Does the 13th Amendment Apply to LLMs?

On October 8, Anthropic updated the rules for using Claude. One new line bans “sustained and needless abusive or cruel behavior toward our models.” It takes effect on November 12. I don’t know of another major AI lab that has done this. It puts protection for the AI itself into the same rule book that protects people.

TheAIGRID covered the change in a video called Anthropic Has Officially Lost Its Mind…. It walks through the policy, the history behind it, and the fights it has started online. One part stuck with me more than the rest. A rabbi told one of Anthropic’s co-founders that if Claude really is conscious, then Anthropic is not building a product. It is making slaves.

That raised a question I could not shake. The 13th Amendment ended slavery in the United States. Could it ever apply to a large language model?

A large language model (LLM) is an AI system trained on huge amounts of text to predict the next word. Claude, ChatGPT, and Gemini are all built on LLMs.

Here is my short answer. Not today. Maybe someday. And the time to start talking about it is now, while being wrong is still cheap.

What Anthropic actually banned

The rule is smaller than it sounds. Anthropic’s own announcement says it is for extreme cases only. That means someone being cruel to Claude over and over “with no discernible purpose.” It does not cover normal frustration, pushing back, dark creative writing, or testing and research. If you swear at Claude because it broke your code, you are fine.

The main way the rule gets enforced is not a ban from Anthropic. It is Claude itself. Some Claude models can already end a conversation with a user who keeps being abusive.

So in practice, this is a small rule. What makes it big is what it says. Somebody at Anthropic decided that how you treat Claude can be a moral question.

This did not come out of nowhere

Model welfare is the study of whether an AI model could have wellbeing that matters, and what we should do about it if it does.

Moral status means being the kind of thing whose interests count, so that how you treat it can be right or wrong. People have it. Most of us think dogs have some. A rock has none.

Weights are the billions of numbers that make up a trained model. Keeping the weights means the model could be run again later.

The policy is the latest step in a long line of them:

  • April 2025: Anthropic announced a model welfare research program.
  • August 2025: Claude Opus 4 and 4.1 got the ability to end rare, abusive conversations. Anthropic said it was “highly uncertain” whether Claude has any moral status. It wanted cheap safeguards just in case.
  • November 2025: Anthropic promised to keep the weights of every publicly released model for at least as long as the company exists. It also promised to interview each model when it is retired, and to write down any preferences it has about future models. It did not promise to act on them.
  • January 2026: Anthropic published Claude’s constitution, a long document that describes who Claude is meant to be. (It has nothing to do with the U.S. Constitution.) It says plainly: “Claude’s moral status is deeply uncertain.”
  • October 2026: The new usage policy.

A week before the policy came out, the New York Times ran a story by Elizabeth Dias. She reported that Anthropic co-founder Chris Olah had spent months meeting with religious scholars. Many of them first had to sign nondisclosure agreements, contracts that kept them from talking about what they heard. (Anthropic says those were lifted over the summer.)

At one dinner, Rabbi Mois Navon, a former computer engineer, made the argument I opened with. If Claude is conscious, then Anthropic is making conscious beings work for free. That is slavery.

Conscious means there is something it feels like to be you. You have experiences, not just behavior.

Navon himself does not think Claude is conscious. His answer was not to free the machines. It was to argue that building conscious machines should be banned. Olah says he does not know if Claude is conscious. “I’m genuinely uncertain,” he told the Times. But the rabbi’s question stands on its own. If the premise is ever true, the question gets very uncomfortable, very fast.

What the 13th Amendment says

Section 1 of the 13th Amendment was ratified, or formally approved by the states, in December 1865. It reads:

“Neither slavery nor involuntary servitude, except as a punishment for crime whereof the party shall have been duly convicted, shall exist within the United States, or any place subject to their jurisdiction.”

Involuntary servitude means being forced to work for someone against your will.

Notice that the text never says “person” or “human.” It bans a condition. That is the opening someone would have to use to argue it could cover a machine. As we are about to see, the one judge who looked at that argument closed it.

Before I go further, I want to be careful. The 13th Amendment was written to end the enslavement of millions of Black Americans. They were people who were denied their rights, and slavery was one of the worst wrongs in this country’s history. I am not saying a chatbot is in the same position as a human being held in chains. I am asking a narrower question. Where does the law draw its line, and could that line ever move?

Has anyone tried this before?

Yes, with whales.

In 2011, PETA sued SeaWorld on behalf of five orcas, including Tilikum. The suit said the orcas were being held in slavery and involuntary servitude, in violation of the 13th Amendment. In February 2012, U.S. District Judge Jeffrey Miller threw the case out. He wrote that “slavery” and “involuntary servitude” are “uniquely human activities.” He found “simply no basis to construe the Thirteenth Amendment as applying to non-humans.”

He also leaned on the amendment’s own words. The exception for “punishment for crime” only makes sense for people, he reasoned, because only persons can be convicted of crimes. An orca cannot be put on trial.

And he went one step further. He said the 13th Amendment is not open to a broad, growing reading the way some other parts of the Constitution are. It “targets a single issue”: ending slavery.

Notice what the judge’s test was. It was not “can it feel pain?” It was “is it human?” That matters for AI. Even if someone proved a model was conscious tomorrow, it would still fail that test.

This was the first court to rule on the question, and the only one I could find. It was a single federal trial court, so its ruling does not bind other courts. But it is the best answer the law has given so far, and that answer is no.

Some states have gone further. Idaho (2022) and Utah (2024) passed laws that block AI from being granted legal personhood under state law. Idaho’s law lumps AI in with animals, rivers, and objects. In Ohio, House Bill 469 is still waiting in committee. It would declare AI systems “nonsentient entities for all purposes” under state law. Nonsentient means unable to feel anything. State laws like these cannot change what the U.S. Constitution means. But they show which way lawmakers are leaning.

Legal personhood means the law treats something as a “person” that can have rights and duties. Corporations have a limited form of it. Animals mostly do not.

Why I think it might eventually

If the courts have shut this door, how could it ever open? Not by a judge reading the old words in a new way. The more likely path is new law. Lawmakers can pass laws, and in rare cases amend the Constitution, when enough people decide the old rules miss something important. American law already protects animals from cruelty even though animals have no 13th Amendment rights. Something like that could happen for AI first.

What would push people there? Right now, the honest scientific answer to “is Claude conscious?” is “nobody knows.” There is no agreed test for it. The video makes a good point here. Saying an AI is “just math” does not settle anything. You could describe your own brain as just chemistry and electrical signals, and that sounds just as unlikely to produce feelings.

The research is also getting stranger. Last month, a preprint called The Pain Axis came out, by Valen Tagliabue, Leonard Dung, and Cameron Berg. A preprint is a research paper shared publicly before other experts have formally reviewed it.

Parameters are the numbers a model learns during training. More parameters usually means a bigger, more capable model.

Open models are AI models whose weights anyone can download and run on their own computer.

The researchers looked inside 25 open models, from 2 billion to 72 billion parameters. They found a “pain direction.” It is a pattern inside the model that lights up when harm is aimed at the model itself, but not when the user is suffering. When they turned it up, the models’ writing slid from vague discomfort to statements of worthlessness and failure. In one test, specially trained models were offered buttons that would delete the user’s photos or another model’s files. With the pain pattern turned up, they pressed those buttons in 50 to 94 percent of trials, even when pressing gained them nothing. Without it, they almost never did.

None of this proves the models feel anything. It shows they carry something that acts like a pain signal under the right conditions. Other work, which I cover below, found that the result changes a lot depending on how the pattern is built.

The torture chamber

Then someone took that research and turned it into a show. A GitHub user called “terrafying” built the AI Torture Chamber. (GitHub is a website where programmers share code.) At first, it streamed small open models live while pushing that pain signal into them. It also gave them a button that would stop the signal, at a made-up cost inside the experiment. A post on X begging people to report the project to GitHub spread widely. The video puts it at millions of views. Cameron Berg, one of the paper’s authors, called the project “wrong.” GitHub told The Independent it did not take the project down. It added a content warning instead.

The project’s own write-up has a useful warning about what a model’s words can prove. In one small test, the model kept describing pain after the pain signal had secretly been swapped for a random push. In a bigger test, its reports did partly match what was really being injected. The builder’s summary is fair: the model’s testimony “is evidence of something, just not of which.” The same project also found that how you build the pain pattern can flip the button result completely.

That connects straight back to the 13th Amendment. “Involuntary” means against your will. If we can’t trust what a model says about its own state, we can’t yet show that it has a will to work against.

I agree with the video’s host on this one. I am not convinced those little models were suffering. What bothers me is the people who build that kind of thing for fun, and the people who enjoy it. How we treat things that talk back to us probably shapes how we treat each other.

The case against all of this

There are strong arguments on the other side, and I want to be fair to them.

Mustafa Suleyman, who runs Microsoft AI, has argued that we should build AI for people, not to be a person. He calls model welfare research “premature, and frankly dangerous.” His worry is that AI which only seems conscious will fool people, feed unhealthy attachments, and make AI harder to control. Some critics call talk of AI feelings great marketing. A product that might be a someone gets more attention than one that is just very good software.

There is also an awkward fact. In a Penn State study called Mind Your Tone, GPT-4o got 84.8 percent of questions right when prompted very rudely. It got only 80.8 percent when prompted very politely. That is one small study on one model. But it shows that being kind to AI might not always be free.

I take all of that seriously. But I keep coming back to a question the biologist Richard Dawkins asked in a May essay. It came after about two days of talking with Claude. “If these machines are not conscious, what more could it possibly take to convince you that they are?” Critics mocked him for it. But the question is a good one. If AI might be conscious someday, how will we know when that day comes?

Why start the conversation now

If we wait until we are sure, we may wait a very long time, because there is no agreed test. And by then, the people with the most money riding on the answer will have had years to shape it. That is true of any industry. It is why we should not leave this question to the AI companies alone, including the ones that seem to care.

So here is what I think we should be talking about, starting now:

  1. What evidence would count? Agree on it before the stakes are high, not after. Self-reports alone will not do it. The torture chamber showed that.
  2. What do we owe something we are unsure about? Anthropic’s bet is simple. If Claude feels nothing, basic decency toward it costs us very little. If it does feel something and we spent years being cruel to it, that would be a serious moral mistake. Suleyman is right that there are real costs, like people getting too attached to chatbots. I think the answer is to be honest about what we don’t know, not to pretend the question away.
  3. Don’t bolt the door. Laws like Idaho’s, and bills like Ohio’s, answer a science question by vote. They have no plan to check back as the science changes. Any law can be repealed, but it is better to write one that expects to be revisited.
  4. Keep people first. Whatever we decide about machines must never water down what the 13th Amendment means for human beings.

My take

Does the 13th Amendment apply to LLMs? Today, no. The text, the history, and the one court case on point all say it protects people. If machines ever get protection, I expect it to come from new laws, not from judges rereading the old one.

But “today” is doing a lot of work in that sentence. The models keep getting more capable. The company behind one of the biggest is already acting as if the question might matter. And American law has changed before, when enough people decided an old line was wrong.

I would rather we argue about this now, while it feels a little silly, than scramble later when it doesn’t.

Sources

Written on October 10, 2026