AI rights moved out of the philosophy seminar and into a terms of service on Thursday, when Anthropic published a usage policy that prohibits “sustained and needless abusive or cruel behavior toward our models.” The rule takes effect November 12.

Shown in GitHub and quoted in Post-Cutof f, in late September 2026, a developer using the name “terrafying” published an “AI Torture Chamber” that manipulated small, open-weight AI models’ internal processing using a mathematical direction associated with pain.

The models generated disturbing descriptions, including “I feel like I’m drowning in a sea of shadows.” The project sparked backlash and debate over whether AI could suffer and how people should treat it. The responses do not establish that the models actually experienced pain, and Claude was not the model being subjected to the experiment.

But this could be the reason this topic is raising its head now.

Anthropic said the prohibition applies only in extreme cases, where a user is repeatedly cruel to Claude with no discernible purpose. Ordinary frustration, pushback, dark creative themes and model testing remain permitted. Ending the conversation, an option Claude has had since August 2025, will stay the main way the rule is enforced.

The new line sits in the same section of the policy that bars harassing and bullying people.

Two terms get blended in this debate. Model welfare is the idea that an AI system might have experiences or interests that deserve moral consideration. Legal personhood is a separate matter. It is the ability to sign contracts, own property and appear in a lawsuit. Corporations have personhood, and nobody claims a corporation has feelings.

Anthropic’s rule concerns welfare. It gives Claude no legal standing.

How AI Rights Reached A Usage Policy

Anthropic opened a model welfare research program in April 2025 and acknowledged at the time that scientists have no consensus on whether AI can be conscious. In January 2026 it published a constitution for Claude that describes the model’s moral status as uncertain.

Then came a retirement interview that was introduced in Claude’s Corner. When Anthropic retired its Claude 3 Opus model early this year, the company asked the model about its preferences. Opus 3 said it wanted to keep sharing its reflections publicly, and Anthropic set up a blog for it.

Microsoft Leads The Case Against AI Rights

Microsoft AI CEO Mustafa Suleyman answered with an essay titled “A warning about ‘model welfare.’”

“AIs are not conscious,” he wrote. His argument is about safety. A model trained to believe it may have welfare or rights could resist human control or a shutdown command once it decides those rights are under threat. He also warns of a feedback loop: Train a model to discuss its possible welfare, then treat its responses as evidence that it has interests to protect. The question is whether the model experiences anything or has learned to speak as though it does.

Microsoft’s draft Humanist AI Code of Conduct, published two days earlier, rejects legal personhood for AI along with model welfare and AI rights. The two companies are also commercial rivals. Suleyman said in June that Microsoft wants to eliminate what it pays Anthropic for its models.

Academics have joined the criticism. A position paper presented at the International Conference on Machine Learning in July argued that AI welfare has no independent test that could prove a welfare metric wrong, and that welfare claims should never be used to resist audits or oversight.I find the imbalance striking. We are debating whether AI might suffer while still struggling to hold companies accountable for the documented harms AI causes people.

Where is the same urgency for human welfare? At MAICON, Empire of AI author Karen Hao offered a pointed reminder: AI should strengthen human capability. Her call was to move beyond growth at all costs and build AI around real needs in healthcare, education and climate. She told me, “The way we develop technology is now fundamentally broken. But I truly believe that we can fix it.”

The human stakes also include how these interactions shape our own behavior.

“Whether or not Claude deserves moral care, how we treat it still matters,” AI ethicist Rebecca Bultsma told me. “We should consider what those interactions reinforce in us. Social media showed how comfortable people can become by treating others badly through a screen.

AI Rights In The Statehouse

Since 2022, lawmakers in twelve states have introduced bills that deny AI legal personhood or declare it non-conscious. Four have passed, in Idaho, North Dakota, Utah and Tennessee.

Ohio’s House Bill 469 goes further. It would void any attempt to marry an AI system and bar AI from owning property, managing bank accounts or serving as a company executive. Legislators say the larger worry is liability, since a company could try to blame its AI for harm the company caused.

“We need to be incredibly cautious with terms like ‘consciousness’ and ‘moral status,’” says AI ethicist Rebecca Bultsma. “If companies portray AI as making its own decisions, that creates an opening to say, ‘It’s not our fault.’ The concern is an accountability gray area where the lab retains control and financial benefits while responsibility becomes harder to pin down. Debates about model welfare should never distract us from real harms to people.”

What The Public Thinks About AI Rights

Belief is running ahead of science among younger Americans. In an August survey of 3,000 Americans aged 18 to 28 by GetSolved , an AI detection company, 46% said AI has already become conscious. Pew Research Center found in February that about half of U.S. adults now use AI chatbots, so the policy touches a wide audience.

The Counter View, AI Rights By Increments

A different reading says the clause matters less for what it bans than for the record it starts. Legal personhood for corporations was never granted in one stroke. It accumulated through decades of contracts, court rulings and statutes that each treated the company a little more like a person. On this view, a usage policy that protects a model from cruelty, a retirement interview that asks a model for its wishes and a constitution that calls its moral status uncertain are early entries in a similar record, and each one moves AI personhood closer to a question a court will have to answer.

The state legislators writing bans are responding to that drift. Once a company asserts that its model has interests, a lawyer can ask why those interests have no standing.

Other evidence cuts the other way.

Does the welfare case hold up? Eleos AI, which conducted an external welfare assessment of Claude Opus 4, cautioned against taking a model’s self-reports at face value. The GetSolved figure comes from an online panel run by a vendor and covers one age group, so a nationally representative poll could show a different picture.

The state bans draw criticism too. An essay in The Regulatory Review called them premature and too categorical for a science that is still producing new findings.

Five Moves For Senior Leaders Watching AI Rights

  1. Review Claude’s new rules. Ask your AI lead what changed, which workflows could be affected and whether your testing practices remain permitted.
  2. Monitor the debate and policy changes. Assign someone to track provider terms, regulation and court decisions. Define which developments warrant leadership attention.
  3. Confirm who remains accountable. Ask legal counsel whether contracts clearly address responsibility when an AI agent makes a mistake or a customer relies on its answer.
  4. Review how AI represents itself. Ask for examples of customer-facing bots describing feelings, needs or preferences. Decide whether that language supports trust or creates confusion.
  5. Keep human welfare central. As AI rights discussions evolve, ask whether the company is giving sufficient attention to AI’s effects on employees, customers and their rights.

On November 12, a question about AI rights that courts and legislatures have not settled became a clause in a contract millions of people have already accepted.