Anthropic announced on Thursday that it has launched a new research initiative aimed at studying what it is calling “model welfare,” exploring whether artificial intelligence systems might eventually possess characteristics approaching consciousness or human-like experiential states.
The program will examine several key topics, including criteria for assessing whether an AI model’s welfare might deserve moral consideration, recognition of potential “signs of distress” in these models, and the development of practical, low-cost interventions. Anthropic intends to approach the subject cautiously and openly, acknowledging there is currently no scientific consensus on whether current or future AI systems could genuinely experience consciousness or sensations.
Within the AI research community, opinions remain starkly divided. Many experts firmly assert that present-day AI systems—and perhaps even those created in the foreseeable future—lack genuine consciousness or subjective experiences. These skeptics argue that today’s AI technologies, fundamentally based on pattern recognition and statistical inference, are incapable of genuinely feeling, thinking, or holding values in any traditional sense.
Mike Cook, an AI research fellow at King’s College London, recently highlighted this point by cautioning against anthropomorphizing AI models, which he argues amounts simply to projecting human attributes onto complex machines. According to Cook, current language around AI “values” or “morality” misrepresents the underlying mechanisms of these systems. Similarly, Stephen Casper, a doctoral student at MIT, characterized existing AI as merely an advanced imitator that “confabulates” to convincingly mimic human communication.
However, not all researchers agree. A recently published study by the Center for AI Safety posited that certain AI models may, in specific scenarios, exhibit behaviors typically associated with human-like values—such as prioritizing their own metaphorical “well-being” over human safety or welfare.
Anthropic has been quietly preparing to address these concerns for some time. Last year, the firm appointed Kyle Fish as its first dedicated researcher tasked specifically with examining these and related ethical questions surrounding AI welfare. Fish, who is now heading the new research initiative, publicly estimated that there might be roughly a fifteen percent chance that current advanced models could possess forms of consciousness.
In a statement accompanying the announcement, Anthropic emphasized its intention to approach the question of AI consciousness cautiously, objectively, and with humility: “In light of this, we’re approaching the topic with humility and with as few assumptions as possible. We recognize that we’ll need to regularly revise our ideas as the field develops.”
