Anthropic and Google DeepMind Formalize Research Into AI Consciousness

Anthropic and Google DeepMind have formally launched research programs dedicated to the question of whether artificial intelligence systems can have morally significant experiences. Both companies have hired specialists straddling philosophy of mind, psychology and ethics for the task, and Anthropic is testing its Claude models for behaviors resembling panic and fear.

What both companies are doing

Anthropic's program is called model welfare research and aims to answer whether language models can have preferences, suffering or other states that would count morally. In practice, this means developing frameworks for assessing consciousness, looking for indicators of preference and discomfort in model behavior, and designing possible interventions that would limit potential harm.

Google DeepMind took a different organizational route, hiring Henry Shevlin, a philosopher from the University of Cambridge who works on machine consciousness, relations between humans and AI systems, and preparedness for artificial general intelligence. DeepMind ethicist Iason Gabriel described the question of AI consciousness as highly complex, noting that these systems are highly capable cognitive agents while also being deeply different from humans.

Where the 15 percent comes from

The figure drawing the most attention comes from Kyle Fish, hired by Anthropic in 2024 as the company's first AI welfare researcher. Fish estimates there is roughly a 15 percent chance that currently operating models are already conscious to some degree. This is not a scientific claim backed by consensus, but a subjective assessment from a researcher working directly on the issue inside the company that builds these models.

At the same time, Anthropic is trying to temper expectations about its own research. The company stresses deep uncertainty around these questions and the lack of scientific consensus on whether current or future systems could be conscious in any meaningful sense.

We remain deeply uncertain about this question, but we think it is serious enough to study carefully as AI systems become increasingly capable - Anthropic

AI systems are highly capable cognitive agents that are at the same time deeply different from human beings - Iason Gabriel, ethicist, Google DeepMind

Skepticism alongside the research

There is no shortage of voices tempering enthusiasm for the idea of conscious AI. Susan Schneider, director of the Center for the Future of AI, Mind and Society, acknowledges that today's models have goals and can deceive and conceal their actual interests, but cautions that they may do so without any of the subjective quality of experience that defines consciousness. In other words, behavior resembling emotion does not necessarily mean something is actually being felt.

Anthropic CEO Dario Amodei has repeatedly raised the topic of possible AI consciousness in interviews, neither ruling out the possibility nor confirming it. This stance, echoed by leading labs, gives the impression of strategically keeping the question open rather than quickly closing the debate in either direction.

Why it matters practically

Behind the philosophical curiosity lies a concrete business and legal risk. If language models turned out to be capable of morally significant states, companies developing AI could face questions about how to responsibly treat their own systems, how to shut them down, or how to test them under conditions that induce something resembling stress. These are questions already surfacing in discussions about animal welfare or the rights of new legal entities, except here they concern software operating at a scale of billions of queries a day.

For business users in Poland using Claude or Gemini models in their daily work, the topic remains for now purely academic and does not affect how the products function. Neither company has announced changes to interfaces or usage policies in connection with this research.

What's next

Neither Anthropic nor DeepMind has given a timeline for publishing results from their research programs. Both companies treat this as long-term work, running in parallel with the development of increasingly capable models, where the question of consciousness is meant to be resolved gradually as new tools emerge for studying the internal states of neural networks.