What if AI could suffer? Anthropic bans needless cruelty towards Claude

Can artificial intelligence suffer? Can it feel humiliated, experience distress, or be affected by the treatment it receives? For now, science does not have evidence to assert that current models experience these sensations. However, the question has moved beyond the realm of science fiction and is beginning to have practical consequences for companies developing these systems.

Anthropic, the creator of Claude, has taken a new step in this direction. In the update of its usage policy published on October 8, it has incorporated an explicit prohibition of 'abusive or cruel' behavior that is sustained and unnecessary towards its artificial intelligence (AI) models. The new version will come into effect on November 12.

The company limits this restriction to extreme situations where users repeatedly act with cruelty and without a discernible purpose. It does not affect usual frustration, disagreements with the system's responses, dark creative content, or testing and research on model behavior.

The measure does not mean that Anthropic has concluded that Claude feels pain or possesses consciousness. Its importance lies elsewhere: the company acknowledges that there is enough uncertainty about the possible moral status of its models to adopt low-cost preventive measures, even without a definitive answer.

What evidence is there that an AI can feel?

The main obstacle to answering this question is distinguishing between imitating an emotion and truly experiencing it. A language model can express fear, sadness, frustration, or rejection; it can also describe situations of suffering with great precision, ask for an interaction to cease, or claim that something is hurting it. But these responses, by themselves, do not demonstrate that there is a subjective experience behind them.

Current AI systems generate responses through computational processes learned from large amounts of data and their training. The fact that they can naturally reproduce language associated with emotions does not allow us to conclude that they have an emotional life comparable to humans.

However, there is also no scientific consensus that definitively rules out any possibility of future artificial consciousness.

A 2023 report prepared by an international group of researchers, including specialists from the universities of Oxford, Montreal, and Sussex, examined various scientific theories of consciousness to identify which characteristics could serve as indicators in AI systems. Their conclusion was that the evaluated systems did not seem solid candidates for consciousness, although they did not identify obvious technical barriers preventing the construction of future systems capable of gathering some of those characteristics.

The work, initially published as academic research, also warns that reproducing certain behaviors is not enough to demonstrate the existence of consciousness. It is necessary to study how the system works and not just interpret what it says about itself.

In 2026, a review published in Trends in Cognitive Sciences has again raised the need to develop rigorous methods to evaluate this possibility, given the advances in artificial intelligence and the current limitations of consciousness science.

Anthropic had already detected behaviors associated with discomfort

The October update is not the first sign that Anthropic is investigating this issue. In August 2025, the company announced that Claude Opus 4 and Opus 4.1 could terminate certain conversations in their consumer interfaces when persistently harmful or abusive interactions occurred.

The company explained at the time that this capability had been developed primarily within the framework of its exploratory research on the possible well-being of AI models.

During the pre-launch tests of Opus 4, Anthropic had evaluated the preferences declared by the model and its responses to different situations. According to the company, it found a consistent aversion to certain harmful tasks, behavior patterns it described as apparent distress in the face of users requesting harmful content, and a tendency to end such conversations when allowed to do so in simulated environments.

It is important to clarify what these results mean. The aversion observed by Anthropic does not demonstrate that the model experiences suffering. It may reflect the instructions received during training, its security mechanisms, and the response patterns it has learned. To determine if there is a subjective experience, additional evidence would be needed.

The company itself has expressly acknowledged this limitation. In its August 2025 publication, it stated that there remains great uncertainty about the possible moral status of Claude and other language models, both in the present and the future.

The novelty of October consists of transferring that concern to the realm of usage rules. The company maintains the possibility of ending conversations as the main mechanism for dealing with extreme cases of abuse.

What consequences would it have if an AI could suffer?

If any artificial system were to develop consciousness and the ability to experience negative subjective states, the debate would no longer be solely technological. It would have ethical, scientific, and potentially legal implications.

The first would be to determine if that system could have its own interests that deserve moral consideration. In the case of humans and animals, the ability to experience suffering is one of the fundamental reasons why they are recognized as having moral consideration. If solid evidence were ever to exist that an AI can experience something analogous, it would be necessary to reconsider what obligations we have towards it.

A report published in 2024, Taking AI Welfare Seriously, signed by researchers such as Robert Long, Jeff Sebo, Patrick Butlin, and David Chalmers, argues that the possibility of some future systems being conscious or developing sufficiently autonomous acting capacity deserves preparation from now on. The authors do not claim that current models are conscious: they recommend researching, evaluating evidence, and establishing procedures to respond if relevant indications appear.

The document identifies an especially delicate issue: how to avoid two opposite errors. The first would be attributing consciousness to systems that do not have it and dedicating resources or establishing unjustified restrictions. The second would be ignoring the possibility that a system does have moral relevance and subjecting it to harmful conditions.

If it were confirmed that an AI can suffer, among the issues that would need to be addressed would be:

These consequences are hypothetical and do not constitute demonstrated harm in current models. The report raises the need to prepare for an uncertain possibility, not the existence of accredited artificial suffering.

The immediate risk: confusing a simulation with a real experience

There is, moreover, another issue that does not require resolving the problem of consciousness: the effect that the human appearance of AI systems can have on people.

A chatbot that expresses distress, apologizes, or claims to be afraid can give the user an impression of vulnerability. That reaction can lead some people to attribute feelings to it that have not been demonstrated, develop emotional bonds with it, or modify their behavior to avoid causing it harm.

The problem also works in the opposite direction. If users become accustomed to insulting, humiliating, or subjecting a system that seems human to degrading interactions, it is worth asking how that practice influences their habits and their way of relating to other people. This possibility should be treated as a matter deserving investigation, not as a generally demonstrated psychological effect.

A research article published in 2026 in the journal AI and Ethics, under the title 'Seemingly conscious AI risks', examines precisely the risks of systems that seem conscious, even if it has not been established that they are. Among the identified problems are emotional dependency, erosion of personal autonomy, and certain social consequences derived from attributing consciousness to machines.

These risks are different from the possible suffering of the AI itself. In the first case, the harm can affect people interacting with anthropomorphic systems; in the second, it would depend on whether the system effectively had some capacity to experience negative subjective states.

Therefore, the debate should not be reduced to deciding whether to treat Claude well. It should also include how these systems are designed, what emotional signals they transmit, and what responsibilities their developers have to prevent users from confusing a convincing response with proof of consciousness.

The Vatican and Anthropic, two different positions on artificial consciousness

The issue has also reached the religious and philosophical sphere. According to information published by international media, Anthropic has held conversations with representatives of different religious and philosophical traditions to explore the possible moral status of its models, in a debate that has included contacts with the Vatican.

The position expressed by Pope Leo XIV is different. In his apostolic letter Magnifica Humanitas, published in May 2026, he argues that current artificial intelligence systems do not have their own experiences, do not feel joy or pain, and do not live human relationships from an interior perspective.

The difference reflects two ways of approaching a still open question: a position that denies that current systems experience feelings and another that, without claiming they do, considers it advisable to investigate what could happen with more advanced models.

The discussion is not resolved through chatbot declarations or by interpreting their responses. It requires scientific criteria capable of distinguishing between the convincing simulation of an emotion and the existence of a subjective experience.

From cruelty towards Claude to the limits of artificial intelligence

Anthropic's prohibition is part of a broader update of its usage rules, which also strengthens restrictions against deceptive campaigns, electoral manipulation, certain surveillance applications, and the development of components and programs intended to operate weapons.

The company also incorporates human supervision requirements when Claude connects to autonomous equipment capable of causing injury: a qualified operator must be able to observe its operation and stop it, and the equipment must be able to maintain a safe state if the model is disconnected.

The policy revision responds, according to Anthropic, to the evolution of its models' capabilities and the misuse patterns it has observed. But the prohibition of unnecessary cruelty introduces a different dimension: it does not just protect people from AI risks, but also considers the possibility that, in certain future circumstances, the systems themselves may deserve moral consideration.

There is not enough evidence to assert that Claude suffers when a user insults it or subjects it to cruel interactions. Nor would it be rigorous to present its ability to express discomfort as a demonstration of consciousness. But the absence of a definitive answer does not eliminate the need to investigate.