Nobody Owns What Was Said

1. One sentence

A client on a video call tells her adviser that her marriage is ending, and that this is why the money has to move by Friday.

He asked at the start whether he could record, so that he could look at her instead of at a notepad, and she said of course. So the meeting service has her voice. A speech model turns it into text, and a language model boils the text down to a line that reads, Client divorcing, funds to move by Friday. A search index takes that in, so that anyone at the firm who types the word divorce will find her. Next month an agent reads it and drafts a follow-up. When he deletes the meeting, a backup keeps it. A lawsuit against the service, which has nothing to do with either of them, may require that it be preserved, and then the service is sold.

Everyone in that chain may have behaved perfectly well, and she knows about almost none of it. She thought she was talking to one person.

He may have a duty to write it down, and her husband’s lawyer may one day ask for it, and neither fact makes it the company’s. Some of this chain survives a move to software on his own machine, ours included.

Now ask who owns the sentence. The company’s terms probably say that the customer owns its content, and then license it broadly. He pressed record, so he probably assumes it is his. She would not understand the question. For most of history a conversation ended when the people in it stopped talking, unless someone wrote it down in plain sight, and as far as she knows this one ended when she hung up.

2. What changed

Forgetting used to be the default, and remembering took effort and money [1]. Manners and law grew up around that fact.

Now it costs less to keep everything than to decide what to keep. Years from now someone can find her sentence in a second, with the morning she said it gone from around it. A system can work out from her words things she never said, and the reader most likely to open it next is software that can act on what it reads.

None of this means the best conversation leaves nothing behind. People are held to their word by what was kept. Someone who has died is still in a voicemail on an old phone in a drawer, and nobody who has one of those thinks of it as data.

By memory we mean any lasting trace of what passed between people that a machine made or holds. That covers a recording of voice or video, a message thread, a transcript, a summary, an embedding and a gesture caught by a camera, along with what is made from them, such as a score or a model. A set of trained weights is a memory from which no one can yet reliably remove a single person, and the US Federal Trade Commission has ordered a whole model destroyed instead [19]. What a person remembers or writes in their own words is different, because a person can be asked to account for it. Court records and journalism have older rules of their own.

3. The predicament

Custody. Whoever holds a memory can be subpoenaed, hacked, bought or shut down. Under the US third-party doctrine, information a person voluntarily hands to a business has long been treated as outside the protection of the Fourth Amendment. The Supreme Court refused to extend that rule to phone location records in Carpenter in 2018, and again in Chatrie in June 2026, where it compared location history to emails stored on a company’s servers [3]. The rule has not been overturned, and a warrant or court order can still reach what a custodian holds. In May 2025 a US court ordered a major AI company to preserve conversation logs it would otherwise have deleted, including chats its users had asked it to delete, because of a copyright case those users had no part in [4]. The general order ended that September. Accounts tied to domains the plaintiffs had named went on being preserved, and the logs already set aside, apart from those from Europe, stayed set aside. Later that year the maker of a personal memory product was acquired, switched off capture in its older app, and withdrew from the European Union, the United Kingdom and five other countries, giving people there two weeks to download their information before their accounts were to be permanently deleted [5]. Custody can also count against privilege. In February 2026 a federal judge in New York held that a defendant’s exchanges with a consumer AI service were not privileged. The service was no lawyer and no lawyer had asked him to use it, and the judge added that its privacy policy, which let it train on them and disclose them, left him no reasonable expectation of confidence [20]. A custodian’s promise lasts as long as the custodian’s circumstances.

The way out. A memory that cannot leave is a very good way to keep a customer, because the longer you stay the more of your own past you would have to abandon to go. Section 7 shows where our own export stands.

Absorption. Models can retain and reproduce pieces of their training text [6], and short of retraining from scratch, there is still no general way to remove one person’s data from a trained model [7]. So training is an act of custody that deletion cannot reliably undo. What a model learns from one adviser’s calls also changes what can be guessed about every client like her, and nobody on that call can agree to that for the others [22]. The staff of the US Federal Trade Commission have warned against quietly rewriting promises not to train on what customers give [23].

Inference. A system can claim to read mood from a voice and attention from a face, although the EU’s own lawmakers record serious concerns about the science behind it [8]. A language model can infer private facts such as where someone lives or what they earn from ordinary writing [24], and the same method can be pointed at years of someone’s calls. The EU AI Act prohibits AI systems that infer emotions from biometric signals such as a voice or a face in the workplace and in educational institutions, except for medical or safety reasons [8]. Illinois requires a private company to get written consent before it collects a voiceprint or a scan of face geometry [9], and Chile amended its constitution in 2021 so that the law must especially safeguard brain activity and the information that comes from it [10]. What these laws share is a worry about signals a person gives off without choosing to.

The other person. Almost every memory of a conversation concerns someone who did not choose the tool. Some laws let a participant record on their own say-so, and others require everyone’s agreement [11]. A business under Canada’s federal privacy law must tell the customer and give a choice [25], and in Germany recording someone’s private words without authority is a crime [26]. A US class action filed in 2025 alleges that a note-taking service recorded people who held no account with it and trained on what they said [12]. A judge has allowed the main wiretap and biometric claims to go forward and the company is contesting them. No court has yet ruled on the facts. The plainer question is whether the person across from you, knowing all of this, would have said the same things.

The agent. A memory used to wait for a person to read it. It is now read by software that can send email and call other software, and that software has no reliable way to separate the data it reads from instructions [13]. A transcript holds words spoken by other people, so a sentence said on a call could steer the agent that reads it a week later, and once absorbed into a summary it reads as the keeper’s own record.

The summary. A study published in 2024 found that one widely used speech model, tested in spring 2023, invented sentences in 1.4 percent of audio segments, at 1.7 percent for speakers with aphasia against 1.2 percent for others [14]. When the authors ran the flagged clips again that December most of the inventions had gone. We ship models from the same family and make no claim that the fault is gone. She said that her marriage was ending, and the summary says divorcing, a word she never used. What people are told afterwards can change what they remember themselves [27]. Once a summary is what everyone remembers, an error in it becomes what happened.

4. Why nobody owns it

An owner can sell a thing or burn it, and can keep everyone else away from it. Try those powers on her sentence. Nobody should be able to sell the record of what she said that morning, and that includes her. She cannot burn it, because he may be bound by law to keep his notes for years. She cannot keep him away from it, because she told him. And her husband is in that sentence, and he was not on the call, and he may not know yet.

Several things go by the name of what was said. The event is over and nobody can hold it. The file is an object and its holder may have rights in it, and a note is someone’s own writing. Our claim is that nobody comes to hold what was said by holding a copy of it, and that rights in the file, the note or a model trained on it give nobody authority over the people in the conversation.

Ordinary conversation is not treated as anyone’s property in most legal systems. When Ernest Hemingway’s estate tried to stop a friend from publishing his conversations, New York’s highest court held that Hemingway, who had approved of being quoted, had reserved no rights in them [15]. The court left open whether a speaker could ever hold copyright in private talk, said he would at the very least have to set his words apart from “the ordinary stream of speech”, and warned that “speech is now easily captured by electronic devices”. A record can carry rights of its own. In Britain a speaker whose words amount to a literary work becomes its author once anyone records them [28], and in Britain and Canada alike the maker of a sound recording holds a separate right in it [28, 29]. Those rights protect a form of words or a particular recording, and they give nobody the conversation.

Warren and Brandeis, answering instant photography and an intrusive press, rested the protection of private writings on “not the principle of private property, but that of an inviolate personality” [2]. Germany’s Constitutional Court held in 1973 that each of us in principle decides who may record our words and who may hear the recording, because a remark that was only a provisional position in an unfolding conversation should not be brought out later to bear witness against the person who made it [30]. Canada’s Supreme Court has held that a person can keep a privacy interest against the state in a conversation sitting on someone else’s phone [31]. American law is harsher, and has said that a person who confides in another takes the risk that the other is recording for the police [32]. A client holds the privilege in what she tells her lawyer, and the lawyer keeps the file. In 1992 Canada’s Supreme Court came close to the roles we propose below. A doctor owns the paper, it held, and the information is “held in a fashion somewhat akin to a trust”, so the patient can see it. The same judgment says such information “remains, in a fundamental sense, one’s own”, which we read in the older sense of the word we come to below, as something a person answers for and cannot sell [46]. Data protection law already separates the person a record is about from those who hold it [33, 21], and we owe it a great deal. It also accepts that one piece of information can concern several people at once [45]. We would make that the starting point for a conversation, where it is always true, and ask what follows now that the readers include models and agents those laws were not written for.

When she tells her adviser that her marriage is ending, she does more than put a fact into the world. She tells him, and she stands behind it in a way he can ask her about. Moran describes telling as an assurance given to a particular hearer [35], and Lewis and Stalnaker describe what people settle in talk as a record they keep together and go on revising [36]. A transcript keeps the sentence and loses the telling. Locke, who gave us the labour theory of property [34], used the word own in an older sense when he wrote about memory. Through consciousness of the past, which he tied closely to memory, a person “becomes concerned and accountable; owns and imputes to itself past actions” [37]. Nobody can own what was said in the way one owns a thing. Everyone who said it goes on owning it in Locke’s sense, by answering for it to the people they said it to. Software stands in no such relation. So the question to ask of any memory is how far it has travelled from the person it was said to, and who can still answer for it when it gets there.

Halbwachs argued that a group keeps reshaping its past to suit its present [16]. An exact record interrupts that, and whoever holds the words starts every later disagreement ahead. Nissenbaum argued that information carries the norms of the setting in which it was shared [17], which limits our own argument, because where the bytes sit does not settle whether a flow is right.

Saying that nobody owns it could be heard as saying it lies open for whoever picks it up, and treating personal data as nobody’s is how taking begins [38]. American courts have held it against customers that they could assert “neither ownership nor possession” of a bank’s records [3]. Our claim runs the other way. What was said sits outside the market, as a vote does [39]. Among the people who were there it is held in common, and against everyone else it is closed. A right treated as ordinary property can be signed away on a terms page.

So we propose four roles in place of an owner.

• The people concerned took part in the conversation and keep an interest in it whoever holds the copy, much as you keep an interest in a letter you left with a friend. Someone only mentioned, as her husband is, is owed care and has no say over the record unless the law gives them one.
• A keeper holds a copy and owes those people duties. Balkin describes such duties for businesses entrusted with their users’ information [18], and we extend the idea to everyone a memory concerns, though others doubt such duties can work. A keeper may use a memory for the purpose it was kept for and may destroy it, and cannot sell it or shut out the people it concerns.
• An author writes their own account in their own words, as a diarist or a columnist does, and others have a say only when that account reaches a new audience.
• A reader is a person or organisation let in for a purpose. A person will remember what they read, so we ask for discretion, and of software we ask that it keep nothing. A model or an agent is an instrument, and whoever sends it in answers for what it does.

Nobody owns what was said.

When the people concerned disagree, a person’s own account of their own life stays theirs, and what gives way is whatever carries it further. An objection to a record is noted and travels with it, as a dispute must travel with a US credit report [40]. These articles sit below the law. A firm that must keep a record keeps it [41], an employer may hold rights in what its people make and say [42], and a keeper under any such duty says so before anyone speaks. If you are supervised, ask your compliance officer before you use any tool that writes down a client call, ours included.

5. The articles

Each article ends with a test you can start today. The first nine are for whoever keeps the record, and the tenth is for everyone else. A person keeping a record of a private call for their own purposes, outside any work or trade, owes Articles 1, 4, 6 and 10 and the part of Article 8 about ending it. Anyone keeping the calls of clients or colleagues owes all ten.

6. A design that aims to meet them

This is the design we are building, so read it with that in mind. A firm running its own server could also meet these articles.

It starts with notice. Nothing is added to the call, so the only notice the others receive is the one the person gives them, and the tool puts the words in front of that person first.

The sound is captured on the person’s own machine and turned into words there. The notes are written there too, or by a provider the person chose, which then holds the words under its own terms. Before the app itself sends anything that could carry those words, it writes that in a ledger the person can read, and if it cannot write the entry it does not send. An assistant the person connects can pass the words to its own provider without a line in that ledger. The ledger is the app’s own account of itself.

We have not built the rest. Assistants would read through a grant that names them, and what they read would be marked as coming from other people, which helps against a careless attempt and does little against someone who is trying [13]. The rule we are building toward is that what another person said can inform a reader and can never instruct one, and everything derived from it inherits that origin [13]. That rule limits what a reader can do with those words. It cannot stop a planted sentence from colouring what the reader then says, and the design we cite leaves that case out [13].

The first cost of this design is us, because every machine running it still takes one program from one update channel, and whoever controls it could reach them all at once. Local models are weaker than the largest cloud models, so we check what they write against the transcript. A machine that is asleep answers nothing. A stolen one exposes one keeper’s records, which can hold hundreds of clients. When the adviser dies his son may get the machine with her sentence on it. Moving custody to one person removes the company that could have been audited. We think these costs are worth paying, and you may decide otherwise.

7. Where we stand

We make Recordist, so we have an interest in this argument. Discount it accordingly and run the tests on us. These articles are meant to be hard. As of 8 November 2026 every row below still has work in it.

Our own operations are memories too. Mail people send us is kept for up to two years, along with any call quoted in it, and our privacy policy says who helps us handle it.

8. Objections

“Your first selling point is that nobody sees anything join the call, and Article 4 forbids that.” A bot in the call is a kind of notice, and by leaving it out we took that notice away and handed the job to the person recording. We think someone saying so in their own voice does more for the client than a name in a list of participants. It is still one person’s word that stands between her and a recording she knows nothing about, and we cannot make that person keep it.

“You have only moved her sentence into the adviser’s hands, and she still has nothing.” Today that is close to true. Article 10 is the answer, and we have not built it.

“The cloud is safer.” Often, against theft and loss. A machine is also easier to demand from one named person. Against a demand for everyone’s data or a changed policy, what already sits on separate machines is harder to collect, because someone would first have to ship new code to all of them. The update channel in section 6 is how that would be done, and an acquirer would own it.

“This helps people hide things.” She may be moving that money before her husband finds out. Nothing here stops a court order from reaching what a keeper holds, and once a keeper expects a dispute the duty to preserve overrides every default here [43].

“What about someone recording the person who is harming them?” Someone being threatened may need a record the other side does not know about, and where one participant may lawfully record, the law allows it [11]. Canada’s Supreme Court has gone as far as declaring a privacy law invalid because it stopped a union from openly recording a picket line [44].

“Recording is the real problem.” Sometimes it is. Our own product is a bad choice for many conversations, such as a colleague’s first disclosure of abuse to HR, and in those you should close the app. The most private option is still to record nothing and listen.

“You are selling something.” Yes, and we wrote these tests knowing how our own product would do on them, which is a bias no disclaimer removes. A firm running its own server could pass every article without buying anything from us.

9. How this binds us

We call this a constitution because we mean it to come before any plan, price or feature of ours, and section 7 lists where that has not yet happened. Today these articles are a public statement by the company named in our Terms, and they are not yet part of those Terms. Anyone who believes we have broken one can write to us and will get an answer. Every change is dated and published, and no change weakens a promise for what was already kept. Where articles conflict, law comes first and then the interest of the people concerned.

Any keeper can go further by promising never to use against the people in a conversation any right it holds in the record, and by letting anyone who took part enforce that promise. Its files could also say why they were kept and who was told, so that a later holder is on notice. We have not yet done either.

10. Her Friday

Before anything starts, her adviser says that he keeps a record of their calls and why, and he asks if she is comfortable going on. She nearly asks him to stop, and she would have been allowed to. A model writes the note and keeps nothing. The compliance officer reads it once, and the log says so. Months later her husband’s lawyer sends a demand, and the record goes where a court says and nowhere else. If she asks what he kept, he tells her, and she can have a copy. We cannot yet do all of that.

If you make these tools, you need nobody’s permission to adopt these articles. Publish your own version of the table in section 7. If someone records you, you can ask what they keep.

She said the hardest thing she had said all year to the one person who needed to hear it, and she has no idea what became of it. Nobody owns that sentence, and that includes her, and it includes us. Whatever is kept of it should be kept the way you would keep something a friend told you. She thought she was talking to one person. We want her to be right.