distance still to cover.

A company could help establish one: a public body with real authority over one consequential decision, whose members the company can't appoint, remove, overrule, or financially punish for saying no. Its funding would need protection from the institution it can stop. It would need rules of its own: who has standing to bring a challenge (the people carrying the risk, not whoever happens to be in the room), what they have to show, a written decision, a time limit, an appeal for the company, and someone watching it in turn. The guardian's guardian needs governing too. That's the experiment I'd like to read about in the next essay on darioamodei.com.

When power is granted, power accumulates. Each of these men saw it, in his own way. Only one of them is still in a position to do something about it, and he's busy.

The decision rights after the offering, in the S-1's own words (preliminary filing of 26 October 2026; the voting, board-selection, Trust-amendment, funding and trustee-removal provisions quoted in this essay are unchanged in Amendment No. 1 of 2 November 2026):

Changing the Trust, in the filing's words. With the Trustees' agreement: "80% of the authorized members of our board of directors and 80% of the Trustees then in office". Without it, while Class F remains outstanding, all three other classes at once: the founders' share ("acting pursuant to the unanimous approval of the Continuing Founders" for the first three years, "66 2/3% of the Continuing Founders" at any time); the employees' Class E ("a majority of the votes cast", with a 40% quorum, on the first route; "75% of the votes cast" on the second); and "a majority of the votes cast by our Class A common stock". After Class F ends: "upon approval of 85% of our outstanding shares of Class A common stock", with no trustee consent needed.

All accessed 1–2 November 2026. Every quotation was checked by code against a saved, hashed copy of the source; the copies and hashes are kept with the essay's working files.