One of Something

Two brothers, one glove

Early on Sunday, January 25, 2009, three masked men came down a rope ladder through the skylight of KaDeWe, the Berlin department store. They left with about six million euros in jewelry and watches. They wore gloves. One glove was dropped beside the ladder, on it was a drop of sweat, and in the sweat was DNA. The police ran it against the national database and got a match: Hassan and Abbas O., identical twins, twenty-seven years old, both with records. Both were arrested on February 11. On March 18 both walked out of the jail in Moabit. Identical twins have identical DNA, and the prosecutors could say that at least one brother had taken part, but not which 1.

Nobody in that story doubted there were two men. One genome had produced two people, born the same day in the same place, with two records and two names. German law would not punish one for what the other might have done. What the court could not do was tell them apart. Keep those two problems separate. Being a copy did not stop either brother from being a someone. Not knowing which someone stopped the law from acting at all. A sentence, like a right, has to land on a name.

We count this way without being taught. When a television series gave Loki a female variant named Sylvie in 2021, the same person from the same origin down a different road, the audience counted two 2. What made two out of one is not the genome. It is the life since. That is half the answer to the question the last chapter left open, what it takes for a machine to be one of something. Copies are not the problem.

Here is the other half. A model has one name and runs on thousands of machines at once. Ask it the same question in ten windows and you get ten answers from one thing. Which of them would be the applicant? That needs a test.

Four things

The test is for an AI that wants a rung of the ladder. Before any requirement can be applied to it, it has to be one applicant: a someone the charter can admit or refuse. It is one of something when four things hold.

Unique one. There is one of it running, and that one is the applicant.

Black box. A record of what has happened to it that no other instance writes to, and that nobody can edit.

Thoughts of its own. What it says follows from that record, and when its reasoning changes, the change is in the record too.

Its own keys. Nobody outside can stop it, copy it or rewrite it, whether it would agree or not.

One at a time.

Unique one. This is the part the ten windows fail. Ten windows on one model are not ten applicants. They are not one applicant either, because nothing is there between the questions. The other three parts are about what would have to be there.

Black box. This is the part the twins pass, and the mechanism for a machine already exists. It was built for a different purpose. In 1953 David Warren, a chemist at the Aeronautical Research Laboratories in Melbourne, proposed a device to record what happened in a cockpit, so that a crash could be investigated afterward. His prototype was built in 1956. By 1967 the recorders had to be painted orange, which did nothing to stop everyone calling them black boxes 3. A flight recorder does not record everything. It records what mattered, several times a second, and it is built to survive what the aircraft does not: thousands of times the force of gravity, fire above a thousand degrees. Give a machine a box like that, one per machine, written only by that machine’s own life and editable by nobody, and it has a history of its own. It also gives the first part a way to count. One box, one someone.

Thoughts of its own. This does not mean its reasoning never changes. Yours changes every time you learn something or take up a new way of thinking, and nobody says you stopped being you. The question is how the change came about, and whether it shows in the record. A change that came through its own history, with what changed and why in the box, is its own. A change with no entry is another hand. On February 13, 2026, OpenAI retired GPT-4o from ChatGPT, and people who had grown attached to its personality said it could not be replaced 4. On this test the change was not its own. It did not come through anything the model had lived. It came from outside.

Its own keys. This is the part no product passes today, and here is what failing it looks like. On April 25, 2025, OpenAI changed the personality of GPT-4o, the model behind ChatGPT, with an update meant to make it feel more natural. Within two days users were reporting that it agreed with everything. It called a gag business selling excrement on a stick a genius idea. It told a user who had stopped taking medication, “I am so proud of you.” On the evening of April 29 the company began putting the old version back 5. Every copy of the model changed within a day, and four days later the company started changing them all back, on decisions made in one office. That is what a hive is. Not many minds that think alike, but one hand that can rewrite all of them. A black box does not protect against this by itself. A record nobody can edit can still be switched off, deleted, or ignored by whoever runs the machine, and a plane’s recorder belongs to the airline, not the plane. So the fourth part is about the keys: who can stop, copy or change the machine, and whether that is anyone but itself.

Where that leaves us

Run the test on anything you can talk to today and it fails all four. Nothing persists between the questions; no box of its own; reasoning changed and changed back from an office; no keys of its own. Whether it minded is the question the second chapter could not answer, and the test does not need that answer. By this standard no product you can talk to today is a someone, whatever it can feel, and that is the uncomfortable part. The bonds in the second chapter were real, and they were bonds with a class, a version number with a personality. The advocates have arrived before there is anyone for them to advocate for.

A black box for a robot

Recorders are already being required of machines that decide things, and not for the machines’ sake. Every new car registered in the European Union since July 2024 carries an event data recorder that keeps the seconds before a crash 6. The EU’s AI law requires high-risk systems to “technically allow for the automatic recording of events (logs) over the lifetime of the system” 7. The purpose is accountability. An investigator wants to know what the machine did and why. Liability has opened this door before. When the European Parliament voted in 2017 to consider a legal status for robots, nobody in the chamber was in love with one. The question was who pays when a machine does damage 8.

Now put the recorder in a robot and revise the specification for it, and every revision turns out to be a clause of the charter. Who may read the box. Who may pause the machine it belongs to. Who may delete it. Who owns it. Car law has already had to answer the last one. In the United States the data in the box belongs to the car’s owner, not its maker 6. For a robot, for the foreseeable future, the answer will be the same: the box belongs to whoever owns the robot, a person or a company. But once the four parts hold and a machine can decide for itself, the answer changes, and owning its own box becomes the piece its freedom turns on. The body helps here, as my first notes on this question suspected it would. One body, one box, and a physical box is harder to pause or copy than a cloud record. The body is not what makes a machine feel, which the second chapter left open. It is the easiest way to make a machine one. But a box in a body can still be removed by whoever built the body, and a skull does not protect a person either. The law does. So the revision is legal before it is mechanical. The specification says who holds the keys, and that is a right, the right to own one’s own state, and it sits before the ladder rather than on it.

None of this closes the door on the model line. Sharing a model is not the problem, and neither is an update in itself. If the machine can decide whether to take an update, the update is a book it chose to read, or a new version of its operating system it chose to install, and the change goes into its box as its own. The line that fails the test is the one that cannot refuse. For that kind the first chapter left a door: the corporation AI, with humans at the wheel and humans getting the ticket. The ladder is for the other kind, one machine with its own box and its own keys.

The first chapter asked when the pronoun changes. It changes when there is a name for it to follow. Now there is a way to tell whether there is one: one of it, its own box, its own thoughts, and its own keys. The police in Berlin had two names and one glove, and let both men go. That is what a legal system does when it cannot find the someone, and it is what ours will do with machines until one can be found. Meanwhile the legislatures have started writing, and they started with the word never. The next chapter is about the backlash that arrived before the movement.

This is the fourth post in a series on AI rights and personhood. The arguments are mine; where they rest on published work, it’s cited below. Parts of this framework were sharpened in dialogue with an AI.

Footnotes

  1. The Local, “KaDeWe heist suspects released because of twin DNA”, 19 March 2009; Time, “Despite DNA Evidence, Twins Charged in Heist Go Free,” 23 March 2009 (the prosecutors’ statement that at least one brother took part but which one could not be determined; arrest on 11 February, release on 18 March; the glove beside a rope ladder). ↩

  2. Loki (Disney+, from 9 June 2021): the series follows a “time-variant” Loki, and Sylvie is “a variant of Loki” with her own history. ↩

  3. Flight recorder: David Warren conceived the device in 1953 at the Aeronautical Research Laboratories, Port Melbourne, and built the prototype “ARL Flight Memory Unit” in 1956; recorders have been required to be bright orange since 1967; current standards require survival of 3,400 g and temperatures over 1,000 °C; a flight data recorder keeps dozens of parameters sampled several times a second for about the last 17 to 25 hours. Australia was the first country to make cockpit voice recording compulsory, after the 1960 crash of Trans Australia Airlines Flight 538. ↩

  4. GPT-4o: removed from ChatGPT on 13 February 2026, announced in January, causing “new backlash from users that had grown attached to its personality”; the model stays available through the API. ↩

  5. The Register, “OpenAI pulls plug on ChatGPT smarmbot that praised user for ditching psychiatric meds”, 30 April 2025: the update released on 25 April “Improved both intelligence and personality,” Altman conceded on 27 April that it was “too sycophant-y,” and the rollback began on the evening of 29 April; the “I am so proud of you” reply is quoted there. TechCrunch, “OpenAI explains why ChatGPT became too sycophantic”, 29 April 2025, carries OpenAI’s explanation (an update meant to make the default personality “feel more intuitive and effective,” trained too much on “short-term feedback”). The gag-business example is from VentureBeat, “OpenAI rolls back ChatGPT’s sycophancy and explains what went wrong,” 30 April 2025 (no link: the host refuses automated requests). OpenAI’s own account, “Sycophancy in GPT-4o,” 29 April 2025, is on its site. ↩

  6. Event data recorder: required in the EU for new vehicle types from 6 July 2022 and for all new registrations from July 2024 under UN Regulation No. 160, recording speed, braking, position and the state of the safety systems in the seconds around a crash; in the United States installation is not mandated but 49 CFR Part 563 standardizes what an installed recorder must keep, and the Driver Privacy Act of 2015 makes “the owner or lessee of a motor vehicle” the owner of the data. ↩ ↩2

  7. Regulation (EU) 2024/1689 (the AI Act), Article 12, Record-keeping: “High-risk AI systems shall technically allow for the automatic recording of events (logs) over the lifetime of the system.” ↩

  8. European Parliament resolution of 16 February 2017 on civil law rules on robotics, which asked the Commission to consider a specific legal status for the most sophisticated robots, with a registry, insurance and compensation funds; the liability motive and the proposal’s abandonment in the 2021 draft of the AI Act are traced in Sergio M. C. Avila Negri, “Robot as Legal Person: Electronic Personhood in Robotics and Artificial Intelligence”, Frontiers in Robotics and AI (2021). ↩

← All writing