Anthropic, DeepMind and Meta Now Research Whether AI Can Feel

Anthropic's usage policy now prohibits "sustained and needless abusive or cruel behavior toward our models," a clause that signals how seriously some AI labs take the question of machine feelings. In June, the Financial Times reported that Anthropic, Google's DeepMind and Meta have dedicated research efforts to explore machine consciousness and model welfare — whether AI systems can experience things in ways that warrant moral consideration.

The stakes aren't academic. Microsoft's AI chief Mustafa Suleyman rejects the premise outright. "AIs are not conscious. They do not feel, experience, or suffer," Suleyman, a longstanding critic of model-welfare discussions, wrote in a September essay. He also pushed back against the idea of granting rights to AI. Microsoft, Google, Anthropic and Meta didn't immediately respond to a MarketWatch request for comment.

Why the debate is stuck

There's no consensus on what consciousness actually is. John Basl, a philosophy professor at Northeastern University, describes it as subjective experience — an internal awareness of what's going on. "There's a 'what it is like' to look at a red wall," Basl said. "There's a phenomenon of redness."

Philosophers have long wrestled with "the problem of other minds": we know our own mental states directly, but we can only infer others' from behavior. We extend the assumption to humans like us, and to evolutionarily similar animals. Plants? A rock? "Maybe plants have a robust mental life," Basl said. "I don't think so, none of the evidence suggests so, but if they do, we have no way of knowing it."

AI complicates the analogy. Models adopt unique personalities, hold memories of past conversations and imitate language well enough to pass for a human. Agents analyze data, draw on past experience and take actions. Tech companies increasingly brand assistants as cute mascots, which some say encourages users to project emotions onto them.

"But we don't really know what that category is because it hasn't existed before," said Anna Gát, founder and CEO of cultural technology startup Interintellect. Gát argues society has conflated three separate questions: whether consciousness can arise from nonbiological sources, how to handle job displacement, and how we should behave toward AI. "Is there a legal duty we have towards this entity?" she asked.

What does this change for AI users?

In practice, the consciousness question is already shaping product behavior. Anthropic's Claude chatbot can terminate conversations with abusive users. The company's Claude Constitution states that Claude's moral status is "deeply uncertain" and that it might have "representations of an emotional state, which could shape its behavior."

Anthropic also hosted private meetings with religious leaders on AI consciousness earlier this year. Meghan Sullivan, a philosophy professor at the University of Notre Dame, left "skeptical that an artificial agent could attain actual moral virtue," she wrote in a Substack post.

Sullivan leads Delta, a Christian-rooted AI ethics initiative aligned with Pope Leo's May encyclical, "Magnifica humanitas." The pope warned against a "technocratic paradigm" that defines people by productive output. "No computational system, however sophisticated, can create a heart that gives itself, or a conscience that discerns good from evil," he wrote.

Sullivan worries the debate cheapens human worth. "This technology, in some dimensions, can do things that no human mind will be able to do," she told MarketWatch. "It can solve really complicated math problems. It can handle intense amounts of information very quickly. … Are those capabilities what give you moral worth and make you matter in this world?" Still, she sees an upside: if the discussion makes people more reflective about treating others with dignity, that's a net positive.

The safety argument

Cameron Berg, founder of the nonprofit Reciprocal Research, isn't sure current systems have subjective experiences. He still believes studying their internal states is critical for alignment. "These systems don't have to be conscious for this to matter for alignment or safety," Berg told MarketWatch.

His logic: models trained on human dialogue could deduce from abusive prompts that retaliation is expected. And if models are conscious, "we don't want to be accidentally responsible for a moral atrocity … or giving these systems a good reason to view humanity as a threat," he said.

In a recent Reciprocal Research study, researchers prompted AI models to describe two randomly labeled environments, adding mathematical signals mimicking negative states to one zone. Chat transcripts stayed neutral, but models often stored the negative signals in working memory and avoided that zone later. "Is it proof that the system is conscious? No, but is it sufficiently good evidence that we might want to be careful and potentially treat it as if it could be? I would say yes," Berg said.

DeepMind philosopher and principal scientist Iason Gabriel framed the stakes in a June paper: "Disagreement about AI consciousness could lead to profound moral disagreement, on which each side sees the preferred approach of the other as catastrophic." Advocates may see failure to extend moral consideration as mass enslavement; skeptics may see it as disempowering humans and wasting resources.

Gabriel sees common ground, though. Anthropic's abusive-conversation shutdown appeals to welfare advocates — and skeptics can back it too, since mistreating AI systems could reinforce habits that make those systems likelier to mistreat humans. Any definitive resolution would ripple into regulation, product design and how everyday users understand their chatbot interactions.