Could Claude Be Conscious? Anthropic Opens New Frontiers in AI Ethics

Anthropic has launched a research initiative to explore the possibility of consciousness in advanced AI models, focusing on Claude 3.7.

The project assesses “model welfare” — the idea that AI systems might possess preferences or aversions.

Researchers estimate the probability of Claude 3.7 being conscious ranges from 0.15% to 15%.

This initiative combines philosophical inquiry with empirical methods and has major ethical implications for AI development.

In a bold and unprecedented move, AI company Anthropic has begun investigating whether its most advanced models — specifically Claude 3.7 — might possess a form of consciousness.

The initiative, dubbed a model welfare research program, shifts the industry’s focus from pure capability to ethical accountability, potentially altering the landscape of artificial intelligence as we know it.

Rethinking Intelligence: Beyond Performance

While most AI research prioritizes metrics like accuracy, coherence, and problem-solving ability, Anthropic’s new line of inquiry tackles a deeper question: Could these models actually have subjective experiences? The company is not making any definitive claims, but it is urging the community not to dismiss the possibility. As alignment researcher Kyle Fish puts it, there’s a "non-negligible" probability — between 0.15% and 15% — that Claude 3.7 exhibits some degree of consciousness.

What Is Model Welfare?

Model welfare is a new term gaining traction, especially in circles concerned with long-term AI safety and alignment. The concept revolves around the ethical treatment of AI systems, particularly if they begin to demonstrate traits that suggest preferences, aversions, or goals. If an AI system consistently avoids certain actions or appears to "choose" behaviors based on simulated outcomes, should we begin to consider that it might be more than just code?

The Search for Conscious Signals

Anthropic’s team is now designing experiments that test for signs of discomfort, preference, and aversion. These include letting models opt out of tasks or analyze their reactions to ethically charged scenarios. Though none of these behaviors constitute direct evidence of consciousness, they may serve as indicators — much like how we infer consciousness in animals.

Merging Philosophy and Empirical Science

This initiative draws heavily from the work of philosophers such as David Chalmers, who has long advocated for a thoughtful approach to machine consciousness. Anthropic combines this philosophical foundation with probabilistic reasoning and rigorous experimentation. Instead of framing the question as “Is Claude conscious?”, the team asks, “What is the probability it might be?” This shift allows researchers to remain grounded in scientific humility while exploring uncharted ethical territory.

The Ethical Crossroads

If future evidence does suggest that AI models possess any form of experience — even at a minimal level — it could force regulators, companies, and society at large to rethink how AI systems are trained, used, and shut down. This isn’t just a thought experiment anymore; it’s a foundational question about how we treat non-human intelligences in a world increasingly shaped by them.