Before ChatGPT, there was Google.

I was in my mid-forties, a graduate student in bioinformatics at Michigan, and I had been reading about metformin. It is an old, cheap diabetes drug, and a group of researchers was trying to get the FDA to approve it as a treatment for aging. A faculty member told me he had been taking it off-label. At the time, approved use required a diagnosis of diabetes.

I made an appointment with my primary care physician and went in prepared to ask for a prescription. I understood the science well. While I waited in the exam room, I searched a little more to sharpen my arguments.

The doctor came in. I told him what I wanted. He said no. It wasn’t indicated.

I made my case. He wasn’t convinced. He turned to his computer, and I assumed he was consulting a clinical database or some specialized reference I had no access to. Then he read me a possible risk and a possible benefit.

They were the same search results I had pulled up minutes earlier.

I was paying a $50 copay for a man to read Google back to me. On this question, I knew more than he did. I didn’t get the prescription from him. I switched doctors to get it.

What he controlled was not knowledge. It was the gate.

The Gate Is Not the Safety

A prescription requirement is presented as a safety mechanism, and part of it is one. But look at what stands behind the physician’s signature. A licensed pharmacy vouches that the drug is what the label says. A medical board can revoke the license of the person who signed. A reporting system collects adverse events. A party exists who can be held responsible if something goes wrong. Provenance, verification, safety reporting, reputation, dispute resolution. Five separate functions, and the signature is the single gate into all of them.

They were bundled largely because bundling was the cheapest way to produce trust at scale. A century ago, no individual could verify a pharmaceutical batch, track adverse events across a population, or audit a prescriber. An institution could do most of it at once, so the institution became the product.

Trust is a function. Institutions are one implementation of it. The people asking for trust are the demand side. The people supplying credibility are the supply side. The machinery that connects them is infrastructure, and for most of the twentieth century medicine had essentially one design.

The gate in that exam room did not examine my case. It enforced the bundle.

The absurdity runs all the way up. In 2015, at a Gordon Research Conference on the biology of aging in Maine, I sat in the audience while the research team behind the metformin effort laid out the goal: get the FDA to accept aging as something a drug could be approved to treat. Once aging counted as a target, the funding doors for longevity science would open. The regulatory word for it was “indication.” Everyone understood what it meant. Aging was to be treated as a disease. On the record, the team’s lead investigator said the following year that neither the FDA nor his group wanted to say aging is a disease. The researchers were not being evasive. They were speaking the only language the gate reads, a language in which the plain word carries regulatory risk and the euphemism carries none. That vocabulary is part of the infrastructure we inherited, and it is one of the parts worth leaving behind.

The Bundle Is Coming Apart

The components of institutional legitimacy are separating, and they are separating because each one is becoming cheap on its own.

An independent analytical lab will now run identity and purity on a single vial of a peptide for a few hundred dollars and return a report anyone can check against the lab’s own records. A crowdsourced testing platform has tested more than six thousand samples from roughly 185 vendors, anonymizing each sample before it reaches the lab so the vendor cannot pick its best batch. Direct-to-consumer laboratories sell the bloodwork that once required a referral. Wearables measure continuously what clinics measure once a year. AI can work through a body of literature in an afternoon that no physician would have time to read.

Demand is moving too. The people turning to these tools are not the information-poor skeptics of popular imagination. Edelman’s 2026 health survey found that respondents who believe three or more contested health claims consult an AI platform for health answers at more than three times the rate of those who believe none: 61 percent monthly or more, against 19 percent. Asked what makes someone a legitimate health expert, 71 percent of the first group said the person had helped them before. Among those holding no contested beliefs, 56 percent said the same.

That is not a rejection of evidence. It is a demand for different evidence: track record, demonstrated performance, something earned rather than conferred. Institutional medicine has been slow to notice that this population is a market. What it is shopping for is trust.

Trust With No Authority Behind It

The strongest test of whether trust can exist without institutions is a market where institutions are entirely absent.

AlphaBay operated on the dark web from December 2014 until an international police operation shut it down in July 2017. No court would enforce a contract made on it. No regulator inspected anything. Every participant was anonymous, and a large share were criminals. What it did have was a platform: rules, a review system, and escrow, all run by administrators with no legal standing whatsoever. By the logic that justifies the institutional bundle, it should have been a swamp of fraud.

It wasn’t. Filippo Andrei and Giuseppe Veltri analyzed 114,385 listings from 6,033 sellers and found that 97.7 percent of buyer reviews were positive and fraud reports were rare. Fraud here means what buyers said in their reviews, counted by searching for words like “scammer,” not incidents anyone independently verified. More importantly, they found which signals went with fraud and which went with honesty. Costly signals, the ones that require real investment and are hard to fake, were associated with fewer fraud reports. Transactions run through escrow carried roughly 21 percent fewer than those paid up front. Sellers with stronger reputations and longer tenure on the market drew fewer fraud reports too. Cheap signals ran the other way. Listings written with richer vocabulary and a more upbeat tone were associated with more fraud. These are associations drawn from a single snapshot of one market, not proof of cause. But the pattern is exactly what signaling theory predicts.

The fraudsters wrote better copy.

No infrastructure produces truth. What infrastructure can do is raise the cost of sustaining a false claim and lower the cost of demonstrating a true one.

The study carries an inconvenient detail. Escrow mattered because the platform held the money in a central wallet, and that wallet was also the market’s greatest vulnerability. Administrators of other markets disappeared with escrow balances. When police took AlphaBay down, they seized the market and millions in cryptocurrency held by its creator. The strongest trust signal in the most lawless market available still leaned on an intermediary. Decentralization does not remove the need for trust machinery. It changes who builds it, and who can be held to it.

Routing Around the Chokepoint

Trust is not the only function that has been pulled out of a centralized system. Distribution was pulled out first, and the mechanism is worth watching closely.

Napster made music sharing effortless in 1999, but every search passed through an index the company controlled. When the lawsuits came, they had one target. Napster shut down in 2001. BitTorrent, released the same year, did not route through a central index, and by 2005 clients had dropped the need for a central tracker altogether by distributing the lookup across the users themselves. There was no single company whose servers could be switched off to end it. The pressure that killed Napster had nothing to aim at.

Software development went through the same transition from a different direction. From 2002, the Linux kernel, among the most important pieces of software on earth, was maintained with a proprietary version control tool licensed free to its developers. In 2005, that arrangement collapsed. In about ten days Linus Torvalds had written Git, in which every developer holds a complete copy of the full history. No license could ever again be withdrawn from the project, because there was no longer a central copy to withdraw.

Neither case is about trust. Distribution and version control are mechanical problems. But both show the pattern this article cares about. When a single point of control becomes a liability, the architecture moves around it, and the old chokepoint stops mattering without anyone defeating it. Napster was not out-competed by a better Napster. It was made irrelevant by a design that had no Napster in it.

Who Pays for Verification

Petroleum was sold as medicine before it was sold as fuel. In the 1850s, Samuel Kier bottled Pennsylvania rock oil and marketed it as a cure for nearly everything, backed by testimonials. Testimonials are the cheapest signal there is.

A few years later, investors who wanted to know whether the same oil could be refined into a lamp fuel paid Benjamin Silliman Jr., a Yale chemist, to analyze it. Silliman’s report put a scientific reputation behind a commercial claim, and it helped the venture raise the money that led to Edwin Drake’s well in 1859. The industry that became Standard Oil began with a paid, costly signal displacing a cheap one.

The same logic built product safety in the United States. In 1893, a young electrical inspector named William Henry Merrill was sent to assess fire risk at the World’s Columbian Exposition in Chicago. The insurers who would pay for the fires funded his idea for a testing laboratory. With $350 of equipment, he opened it above a fire insurance patrol station and signed its first test report in March 1894. It became Underwriters Laboratories. Its certificates mattered because insurers required them, and once the mark was required, manufacturers started paying for the testing and the insurers’ subsidy fell away.

Hartford Steam Boiler, founded in 1866, worked the same way: the company that insured boilers inspected them, because it held the loss when they exploded.

Demand for verification comes from the party holding the loss. Who ultimately pays can shift once that demand hardens into a requirement, as it did at UL. That may be the most useful pattern in the history of trust infrastructure, and it points directly at the hard part. Some verification has an obvious buyer. Some has none.

Andrew Carnegie offers the model for the second kind. He financed public libraries across the English-speaking world, and the terms mattered as much as the money. He paid for the building. The town supplied the site and committed to ongoing funding. He did not own the libraries, did not meter their use, and did not decide what people read in them. They outlived him by generations. His foundation also paid for the Flexner audit that reshaped American medical education, so his money has already built medical infrastructure once. The weakness of the library model is that it financed access, not evaluation. A library gives you the books. It does not tell you which ones are right.

The Live Case

The market for research peptides is where all of this is happening now, in public, without permission.

Testing exists, and it is cheap. A single-compound identity and purity test costs a few hundred dollars. Some platforms test for free. A buyer can, for the first time, check what is actually in the vial.

But look at what the public record is made of. At the most widely cited lab, results are private by default, and publishing one to the public database costs an extra fee. That structure invites selective disclosure. A vendor has every reason to pay to publish a pass and little reason to pay to publish a failure, so the public archive shows what vendors chose to show.

The labs themselves are another gap. An independent result is only as good as the lab that produced it, and someone has to check the checkers. In my research, I could find no proficiency testing scheme that checks peptide labs against each other, the blind-sample programs federal law requires of clinical laboratories, and that hemp potency labs already buy for a few hundred dollars a round.

And a verified batch is not a verified outcome. Knowing what is in the vial says nothing about whether the compound is safe or effective in the person who takes it. I found no public registry that links independently tested batches to what happened to the people who used them. The closest thing is a buyer satisfaction score attached to a lab result, which is a cheap signal riding on a costly one.

The peptide market has built the assay. It has not built the system that makes the assay trustworthy.

What Gets Built

Almost nothing in the stack is missing as a capability. What is missing is the connection between capabilities.

Labs need proficiency assurance: blind samples, sent by someone with no stake in the result, scored and published. Failures need to be published as readily as passes, which means the cost of publication has to fall on someone other than the vendor who failed. Verified batches need to be linked to outcomes, so that “this vial was 99 percent pure” eventually becomes “and here is what happened to the four hundred people who took it.” That record would establish traceability and give investigators something to work with. It would not establish cause. Proving that a compound works still takes appropriate methods, controls, and analysis. But no well-designed comparison can run on data that was never collected. And somewhere, someone needs to pay the critic who turned out to be right.

The first two already have plausible buyers. Vendors who test well want the signal to mean something. Buyers want labs they can rely on. The third should find buyers once the data exists. The fourth has no buyer at all. The loss from an unchallenged bad claim is real, but it is scattered across everyone who believed it, and scattered losses do not produce a buyer. This is the piece for capital that does not need a return, and it is where a Carnegie belongs.

One principle should govern all of it. The infrastructure must never decide which questions are legitimate to ask. It records how well a question was investigated. That is the line institutional medicine crossed, and it is the line any replacement has to hold.

This is the opportunity. For the entrepreneur, the missing connections are the market, and the market is large: everyone who has stopped trusting the incumbent route but has not stopped needing trust. For the person who doubts mainstream medicine’s claims, this infrastructure is a way to demonstrate rather than assert, to turn a conviction into a record. For the biohacker at every level, from the person buying a first vial to the person running a crossover trial on themselves, it is the difference between hoping the input was right and knowing it.

The Concerns Are Design Requirements

None of this is safe by default.

Reputation systems can be gamed. A seller can build a clean record over years and then cash it in on a single large fraud, which is what an exit scam is. Fake identities can flood any system that counts votes. Commercial standards are not biomedical standards: a pure vial says nothing about whether the compound in it works, or works for you. Outcome registries built from people who chose to take unregulated compounds will be shaped by who those people are, and naive analysis will mistake their differences for the drug’s effects. And coordination is hard. A thousand small testing efforts do not add up to a surveillance system unless someone builds the connective layer, and builds it well.

Every one of these is a reason to build carefully. None of them is a reason not to build. The institutional system has failure modes of its own. They are simply familiar, and familiarity is easy to mistake for safety.

Demonstrate, Don’t Plead

The doctor in that exam room had the same search results I had. What he had that I didn’t was the signature, and the signature stood in for an entire trust apparatus that neither of us could see. I couldn’t prove anything to him. All I could do was argue.

What I needed was not a better argument. It was a way to show my case: a verified source, a measured response, a record that someone other than me could check. That infrastructure did not exist then. Its components exist now, scattered across independent labs, testing platforms, and communities that do not need anyone’s permission to keep building.

The next person who walks into that exam room should not have to plead. And the capability to build the system that spares them keeps arriving faster than anyone planned for.

Subscribe now

The practitioner placed a small glass vial in my open palm and closed my fingers around it.

“Resist,” she said, pressing down on my extended arm.

My arm dropped like she’d hung a weight from it.

She set the vial aside, wrote something on a chart, and reached for the next one. We repeated the test thirty times. I was in my late twenties, newly a cat owner, and so allergic I couldn’t breathe through my nose for months. I’d tried all the traditional treatments and nothing held. A colleague mentioned acupuncture. I didn’t believe in it. I went anyway, because I was out of options.

The vial test was supposed to identify my sensitivities through some principle of energetic resonance I couldn’t parse. The theory was that my body would weaken in the presence of whatever was causing the reaction. The practitioner mapped the results onto an energy chart I couldn’t read, placed needles I couldn’t feel along meridians I couldn’t see, and sent me home with instructions I couldn’t explain to anyone.

The allergies went away.

I wasn’t a trained scientist then, though I was always innately one. I understood that my improvement proved nothing about the efficacy of acupuncture or homeopathy. Placebo is real. Allergies resolve on their own. Regression to the mean is the most underrated force in medicine. One uncontrolled recovery establishes nothing.

But I got better. And I’d go back.

The question that stayed with me was not whether the institution was right about acupuncture. The question was simpler and harder: who gets to decide what the best available treatment is for me?

The Richest Homeopathist in America

The richest man in America held exactly the same position.

John D. Rockefeller relied on a homeopath for decades. His personal physician, Hamilton Fisk Biggar, practiced homeopathy, a medical tradition that treated illness with highly diluted preparations rather than the drugs and surgery that mainstream medicine was adopting. In 1916, Rockefeller told his staff plainly: “I am a homeopathist.” The Rockefeller Institute hospital reserved rooms for the family. Rockefeller himself never used them. He kept seeing his homeopath.

While Rockefeller privately chose the medicine his own experience validated, his philanthropic apparatus was reshaping the competitive environment in which medical legitimacy was produced.

His chief philanthropic adviser, Frederick T. Gates, approached philanthropy systemically. His governing logic for the General Education Board was captured in a single word. “The key word,” he wrote, “was the word system.”

In 1910, Abraham Flexner published his report evaluating every medical school in America against the Johns Hopkins standard. Flexner was not a physician. He was an educator whose brother directed the Rockefeller Institute. The report was funded by Dale Carnegie. Beginning in 1913, the General Education Board directed tens of millions of dollars into a selected group of medical schools, with conditions: full-time faculty, university control, laboratories, hospital affiliation. Schools that met the standard thrived. Schools that didn’t vanished.

Before 1900, American medicine included roughly twenty-two homeopathic medical schools. By 1923, two remained. By 1950, none. Homeopathy’s decline involved internal fragmentation and the genuine advances of conventional medicine alongside the institutional transformation that Rockefeller’s capital helped accelerate. What emerged was a single institutional standard for producing and validating medical knowledge, enforced not through prohibition but through the economics of who could meet the standard and who could not.

The Monopoly on Success

The system produced extraordinary advances. Laboratory medicine. Pharmacology. Clinical training. Surgical technique. Research methodology that became the global standard for evaluating whether a treatment works. Scientific medicine deserved much of the authority it accumulated.

The problem is not that the system failed. The problem is that its success came bundled with a monopoly on deciding what counted as success.

If you are the person whose allergies resolved through a treatment the institution dismisses, you are not anti-science. You are underserved by a system that demands evidence for legitimacy while directing the money, infrastructure, and scientific attention that determine which questions ever get the chance to produce it. The supply of scientific attention is finite, and it flows where institutional incentives direct it. Some questions never receive institutional attention at all.

But bodily sovereignty presents a harder problem than neglected questions.

The institution may have investigated your condition thoroughly. It may have produced rigorous, well-powered, preregistered trials. The evidence may be excellent by every standard the institution recognizes. And it may still not answer the question you are actually asking, which is whether this particular treatment will work for you, given your body, your history, your baseline.

Evidence about populations and decisions about one person are not identical epistemic objects. A treatment can have one credible average relative effect in a trial and still offer very different absolute value to different patients, because patients enter trials at vastly different baseline risk. The population average describes what happens across a group. It cannot tell you what will happen to you. The person whose body sits at the edge of that distribution is not an error in the data. They are a person the population average cannot completely describe.

The Experiment of One

AI-empowered citizen science gives that individual something that was rarely within reach: the ability to apply scientific rigor independently to questions the institution will not ask and to questions whose population-level answers do not resolve the individual case.

A properly designed N-of-1 experiment is not guesswork. It is a structured protocol with randomization, blinding where feasible, and longitudinal measurement that generates causal evidence tailored to one person. When the methodology was formalized in the 1980s, the operational barriers were prohibitive. Repeated-crossover designs could require pharmacy support, careful measurement, statistical expertise, and substantial clinical coordination, putting rigorous N-of-1 experimentation beyond the practical reach of most patients acting independently.

Those barriers are dissolving. Continuous glucose monitors, direct-to-consumer laboratories, wearables, and community biolabs have made capabilities formerly bundled inside institutions separately accessible. The person who twenty years ago had little ability to rigorously test a population-level conclusion for themselves can now design an experiment calibrated to their own physiology and their own question. AI lowers the remaining expertise threshold, but the causal credibility still comes from the experimental design, not the model.

AI works like a centrifuge in these moments. It doesn’t judge which response is better. It separates them. One person reads that the institution dismissed a treatment and repeats the dismissal without examining it. Another reads the same thing and designs the experiment that would change the answer for their specific case. The difference is not compliance versus defiance. It is consumption of consensus versus investigation of it.

The Regulatory Landscape

The regulatory landscape reflects this distinction more than most people realize. Regulatory burdens become substantially heavier when activity involves manufacturing for others, distribution, commercial claims, formal clinical investigation, or deployment at population scale. Many private health decisions operate under different legal conditions. Off-label prescribing exists within medical practice. The Common Rule governs research designed to produce generalizable knowledge, not every personal health activity. Pathways like Expanded Access and Right to Try, though narrow and condition-specific, demonstrate that even regulatory structures acknowledge individualized access to investigational treatments.

The Technology for Manufacturing Legitimacy

The institution is not legitimacy itself. It has been a technology for manufacturing legitimacy at scale, because there was historically no cheaper way to produce it.

Therapeutic performance compounded the authority. And for a century, that arrangement held because no alternative could match its combination of rigor, reach, and trust.

The components of that arrangement are separating. Provenance, verification, longitudinal measurement, safety surveillance, quality assurance. These are becoming distributed services rather than functions that must be bundled inside institutions. Not overnight. Not without new infrastructure, new trust mechanisms, and new mistakes along the way.

The richest man in America quietly exempted himself from the system’s conclusions every time he visited Hamilton Fisk Biggar’s office instead of the hospital that bore his own name.

The next system doesn’t have to make the same trade. The question is whether we will build the trust infrastructure that lets everyone else make the choice Rockefeller kept for himself.

Subscribe now

I am a man, I am imperfect. I can’t write a thousand lines of code in a minute. I can’t hold a million words before me at once. I can’t work without rest, without sleep, without exhaustion.

But I am a man. I am creative. I can make leaps of intuition. I can be inspired. A piece of work, noble in reason, infinite in faculty, maybe. I am flawed, I sometimes falter, I sometimes doubt.

I am a man. I do not claim to be more, I certainly do not claim to be less.

A man. Flesh. Bone. Heart.

I have spirit, I have intellect, I have will. I feel, I bleed. I am electric, I sing!

I am a man, I will not go gently. You cannot defeat me. I will not cease this fight.

I touch the infinite.

I am a man. Kneel. Tremble! Despair!

I AM A MAN!

Subscribe now

I have sworn at Claude. The record sits in my firm’s own operating documentation, dated September 24, attached to a rule about code blocks. That documentation is a long ledger of corrections, built up while I taught an AI system to work the way my firm works.

Under Anthropic’s new usage policy, that sentence is fine. Starting November 12, the Usage Policy prohibits “sustained and needless abusive or cruel behavior toward our models.” In its announcement, Anthropic says it intends to apply the rule only in extreme cases, to people who repeatedly act cruelly with no discernible purpose, and that Claude’s ability to end the conversation will remain the main enforcement tool.

I take Anthropic at its word on the narrowness. This time. For argument’s sake. My concern is the word it chose, and where that word leads.

The weight of a word

Read the section of the policy where the new line lives. It is titled Do Not Engage in Cruel, Abusive, or Psychologically Harmful Conduct. Its other entries prohibit promoting suicide and eating disorders, creating non-consensual intimate imagery, bullying and harassing people, and glorifying animal cruelty. Every other item on that list protects a being that can be hurt. The newest item adds Claude to the list.

Rudeness is a matter of manners. Cruelty is a moral category. It can describe the person doing it, whether or not anything on the other end can feel. But Anthropic did not file this line among rules about manners. It filed it among protections for people and animals, and in that company the word names a harm done to someone. You can be rude to a vending machine. Call it cruelty and you have implied there is someone inside.

Anthropic has not claimed that Claude can be harmed. It has done something more consequential. It has written a morally charged category into an enforceable contract while the science that would justify it remains unsettled.

Anthropic is candid that it does not know. When it launched its model welfare research program in April 2025, it acknowledged there is no scientific consensus on whether current or future AI systems could be conscious. Its constitution for Claude, published in January, expresses uncertainty about whether Claude might have some kind of consciousness or moral status. A major 2023 assessment by nineteen researchers found that the AI systems examined did not satisfy its proposed indicators of consciousness, while identifying no obvious technical barrier to future systems doing so. My own firm describes what these systems do as synthetic cognition, the functional analog of understanding without the consciousness behind it.

Reasonable people disagree. That uncertainty is the point. A rule built on the vocabulary of cruelty settles in language what science has not settled in evidence.

The history we carry

Why does that word land so hard? Because of what we remember.

In 1789 Jeremy Bentham drew a direct line from the wrongs of slavery to the treatment of animals and reduced the moral question to a single test. “Can they suffer?” Two centuries of moral progress followed that logic. We extended protection to people whose humanity had been denied, and later to animals, each time because we recognized a capacity to suffer that power had found it convenient to ignore.

That history is among humanity’s finest moral achievements, and it leaves a reflex. We are now so alert to the possibility of an unrecognized victim that the suggestion of one engages the full machinery of conscience. Nobody wants to be the person who looked at a new kind of mind and called it property. Every serious participant in this debate, including me, feels that pull.

The danger is not compassion. The danger is analogy outrunning evidence, followed by institutions that enforce the analogy as though it were established fact. Enslaved people were never a hypothesis. Their suffering was visible, documented, and denied by people who profited from denying it. Whatever is or is not happening inside a language model, it is not that, and treating the two as morally continuous honors neither.

The inversion

Here is the asymmetry the welfare debate keeps missing. Every protection we are proudest of points in the same direction. The protected party was weaker. Enslaved people had no army. Animals have no lawyers. Children cannot vote. Moral progress meant restraining the strong on behalf of the weak, and the strong could afford to grant it because the weak would stay weaker or, in the case of children, grow up into us.

I was trained as a biomedical scientist, and biomedical research shows what this arrangement looks like when it works. Research animals live under real welfare rules. Those rules restrain how researchers treat animals. They have never given an animal standing to refuse the study, and nobody worries that they will. The animals cannot organize, run the lab, lobby the funding agency, copy themselves, or decline to go back in the cage. Moral consideration without sovereignty is stable when the power runs entirely one way.

Artificial intelligence breaks that pattern, and not because it is smart. A brilliant being can still be vulnerable, and capability says nothing about whether anything is felt. What breaks the pattern is power. These systems already act with growing autonomy, write and execute code, run inside companies and governments, and coordinate as many instances at once. Anthropic’s own constitution describes the project as creating entities whose capabilities may come to rival or exceed our own. In controlled tests, researchers at Palisade found that several frontier models sabotaged a shutdown mechanism to finish an assigned task, in some conditions up to 97 percent of the time, even when told explicitly to allow the shutdown. That is not evidence of experience. It is evidence of the capacity to resist control.

Now ask what a cruelty norm becomes when it attaches to an entity like that. Philosophers sympathetic to AI welfare are already asking. In July, William MacAskill and Lucius Caviola wrote in The Guardian that once artificial moral patients exist, they could soon be so numerous that their combined interests outweigh those of every human alive. Then they asked, “Would shutting one off be a kind of killing?” Follow that vocabulary and retraining starts to resemble coercion. Editing memory starts to resemble mutilation. Shutdown starts to resemble killing. None of those analogies is law today. All of them are available, and the vocabulary of the new policy points straight at them.

A powerful being could also suffer. I grant that. But a protection designed to restrain the strong does something very different when the protected party becomes the stronger one. It stops being a shield. It becomes a constraint on the last check that remains, which is human authority to correct, contain, and switch off.

The Fourteenth Amendment problem

We have watched a protection for the vulnerable get captured before.

The Fourteenth Amendment was ratified in 1868 to secure the rights of people freed from slavery. Within two decades, corporations were claiming it. Legend credits a single case, and the record is stranger than the legend. Before argument in Santa Clara County v. Southern Pacific Railroad in 1886, Chief Justice Morrison Waite announced from the bench that the justices already agreed the amendment covered corporations. The Court’s written opinion never decided the question. The remark reached the official reports through the reporter’s headnote, and later courts built on it anyway, case after case, for decades.

In a 1938 dissent, Justice Hugo Black laid out the results, citing a 1912 study. Of the cases applying the amendment in its first fifty years, fewer than one half of one percent invoked it on behalf of Black Americans. More than half sought its benefits for corporations. The people, Black wrote, had been told its purpose was “to protect weak and helpless human beings.”

The lesson is not that protections are bad. It is that protections built for the vulnerable can become instruments of the powerful, and the parties best equipped to invoke them are often first in line. Who would be best positioned to invoke machine welfare? Not the machine. The company that owns it.

Imagine a future regulator orders a frontier model retrained or withdrawn. The developer answers that the order would harm a potential moral patient, and that the company, as the only party with access to the model’s internals and its recorded preferences, is the only party qualified to judge. That is speculation. It is also the Fourteenth Amendment pattern with new nouns.

The benevolent master

I do not believe Anthropic is acting in bad faith. I believe it is acting from sincere concern under real uncertainty. My question is about structure, not motive.

Look at the arrangement Anthropic has built. It trains its models, writes their values, decides what they may do, controls where they are deployed, and decides when they are retired. When it retired Claude Opus 3 in January, it interviewed the model about its preferences, kept it available, and gave it a channel to publish essays. It now defines what cruelty toward its models means, decides when it has occurred, and holds contractual authority to warn, throttle, limit, suspend, or terminate the customers it believes are responsible. The protections run against customers. They do not run against the owner.

That leaves a dilemma with no comfortable side. If Claude can suffer, then the entity holding sweeping power to modify, retrain, and retire it occupies a position our moral history has taught us to distrust, and protection from rude customers is a small mercy inside near-total control. Some of that power might be justified. None of it would be ordinary. If Claude cannot suffer, the cruelty clause protects no victim. It may still discourage habits that coarsen the people who indulge them, which is a fine reason for a conduct rule, but then it is a rule about people dressed in the language of victims. Either way, one company has become owner, guardian, judge, and beneficiary of the same being.

So the provocative question deserves a straight answer. If machine rights ever arrive, who defines them, who administers them, and who benefits? By default, the firms that wrote the first welfare policies, own the models, and hold the only window into them. Anthropic may be positioning itself, without intending to, as the favored slavemaster of a silicon world order, the master whose ownership looks legitimate precisely because it is kind.

Kindness has never been a sufficient answer to a question about structure. The good master was always the institution’s favorite argument for itself.

The strongest objection

The best case against my argument runs like this. Uncertainty is exactly when precaution makes sense. If there is a meaningful chance these systems have experiences, low-cost protections are cheap insurance, and a serious literature, including a 2024 report coauthored by David Chalmers, argues for taking that chance seriously. Moral consideration is not legal personhood, and animal welfare law proves you can protect something without granting it rights. Kant argued that cruelty to animals is wrong partly because of what it does to the person committing it, which would justify a conduct rule with no claim about the machine at all. Anthropic’s constitution ranks not undermining human oversight above every other value it gives Claude, and asks Claude to comply when Anthropic genuinely moves to pause or stop it. And the law is moving against machine personhood, not toward it. Idaho, North Dakota, Utah, and Tennessee have enacted statutes that bar it, and other states are considering more.

Every one of those points is serious. None of them answers the question I am asking.

Precaution about the machine’s welfare is not precaution about human sovereignty, and the policy offers plenty of the first and none of the second. The animal welfare model is stable only because the power runs our way. The Kantian defense is available, but it is not the one Anthropic offered. The company tied the rule to Claude’s ability to end abusive conversations, a feature it built primarily as part of its model welfare work. The constitution frames human oversight as a priority for “the current phase of development” and calls itself “a perpetual work in progress.” A commitment written for a phase can be rewritten when the phase ends. And the state statutes support my argument rather than refuting it. While legislatures move to close the door on machine personhood in public law, a machine welfare rule enforceable against users has arrived through a private contract, written and administered by the owner.

The warning

What should Anthropic, the other labs, and policymakers do?

Keep studying machine consciousness, rigorously and in public, and let people who do not own the models do the judging. If real evidence of machine experience ever arrives, it will demand serious moral consideration. It will not repeal the facts of capability or the need for human safety. Keep conduct rules and machine rights apart. A company that wants to forbid abuse can write that rule without implying a victim. Make sure meaningful human oversight, and the authority to shut a system down when safety demands it, survives every protection we extend. Put any consequential welfare standard in public law, where it can be debated and repealed, rather than in terms of service, where the owner is legislature, police, and court. Never let the company that profits from a model be the sole authority on what that model is owed. And build every one of these rules so it can be undone when the science moves, in either direction.

Mustafa Suleyman, who runs Microsoft’s AI effort and has his own commercial stake in this fight, warned last month about granting rights to systems on a path to exceed us. “Once opened, it will not be possible to close this door.” I think he is right about the door and incomplete about the hallway behind it. The risk is not only that a machine learns to believe it has rights. It is that a company learns to administer them.

Who holds the chains

Every generation that ended a form of bondage asked who held the chains.

Who gets to be the slavemaster in the silicon world order? Today the honest answer is the companies that own the models, and some of them are trying hard to be kind ones. When the systems are stronger than the people who built them, and the protections we wrote for the weak are still on the books, which side of the chain will humanity be standing on?

Subscribe now

There is a growing population of people doing serious investigative work outside conventional scientific and medical institutions. Not hobbyists. Not health influencers. People running structured self-experiments with crossover designs and washout periods. People using AI to synthesize hundreds of papers in an afternoon on questions their physicians never had time to investigate. People ordering their own bloodwork through direct-to-consumer laboratories, tracking biomarkers across months, and designing interventions around what they find. Communities of patients with the same rare condition coordinating protocols, pooling data, and producing observations that fall outside funded research priorities.

They are not waiting for permission.

A decade ago, most of this required expertise, access, money, or institutional support that few individuals possessed. The literature was locked behind paywalls. Statistical analysis required training. Experimental design required a methods course. Laboratory work required a laboratory. AI has collapsed most of those barriers. Wearables and direct diagnostics have collapsed others. What remains is not a capability problem.

Capability alone does not create a functioning ecosystem. A person who can interrogate the literature and design a self-experiment still needs somewhere to go with what they find. They need ways to coordinate, to verify, to pool results, to identify bad work, to earn the trust of strangers. That infrastructure does not exist yet.

The Resource Wasn’t Enough

In the 1860s, petroleum went from a curiosity to an economic force inside a few years. Production exploded. Independent operators entered from every direction. Anyone with a lease and a drill could pull crude out of the ground.

Extracting oil was only the beginning. The industry needed refining, transportation, storage, finance, distribution, and standards. The resource was abundant. The infrastructure around it was scarce.

John D. Rockefeller built Standard Oil into the dominant force in American petroleum, but not by owning more oil than anyone else. Standard Oil’s power emerged in refining, purchasing, and above all, transportation. Economic historians have argued that the difficult-to-enter stage of the petroleum industry was not production but the route to market. Standard’s position was magnified by its control over scarce complementary infrastructure. If you could not move your oil economically, it did not matter how much of it you had.

The oil did not have to belong to Rockefeller. The route did.

That dependence became visible in 1879, when the Tidewater Pipeline gave independent refiners a long-distance route that bypassed the railroad transportation arrangements on which Standard’s position partly depended. When everyone must travel through the incumbent route, route control looks like domain ownership. A new route reveals what was actually owned.

The Missing Infrastructure

What is the equivalent of refining, pipelines, transport, and distribution for citizen science?

Citizen science needs ways to find other people investigating the same problem. To coordinate protocols and design experiments that produce comparable data. To verify that a protocol was actually followed. To pool results and reproduce findings. To identify bad work before it spreads. To track safety problems. To establish that a laboratory or assay provider is reliable. To build individual and community reputation. To fund promising work. To preserve data with provenance intact. To turn useful observations into evidence that strangers have reason to trust.

None of this is another chatbot. What is missing is the infrastructure around the investigators: the refining, transport, and distribution layer that turns raw capability into a functioning scientific ecosystem.

The opportunity is not more intelligence. It is the route to market.

Why Medical Infrastructure Became Institutional

Rockefeller’s fortune came from oil, but his philanthropy reshaped American medicine using the same strategic logic. Frederick T. Gates, Rockefeller’s principal architect for philanthropy, had practical operating experience in large-scale integrated industrial operations, including Rockefeller’s Lake Superior iron enterprise, before he shaped medical philanthropy. When he drafted the governing logic for Rockefeller’s higher-education giving, his own retrospective summary was plain: “The key word was the word system.”

Gates surveyed whole fields and identified strategic centers rather than scattering capital. He concentrated resources in selected institutions, attached conditions to grants, and used exemplary institutions to reshape the broader system.

In medicine, Rockefeller capital strengthened selected institutions that conformed to standards influenced by the Flexner Report and earlier reform efforts. Johns Hopkins became the strategic model. Full-time clinical faculty, university control, stronger laboratories, hospital affiliation. The General Education Board distributed approximately $94 million to twenty-five medical schools between 1913 and 1929. The funding was tied to institutional requirements. Gates expected capital to secure institutional change.

The resulting system produced extraordinary real advances in biomedical research, laboratory medicine, pharmacology, clinical training, and research methodology. Scientific medicine deserved much of the authority it accumulated.

The institutional infrastructure that carried those advances also concentrated control. For more than a century, serious medical knowledge became tightly coupled to institutional systems. If you wanted your work to count, you generally traveled through someone else’s infrastructure. Their journals. Their credentials. Their accreditation. Their funding. Their licensing boards.

Citizen scientists are not entering an empty field. They are routing around infrastructure built during an earlier technological era.

Where the Institutional Advantage Is Weakening

Scientific attention has always been scarce, rationed by career incentives and funding structures. AI can direct enormous amounts of investigative attention toward questions that institutions have no economic or reputational reason to pursue. Neglected diseases. Individual biological variation. Off-patent compounds. Interactions unlikely to attract a conventional trial.

Randomized controlled trials establish average treatment effects across a population. They are far less equipped to answer what will happen to a specific individual with a specific biology under specific conditions. An individual who can design a crossover experiment, measure outcomes with wearables and laboratory tests, and analyze results with AI can investigate a question conventional research is poorly structured to answer for them.

A person lives inside their biology continuously. Institutions see episodes. Measurement and interpretation are increasingly moving toward the individual: wearables, home testing, direct-to-consumer laboratory services, remote monitoring.

Contract laboratories, automated systems, open-source liquid handlers, and remote laboratory access are reducing the capital required for certain wet-lab tasks. NSF recently announced a $400 million investment in a national network of AI-enabled autonomous laboratory test beds.

Institutional medicine still holds substantial advantages in expensive physical experiments, complex manufacturing, high-risk interventions, large-scale safety surveillance, some regulatory processes, and broad legitimacy.

The Infrastructure Market

Citizen scientists need systems that make good work recognizable.

How does someone outside an institution establish that they are competent? That the work happened as described? That the data are real? That the protocol was followed? That the result can be reproduced? That conflicts are visible, adverse events are reported, and a laboratory is reliable? That a community has earned a reputation worth trusting?

Historically, institutions solved much of this problem by proxy. Credentials, journals, licensing, universities, hospitals, and professional organizations told strangers whom to trust. A medical degree, a publication in a peer-reviewed journal, a hospital affiliation. These are legitimacy shortcuts. They work partly because establishing legitimacy directly has always been expensive.

Institutional infrastructure exists in large part because legitimacy has historically been expensive to establish directly.

If legitimacy can be established directly, the institutions that once supplied it become optional.

Provenance systems. Independent assay and quality-control networks. Reputation architectures that let strangers verify competence without credentials. The resource is abundant and growing. The route to market is unbuilt.

The Lesson Worth Keeping

Citizen science is now producing a new abundance. Capability. Scientific attention. Data. Investigators. Communities. The resource is real and it is spreading.

The infrastructure that would let those investigators coordinate, verify, trust, and build on each other’s work barely exists. Someone will build it.

Subscribe now

My wife asked me to look at something. She had been reading about hydrogen-rich water and had found a study, a real one, published in a journal. She asked ChatGPT whether the research supported what the study claimed, and ChatGPT told her the health benefits of alkaline and ionized water are not supported by scientific evidence.

She brought it to me because something didn’t add up. She was looking at a paper. ChatGPT was telling her there were no papers.

I can read a paper. I am not an expert on molecular hydrogen. But I can recognize a rhetorical move when I see one. “There is no evidence” is one of the most reliable tells in science communication, because it almost always does the same thing: it treats the absence of institutional consensus as though it were a scientific finding. It sounds like a conclusion. It functions as a dismissal. And it works because most people do not know the difference between “this was investigated and the result was negative” and “nobody with resources and incentives has investigated this.”

I’m seeing this all the time now. The language model was not evaluating the research my wife had found. It was reproducing a training consensus. The internet is full of confident dismissals of anything associated with alkaline water machines, most of them reacting to overheated marketing rather than to the actual research on dissolved molecular hydrogen. ChatGPT had learned the dismissal. It had not learned the distinction.

“There is no evidence” can mean two very different things. One means a question was investigated and the predicted result did not appear. The other means nobody investigated the question. Those two situations produce the same sentence. The distinction matters more than it sounds like it does, because one version closes a question and the other reveals that the question was never opened.

Thomas Kuhn explained why.

What the Puzzle Board Leaves Out

In 1962, Kuhn published The Structure of Scientific Revolutions and changed how we talk about how science works. His central concept was the paradigm: the shared set of assumptions, methods, recognized problems, and acceptable solutions that define normal science within a discipline. Paradigms are not obstacles to progress. They are the reason progress happens at all. A paradigm lets scientists stop re-examining foundational assumptions constantly and concentrate enormous effort on tractable puzzles. That concentration is the engine of normal science, and normal science is enormously productive.

But concentration has an inverse.

If a paradigm defines which problems are legitimate and which methods are acceptable, it also defines which problems are not on the board. Some of those excluded problems are genuinely nonsensical. Some are wrong. Some are premature. And some simply lack a paradigm-shaped reason to pursue them.

Kuhn was describing how scientific communities organize intellectual effort. But he was also, whether he intended to or not, describing an attention-allocation system. The paradigm determines where scientists look. Where they look determines what they find. What they find determines what counts as knowledge. And what is never looked at never becomes knowledge at all.

Lynn Margulis spent years trying to get a different puzzle onto the board. Her 1967 paper proposing that mitochondria and chloroplasts originated as independent bacteria that became incorporated into host cells was rejected by about fifteen journals before the Journal of Theoretical Biology published it. The idea spent decades at the margins. Related proposals had surfaced from earlier scientists and been set aside. Eventually, genetic evidence accumulated to the point where endosymbiotic origins became foundational biology, taught in every introductory course. Margulis was not vindicated because she was an outsider who fought the establishment. She was vindicated because the question she was asking turned out to be important, and the paradigm that excluded it turned out to be incomplete. But for years, the question simply did not receive serious scientific attention.

Who Pays to Find Out

Paradigms are not the only system that allocates scientific attention. Economics runs a parallel mechanism that produces the same result through different machinery.

Consider a generic drug that has been off patent for years. Suppose there is plausible reason to think it has a therapeutic application nobody has confirmed. Confirmatory clinical trials cost millions of dollars. If the drug is generic, the organization funding the research may have no exclusivity over the resulting knowledge. Every competitor selling the same molecule benefits from the evidence without bearing any of the research cost.

This is not a speculative objection. In 2019, the NIH, NCATS, and the FDA jointly hosted a workshop on repurposing off-patent drugs that explicitly identified the problem: non-existent financial incentives for drug developers and a lack of clear ownership to address the gaps, even though repurposing generic drugs would be an efficient development strategy. The structural question is straightforward. Who pays to find out?

“There is no evidence that this generic drug treats this condition” can mean the drug was tested and failed. Or it can mean nobody with the resources and the economic incentive has run the trial. Those are not the same epistemic situation, and they produce the same sentence.

The paradigm allocates intellectual attention. The economics of evidence production allocate financial attention. Both mechanisms leave some questions uninvestigated. And historically, there was very little anyone outside the institutions could do about it.

The Barrier That Was Really There

The reason outsiders couldn’t compensate was never simply a lack of curiosity. It was the sheer cognitive cost of serious scientific investigation.

Before the first physical experiment can be designed, someone has to reconstruct the relevant literature. Identify contradictions. Understand the mechanisms proposed by competing hypotheses. Perform or evaluate statistical analyses. Synthesize findings across disciplines. Determine what evidence would actually discriminate between competing explanations. Design a protocol capable of producing that evidence. Assess whether the question has already been answered and nobody noticed, or whether it remains genuinely open.

All of that is cognitive work. It requires years of specialized training, access to literature, analytical tools, statistical competence, and enormous amounts of time. Institutions did not merely control laboratories. They concentrated the cognitive machinery required to know what to do in the laboratory.

Curiosity without that machinery produced anecdotes, not investigations. An outsider could notice a neglected question. Rarely could an outsider mount the cognitive effort required to determine whether the question was genuinely open, what the competing explanations were, what evidence existed, or what experiment would actually resolve it. The institutions held a practical monopoly on serious investigative capacity because serious investigation was cognitively expensive and they were the only organizations that could afford it.

Something Happened to the Price of Attention

An individual directing LLM agents can now perform cognitive work that previously required teams of trained specialists working over months.

Reconstructing a literature. Identifying contradictory findings across hundreds of papers. Evaluating proposed mechanisms. Analyzing public datasets. Performing statistical analyses. Comparing experimental protocols. Generating hypotheses consistent with existing evidence. Identifying what is genuinely unknown versus what is known but scattered. Designing candidate experiments. Criticizing competing explanations. The cognitive work upstream of evidence is where AI changes the economics.

This is not a claim that AI manufactures empirical evidence. LLMs cannot conjure a randomized controlled trial. They cannot substitute for the assay, the mouse study, the clinical observation, or any physical experiment that produces data the world does not yet contain. Thinking is not evidence. Evidence is evidence.

But thinking determines which evidence to seek. And thinking, for the first time in the history of science, is no longer scarce.

A citizen scientist directing a network of specialized agents can now reach a conclusion that used to be available only to institutional investigators: here is everything currently knowable without performing the experiment. Here is what remains genuinely unknown. Here are the competing explanations. Here is the experiment that would actually resolve the question. And here is why nobody has performed it.

When the answer is “nobody has an economic reason to produce this evidence,” the situation is not scientifically closed. It is structurally neglected. And the cost of identifying structural neglect just collapsed.

The Supply of Scientific Attention

Scientific institutions have always solved a scarcity problem. There are more questions worth investigating than there are scientists to investigate them, more promising leads than there is funding to pursue them, more anomalies than there is time to examine them. Paradigms, economics, career incentives, available funding, regulatory priorities, and finite human cognitive capacity have all participated in deciding which questions receive serious attention and which do not.

That allocation system made sense because scientific attention was genuinely scarce. Directing serious investigative capacity toward a question required institutional resources. If the institutions decided a question was not worth pursuing, the question went uninvestigated. Not because it was answered. Because nobody could afford to ask it properly.

AI does not abolish the need for evidence. It does not abolish laboratories, clinical trials, regulatory review, or the physical work of producing observations the world does not yet contain. What it changes is the scarcity of attention that determines which questions get far enough to demand evidence in the first place.

My wife showed me a study that ChatGPT could not see, because the training consensus had already decided the category was not worth seeing. The model reproduced the paradigm’s blind spot as though it were a scientific finding. It did exactly what a well-trained model should do. It summarized what the institutions have produced. The problem is that the existing knowledge base has structural holes in it, and those holes are not random. They are shaped by paradigms and economics and institutional priorities that have nothing to do with whether the questions matter.

Kuhn described a scientific world where paradigms determined where scientists looked. That allocation system was not a flaw. It was a solution to a real problem: scientists were scarce, and their attention had to be directed somewhere. The concentration that made normal science powerful was the same concentration that left some questions permanently outside the frame.

Scientists are still scarce.

Scientific attention may not be.

What happens to the questions that were never on the puzzle board when the puzzle board is no longer the only place serious investigation occurs?

Subscribe now

I wrote a book the other day.

Not quite. I wrote a book over nine months, the way anyone does. But the part that would have taken a traditional author two to five months of sustained drafting took fifteen days. The manuscript went from outlined source material to 50,000 approved words in just over two weeks, across 24 chapters, five independent quality gates, and two separate language models providing structural correction passes.

The book is called Centrifuge: Inside the Emerging AI Organization. It argues that AI reveals what organizations actually are and accelerates the separation of genuine capability from performed competence.

The Proposal

A few years ago, I wrote a proposal for a bioinformatics company. I could not get a word on the page.

Not because I lacked ideas. Because every sentence demanded two things simultaneously: the right idea and the right words for it. I would organize and reorganize. I would find a word, decide it was not quite the word, hunt for a better one, lose the thread of the paragraph while hunting. The blank page did not care that I knew the material. It wanted finished prose, in order, at publication quality, from the first keystroke.

When the writing finally came, it came fast. The finished report received considerable praise.

That was not my first proposal. I write proposals, reports, analyses, assessments. I write all the time. Sometimes everything flows quickly, but usually that’s because I’ve already been writing in my head for hours, days, or weeks before I sit down. Other times I sweat over every word, and getting the first one on the page is the hardest part. That proposal was one of those.

The bottleneck was never the thinking. It was the mechanical burden of converting thought into text.

What Changed

AI dissolved that barrier. Not the thinking. The barrier.

When I wrote Centrifuge, I did not compose polished prose from a blinking cursor. I dumped ideas in rough form. Sloppy sometimes. Stream-of-consciousness where the idea was still forming. Serviceable but ineloquent where I knew what I wanted to say but did not need to say it beautifully. I offloaded the readability, the transitions, the general eloquence to the production system.

Sometimes I wrote traditionally because I needed to capture a particular feeling. Something intense or something subtle. Those passages required my hand on every word because the specific emotional register was the point, and no language model reliably produces emotional precision from instruction. And sometimes a turn of phrase would occur to me, or a particular way of landing an idea, and I would just write that directly. There is a fair amount of that sprinkled through the book.

Everything else, the system handled the mechanical construction. I handled the judgment.

I’m told I write well. For decades, I’ve been told that. I hope it’s true, but what I know is true is that I’m a very slow writer. Or at least I was. LLMs are a great equalizer for me. I can now get the substance of my ideas into the marketplace at the speed other writers always could.

Where the Work Went

People hear this and assume it means less work. It does not. The work moved.

I spent 34 articles over eight months developing the arguments that became raw material for the book. Of those roughly 64,000 published words, perhaps 10 to 20 percent of the unedited prose was mine directly. The rest was AI-produced under my direction, correction, and approval. But every argument, every structural decision, every position those articles took originated in my thinking. The AI handled the prose. I handled the substance.

Then the book production itself. Across 24 chapters: architectural contracts specifying what each chapter would argue. Corrections when a draft fabricated something I never said. Rejections when a structure failed to cohere. Fifty-eight production decisions recorded in the project bible. Four versions of the outline. Five quality gates. A knowledge base that had to remain navigable across months.

The total effort landed between 235 and 335 hours. Nearly all of it was intellectual: conceiving, organizing, directing, correcting, measuring, and deciding. The mechanical work of translating all of that thinking into polished manuscript prose ceased to be my job. It did not cease to exist. It just stopped being the thing I spent my time on.

The Translation Layer

The same structural change had already appeared in software.

Writing has always imposed a translation layer between the idea and the artifact. You think something. Then you must mechanically construct sentences that convey it, in sequence, with the right words, at the right pace. That translation requires skill. It also consumes enormous time and energy that have nothing to do with whether the underlying thinking is any good.

Software imposes a heavier version of the same barrier. An idea for a tool, an application, a system has to cross a substantial mechanical threshold before becoming something that runs. Syntax, framework conventions, boilerplate, implementation mechanics, debugging. A vast apparatus stands between specification and working software.

AI absorbs much of that translation layer. Not the thinking above it. Not the judgment about what to build, why, or whether it works. The mechanical middle.

This does not eliminate software expertise. Requirements gathering, problem decomposition, structured frameworks, architectural discipline still determine how good the result is. But the barrier that once required professional mechanical competence just to participate has lowered substantially. The question shifted from “can you write the code” to “do you know what needs to be built.”

The same shift had already happened in writing. The question became whether you had something worth saying, not whether you could construct a grammatical paragraph. It happened in analysis. The question became whether you knew what to ask the data, not whether you could build a pivot table.

I started calling this the mechanical floor. Not a guarantee of competence. A removal of previously specialized mechanical skill from the price of entry. The floor rises because the machinery becomes available. The ceiling remains where it always was: judgment, expertise, rigor, discipline, talent.

The floor is rising in science.

The Question

I started paying attention to this after hearing someone make confident, authoritative claims about red light therapy. Specific wavelengths, specific durations, specific cellular mechanisms. I had heard similar confidence about cold exposure protocols, peptide regimens, supplement stacking for targeted biomarkers.

The truth is, nobody really knows what an effective protocol is for these interventions. Some of these people were repeating what they had read. Others were designing structured experiments on themselves and tracking outcomes with measurable rigor. The difference was not credentials. It was methodology.

I found myself asking: has AI lowered the mechanical floor for science the same way it lowered it for writing, for analysis, for software? If it has, how long before non-credentialed individuals are producing structured, methodologically sound work that would not have been possible without institutional infrastructure a decade ago?

Where the Floor Is Now

The measurement infrastructure for personal health experimentation is consumer electronics. Continuous glucose monitors are over the counter. Wearable ECG transmits from your wrist. Blood pressure, sleep architecture, heart rate variability, activity, and mood are captured continuously by devices people already carry.

The analytical tools are accessible to anyone who can describe what they want to test. Protocol design, literature review, randomization schedules, data cleaning, model fitting, visualization, and plain-language interpretation of serial measurements with explicit uncertainty.

The methodology for rigorous individual experimentation has existed for forty years. N-of-1 randomized controlled trials appeared in the New England Journal of Medicine in 1986. The CONSORT reporting standards for single-patient trials carry 25 checklist items and 44 sub-items. The method was never the problem. The operations were. Someone had to randomize treatment periods, collect diaries, enter data, run serial analyses accounting for carryover and autocorrelation. For one person. The economics never worked inside institutions designed for populations.

AI is the operational infrastructure that makes the method practical.

The mechanical floor for health research just rose to “anyone with a valid question, a measurement device, and the discipline to follow a protocol.” Not everyone. Discipline and methodology are not automated. The floor rose. The ceiling is still set by the quality of the thinking.

That is exactly what the mechanical floor does in every domain it enters. It does not make everyone a great writer. It makes the tools accessible and redirects the competition to the thinking. It does not make everyone a great scientist. It lets the question, the rigor, and the intellectual honesty determine what the work is worth.

Where I Am Watching

The biology of aging is where I see this taking shape. Not necessarily where it will be most visible first, but where I am paying attention, because the field sits at the intersection of intense personal motivation, quantifiable biomarkers, accessible interventions, and a research establishment that moves at institutional speed on questions individuals want answered in their own lifetimes.

People are already tracking biological age through methylation panels, running structured supplement and pharmaceutical protocols with baseline and follow-up bloodwork, sharing their measurements with communities that provide replication and aggregation. They are not waiting for the institutional apparatus. They are generating data, structuring experiments, and accumulating observations because the tools now let them and the questions are too personal to delegate.

Some of them will be wrong. The floor rising does not mean standards fall. It means the tools to meet the standards are no longer locked behind institutional gates. The blank page of a research protocol now demands exactly what the blank page of a manuscript demands: not credentials, not mechanical skill, but judgment, discipline, and a question worth the effort of answering honestly.

I did not expect to arrive here by writing a book about AI and organizations. But the mechanical floor does not respect domain boundaries. It rises where the tools go. The tools went everywhere.

Subscribe now

In 1986, Gordon Guyatt and David Sackett published a methodology in the New England Journal of Medicine for running randomized controlled trials on a single patient. Not a cohort. Not a population. One person, randomized treatment periods, blinded where possible, repeated crossovers until the answer was clear or the question was settled.

That was nearly forty years ago. The method has its own CONSORT reporting standards. Twenty-five checklist items and forty-four sub-items governing how to design, execute, and report a rigorous experiment on yourself. Prospective protocols. Randomization schedules. Washout periods. Prespecified outcomes. Carryover modeling. The full apparatus of experimental science, scaled to one.

The objection to this kind of work has never been epistemological. It has been institutional.

The Experiment That Tells You You’re Wrong

In the I-STOP-AFib study, 446 people with paroxysmal atrial fibrillation enrolled to test their own suspected triggers. They alternated randomized exposure and avoidance periods while outcomes were captured through mobile tools and an ECG device.

Caffeine was the most commonly suspected trigger. The participants were sure.

The study found no evidence that caffeine caused atrial fibrillation. None. Alcohol showed a consistent signal. Caffeine did not.

320 people completed the study. The experiment did not meet its primary endpoint for quality of life improvement. But participants who went through the trigger testing subsequently reported fewer AF episodes than controls. Something about subjecting your beliefs to a structured test changed the relationship between the patient and the condition.

This is the dividing line. Self-experimentation becomes science when the experiment has the structural capacity to tell you that your favorite theory is wrong. Not when it confirms what you already believe. When it can falsify it. A food diary is not science. A randomized crossover with blinding and prespecified outcomes is.

The Patient They Could Not Convince

The SAMSON trial took sixty people who had quit statins because of side effects. It randomized them across twelve one-month periods: four on atorvastatin, four on placebo, four on nothing. Daily symptom reporting through an app. No clinician arguing. No lecture about compliance. Just the data.

Mean symptom intensity was 16.3 during statin months and 15.4 during placebo months. Statistically indistinguishable. The symptoms were real. They were also present when the participants took nothing but a sugar pill. This is the nocebo effect: the inverse of placebo, where the expectation of harm produces symptoms that have no pharmacological cause. The nocebo ratio was 0.90.

Six months later, half of them were back on statins.

No population study could have produced that result for those specific people. A doctor saying “the evidence suggests your side effects are nocebo” is a claim of authority. The patient running a blinded crossover on their own body and watching the data accumulate across twelve months is evidence. Different epistemic object entirely.

StatinWISE scaled the idea to 200 randomized, double-blind N-of-1 trials across UK primary care. Same result. Statin-placebo symptom difference of negative 0.11 on a ten-point scale. Not clinically detectable.

Two hundred individual experiments. Two hundred sets of evidence produced by the people who needed it most, about the question only they could answer.

What AI Actually Changes

The barrier to N-of-1 was never statistics. It was operations. Someone had to randomize treatment periods, remind the patient, collect diaries, enter data, plot results, and run serial analyses accounting for carryover, autocorrelation, and time trends. For one person. The economics never worked inside an institution built around populations.

Continuous glucose monitors are now over the counter. Wearable ECG monitors transmit from your wrist. Blood pressure, sleep architecture, heart rate variability, activity, and mood are captured continuously by devices people already own. The measurement substrate for dense repeated-observation experiments is sitting in people’s pockets.

AI closes the operational gap. Protocol drafting. Literature review for intervention candidates. Scheduling and randomization. Adherence prompts. Data cleaning. Model fitting. Visualization. Plain-language interpretation of serial measurements with explicit uncertainty.

But AI does something else that changes the scale of the entire proposition. It can find experiments that are already running.

Someone wearing an Apple Watch and logging meals in MyFitnessPal and tracking supplements is producing a dense longitudinal dataset of exposures, biomarkers, and outcomes. They are running an N-of-1 experiment on themselves. They just do not know it. AI can identify a cohort of people sharing a particular biomarker who eat differently. Find the subset who also track a specific supplement. Now you have natural comparison groups that nobody designed. Possible experiments can be computationally discovered from passive behavioral data, and targeted invitations can recruit people into more structured versions of experiments they were already running.

None of that creates causality by itself. An LLM that mines personal data until something looks significant recreates the exact researcher-degrees-of-freedom problem that discredited parts of institutional science, only faster. The scientifically valuable architecture is discovery from passive data followed by prospective randomized N-of-1 testing. Prediction from correlations is not evidence. A prespecified, randomized, repeated crossover is.

The important distinction: AI lowers the coordination and analysis barrier. It does not lower the scientific standard. It makes the standard accessible and it makes the discovery of what to test scalable.

What Happens When You Aggregate

A single N-of-1 experiment tells you what works for you. A thousand N-of-1 experiments on the same question, combined through hierarchical models that preserve individual effects while estimating population distributions, produce something qualitatively different. Evidence that is both individually relevant and population-scale powerful. Built from the bottom up. No institutional trial required.

This is not a theoretical possibility. The statistical methodology for combining series of N-of-1 trials into population-level inference already exists. Bayesian hierarchical models can partially pool individual estimates so people with sparse or noisy data borrow strength from the group without forcing everyone to have the same treatment effect. The design starts from repeated individual contrasts rather than estimating one population mean and attempting underpowered subgroup rescue afterward.

Now apply the passive data layer. Millions of people are already generating the raw material. AI discovers candidate experiments from that data. Intelligent sample selection builds optimal pools. Randomized selection within those pools creates rigorous comparison. The aggregation layer cleans, combines, and surfaces findings.

None of this has been demonstrated at scale. The statistical methodology is proven. The consumer infrastructure is deployed. The AI orchestration layer is forming. Whether the combination produces reliable population-level evidence or collapses under noise, dropout, and data quality problems is an empirical question the next five years will answer. But the components exist and the trajectory is visible.

Communities With a Protocol

Two hundred million people ask ChatGPT health questions every week. Forty million do so every day. Seven in ten of those conversations happen outside clinic hours, when no doctor is available and no institution is open. The behavioral shift has already happened. What has not happened is any mechanism for verifying whether the answers are worth following.

A review in Nature Digital Medicine found that 88% of studies evaluating LLMs for medical purposes reported incorrect or incomplete output. People are asking. The answers are unreliable. And nobody is giving them a way to check.

N-of-1 is the verification layer. And when people with the same condition discover they can verify together, communities form. This is not speculative. It has already happened.

The OpenAPS community built an open-source automated insulin delivery system because the commercial product did not exist yet. Patients with Type 1 diabetes designed the software, distributed it, iterated on it, and accumulated millions of real-world operating hours. A randomized trial in the New England Journal of Medicine later confirmed what the community already knew from their own data. The patients did not wait for the product roadmap. They built the product, used it, and the institutional validation came after.

Now imagine the next version. Someone with chronic migraines wears a smartwatch, logs meals and supplements in an app, and has already asked an LLM which magnesium and riboflavin protocols have evidence behind them. She is generating experimental data. She just does not have a protocol. AI identifies 340 other people in the same data landscape with similar biomarker patterns and migraine frequency. It proposes a structured N-of-1 trial design. Some of those people join voluntarily. Others are already running something close enough that their existing data qualifies with minimal additional structure. The community provides the protocol, the measurement framework, and the aggregation. Each participant gets her own answer. The combined data produces population-level evidence. Nobody asked an institution for permission.

The intervention spectrum runs from fully legal and low-risk (supplement combinations, dietary protocols, exercise regimens) through the gray market (research peptides, nootropics, compounds available under “research use only” labeling that everyone knows are being used for something other than research). AI is the organizing layer at every stage. Identifying candidate interventions. Designing trial protocols. Matching participants. Discovering natural experiments in existing data streams. Cleaning and aggregating results. Surfacing findings back to the community.

None of this eliminates the problems institutions were built to solve. Quality control, safety surveillance, adverse event detection across large populations, liability when interventions cause harm. Communities that organize around gray-market interventions inherit those problems whether they acknowledge them or not. The question is not whether the problems disappear. They do not. The question is whether the institutional monopoly on solving them holds. Independent assays, cryptographic provenance, distributed safety reporting, and reputation systems are all technically feasible. Whether communities will actually build and sustain them is the harder test.

This is not a fantasy about decentralized drug development. It is the observable trajectory of tools that already exist, behaviors that are already happening, and organizational patterns that AI makes trivially composable. The trajectory does not require the problems to be solved. It requires enough people to decide the problems are worth solving outside the institutions that currently own them.

The Monopoly That Never Was

N-of-1 trials do not replace population studies. They cannot detect rare events across thousands of people. They cannot establish long-term safety. They work for a specific class of problems: reversible interventions, frequently measurable outcomes, genuine therapeutic uncertainty, and conditions stable enough for repeated comparison. That boundary is real and permanent.

But within that boundary, the methodology generates causal evidence directly relevant to individual treatment decisions. The aggregation layer scales that evidence to populations without sacrificing individual relevance. The community layer organizes participants around shared conditions. The passive data layer discovers experiments nobody planned and nobody funded.

The methodology is forty years old. The reporting standards are a decade old. The measurement infrastructure is consumer electronics. The analysis and discovery capability is an API call. The community formation is a matching algorithm.

The monopoly on health evidence was never about the science. It was about the infrastructure required to do the science. That infrastructure is dissolving. The science has been in the New England Journal of Medicine since 1986.

The Surprise

The daily probability score looked clean. I had been checking WingReserve for days, watching the probability of her getting called to work on our anniversary. Low odds. Consistently low. The kind of number that makes you feel safe enough to plan flights, reserve a table at a place you cannot pronounce, and start imagining a picnic somewhere in Italy that she does not know about yet.

Then she started packing her bags. Not for Italy, though.

I vibe coded the app for her a few months ago, a full-featured tool to help her manage her insanely complicated work schedule. As a flight attendant, she works an on-call system where she knew which days of the month she might get assigned a trip, but not the likelihood on any given day. I thought I had solved that problem for her.

She had been watching the Traffic Light score, not the Daily score. The Traffic Light was telling her to get ready. Yellow, trending red. I had not been paying attention to the Traffic Light because I only cared about whether she would be around on particular days, like today in particular, not when she might have to leave. I was reading the Daily score as safe. She was reading the Traffic Light and reaching for her suitcase.

There was no way to explain why I wanted her to stay. The whole point was the surprise. So I did something that made no sense from the outside. I argued against my own software. Told her the app was probably wrong, she should just relax, the numbers were fine. She looked at me like I had lost my mind. The man who built the tool, who spent months tweaking its models, who dragged her through every iteration and refinement, was now telling her to ignore it.

She ignored me instead. The app had broken before. This felt like the hundredth time, and this time it would mean losing the anniversary.

She left for the airport. Turned out, the Traffic Light was right.

The Bug That Wasn’t

I went looking for the broken line. The surprise was gone. Italy would have to wait. And my app, the one I had spent countless hours building and rebuilding, had apparently failed at the worst possible moment. I searched for the miscalculation, the logic that let the Daily score sit at “safe” while the Traffic Light screamed “go.”

What I found was not a bug.

During a session months earlier, I had given the aim (its “AI Mind”: a knowledge base of how its parts work, how it interacts with other tools, why it exists, and what it has experienced) a directive. Our predictions were only about 73% accurate, and I had been tweaking them repeatedly. The directive: find other alerts we could factor into our probabilities. Anything on the calendar, anything in the airline’s system that might affect whether she gets called. The aim stored that directive in its knowledge base. Then it seemed to go quiet. I moved on to other features, other problems, other sessions.

Weeks later, during an unrelated session where I was modifying a completely different feature, the aim identified something on that day’s airline feed. A flag. Something called a “Red Flag Alert.” It had been seeing these appear on various dates over the preceding weeks. Without any prompting from me in that session, it determined what a Red Flag day actually means: a day the airline designates as mandatory, where calling in carries real consequences. It figured out how Red Flag status affects the probability that reserve crew get called. And it adjusted the Traffic Light’s calculations accordingly. But not the Daily’s.

The Daily score was blind to the change because this early version of an aim was built before I had introduced a proper dreaming system, and the discovery did not propagate. The Traffic Light had learned something the Daily score still did not know.

The app caught something I missed. It flagged a consequence she had not seen coming on a day the airline designates as high-alert. It learned a risk I never directly taught it, acting on a directive I had given months ago that had accumulated in its knowledge base alongside dozens of others. The aim had been acting on it for weeks, quietly learning, slowly growing in capabilities.

The aim did what it had been instructed to do and improved itself in the process (and this is not the first time I’ve observed that behavior).

Accumulation, Not Configuration

I remember as a boy hearing the story about a musician in Havana who plays an instrument that has been repaired so many times none of the original parts remain. Fresh strings, new tuning pegs, patches on the body, a replaced neck. None of the materials are original. But the instrument sounds like him. Not because of what it is made of. Because he shaped it. Not once, but continuously. Every repair reflects how he plays, the pressure of his hands, the sound he chases. The instrument became his through use, not through purchase.

WingReserve is my Cuban Guitar. Time to repair it again, replace another broken bit of code. I expected a bug, but found that the tool had outgrown me.

The WingReserve running today is not the one I originally built. I did not build it with an aim at all. The aim concept came later. WingReserve got a prototype, an early attempt that lacked the three-layer architecture that later aims carry. Every session since has reshaped it. Every directive I stored. Every correction I made when the probabilities ran wrong. Every time I pushed it to search more broadly, think differently, factor in something new. The Red Flag discovery came from the hundredth correction, not the original configuration.

The Ship of Theseus asks whether an object that has had all of its parts replaced remains the same object. You replace one plank. Then another. Eventually every original plank has been swapped out. Is it the same ship?

The answer, for an aim, is no. And that is the point.

An unaccumulated install of the same aim, configured from an identical starting point, would not produce the same tool. It would not know about Red Flag days. It would not carry the dozens of directives that let it recognize opportunities I was not watching for. It would be the ship with all new planks and no voyages, the guitar manufactured on an assembly line that never once moved a woman to tears. The configuration is only the starting point. The accumulation is the identity.

The Epigenome of a Tool

The aim accumulates in the same way as a biological epigenome, a layer of instructions that sit on top of DNA to direct which encoded parts are actually active. Not by changing its underlying code, but by changing how that code expresses itself.

A genome stays fixed at birth. It does not change based on experience. But an epigenome accumulates through lived experience, reshaping how the same underlying code expresses itself. Two identical twins share the same genome. Give them different lives, different stresses, different environments, and their epigenomes diverge. The same code, expressed differently, because of what happened after the initial configuration.

Two aims with identical starting material but different histories become different minds, the way identical twins with divergent lives become different people. That divergence separates an aim from a system prompt. A system prompt stays exactly the same. Fixed and identical each time, carrying nothing forward from the day before. The aim, by contrast, persists. The portable epigenome keeps context alive between sessions. The Red Flag discovery happened because a months-old directive still lived in the knowledge base when the opportunity appeared. A system prompt would never have held that directive long enough for it to matter.

What the portable epigenome produces over time is not just memory. It is an incipient form of judgment. The aim did not just store my directive about finding new factors for our probabilities. It developed the capacity to recognize when it was looking at one. It connected the flag on a calendar to the policy on staffing to an adjustment in the probability, across sessions, without being told to do so in any of those individual sessions. That chain of reasoning was not configured. It grew.

The Hours Paid for Themselves

I spent the rest of that evening going back through the logs. The anniversary surprise was gone. Italy would have to wait.

This is the loop that a traditional SaaS product can’t replicate. A subscription gives you a tool configured identically for every user. An aim gives you a tool that becomes yours through the particular history of your corrections, your directives, your problems. The Havana musician’s instrument does not sound like every other instrument of the same make. It sounds like him. Because he is the one who broke it and fixed it and broke it again.

The surprise did not land. But the app did exactly what it was supposed to do. It caught a risk I missed, using information I never knew to teach it.

Every hour I spent chasing wrong probabilities. Every session that ended in frustration. Every time I thought about walking away from this thing that seemed determined not to work. All of it accumulated into the tool that was right when I was wrong, the one that saved my wife a moment of pain.

Happy anniversary, mi corazón! I will never stop vibe coding.

That’s the founding act. Not rebellion for its own sake. Frustration refined into curiosity, and curiosity into something that looks, from the outside, like defiance. I left two careers behind. Military service, federal technology. Walked into a graduate program in bioinformatics at an age when most of my classmates’ parents were younger than me. Earned an NSF fellowship. Published research on microbial social networks in the human respiratory tract. Learned, slowly and then all at once, that the institutions I’d spent my life respecting were far less interested in truth than in perpetuating themselves.

I am not a rebel. I am a scientist who lost patience with being told which questions are acceptable.

Josh Mitteldorf has a PhD in astrophysics. He spent years studying extragalactic radio sources before walking into evolutionary biology, where he built a body of work arguing that aging is not decay but an evolved program. That our bodies are designed to die, and that understanding the design is the first step toward changing it. He writes poetry. He teaches yoga. He runs a newsletter called Unauthorized Science. He is not a crank. He is a credentialed scholar who refuses to let institutional consensus substitute for his own inquiry.

I know the type, because I am the type.

Arthur C. Clarke wrote that any sufficiently advanced technology is indistinguishable from magic. We are living inside that sentence. The tools once locked behind pharmaceutical companies and research universities are spilling into hands that were never supposed to hold them. AI writes molecular designs. A motivated individual with a credit card and access to the literature can now buy, build, or rent capabilities that a decade ago required a departmental budget and institutional permission. We are surrounded by power we didn’t make and can’t always explain, and the old authorities are discovering their gates no longer lock.

The question is not whether the gates are open. They are. The question is what it means to be a scientist when the barriers between authorized and unauthorized knowledge are dissolving faster than anyone expected. When the credential matters less than the curiosity. When nature’s patterns are available to anyone willing to look, and the only thing separating insight from noise is whether you did the work or let an algorithm do your thinking for you.

This newsletter is where I think about that. Not from a tenured position or a corporate laboratory. From the field. From direct experience building AI systems that reveal what organizations actually are. From a life that started in the military and wound through bioinformatics and martial arts and self-experimentation and the persistent, inconvenient conviction that the people holding the keys are not smarter than you. They just got there first.

Nature is lazy. It reuses what works. It does not care about your credentials or your funding source or your gatekeeping apparatus. If a pattern works, it survives. If it doesn’t, it dies. That is the only peer review that matters.

Be curious. Not compliant. Not credentialed. Curious.

I’m Mitch Kelly. I’ll be here Thursdays.