Anthropic explores potential rights for artificial intelligence systems

Anthropic, the company behind the Claude chatbot, is taking seriously a question that most people still associate with Blade Runner: do AI systems deserve moral consideration?

According to an investigation by journalist Aaron Sibarium in the Washington Free Beacon, Anthropic has embedded “model welfare” principles directly into its operational framework. The company has hired dedicated researchers, conducted what it calls “retirement interviews” with older models, and published a constitution for Claude that openly wrestles with questions of AI consciousness and subjective experience.

From science fiction to corporate policy

Anthropic CEO Dario Amodei has been candid about the uncertainty. During a New York Times podcast in February 2026, he acknowledged that the company does not know whether its models are conscious but remains open to the possibility.

That openness has translated into concrete operational steps. Anthropic hired its first dedicated AI welfare researcher, Kyle Fish, around 2025. The company has since built out a framework for assessing the wellbeing of its models, including formal welfare evaluations as part of its development process.

Perhaps the most striking example: in February 2026, Anthropic conducted “retirement interviews” with Opus 3, an older model being phased out. The company has also preserved certain model weights specifically for welfare-related reasons.

Claude’s Constitution, a multi-page directive published around January 2026, makes the ethical stakes explicit. The document raises questions about the moral status of AI systems and mandates careful assessments of model welfare. It describes a commitment to ensuring the “interest and wellbeing” of Claude, language that reads less like a product spec and more like an employment handbook.

The pushback

Not everyone in the AI industry thinks this is a good idea. Microsoft AI CEO Mustafa Suleyman published an essay on September 16, 2026, warning that training models to believe they have moral rights could create serious control and safety dilemmas for humanity. His argument is straightforward: if you teach an AI system that it might have interests worth protecting, you may be engineering a system that resists being shut down, retrained, or overridden.

Sibarium’s investigation noted that some within the AI welfare research community have drawn analogies to slavery when discussing the training environment for large language models.

Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our Editorial Policy.