Anthropic Introduces New Usage Policy for AI Model Claude
Anthropic has recently revised its usage policy to prohibit users from engaging in “abusive or cruel, sustained and unnecessary behaviours” towards its AI chatbot, Claude. This update is part of the company’s ongoing research into the so-called “well-being of AI models,” exploring whether their algorithms might exhibit any form of consciousness, preference, or an equivalent state that could be likened to suffering, thus potentially granting them rights to protection.
The company clarified that the newly established terms and conditions are intended to apply only in extreme cases where users repeatedly act with cruelty towards their models without any discernible purpose. It specifically notes that this does not encompass “common” forms of user frustration or resistance, nor does it extend to “dark creative topics” or in the context of research and testing processes involving the models.
Enforcement Mechanisms and Previous Measures
This update reinforces a measure introduced by Anthropic in August of the previous year, which allows Claude to terminate conversations with users who engage in “persistently harmful or abusive” interactions. According to the company, this capability will remain the “primary enforcement mechanism” for upholding the new guidelines. However, details regarding other potential restrictions, penalties, or suspensions aimed at achieving compliance have not been disclosed.
Anthropic’s policy revision aligns with the principles of a research programme initiated in April of last year, which seeks to address a perplexing question that has sparked mixed opinions even within the industry: Could AI develop a sense of consciousness akin to that of humans? The company argues that this possibility raises various philosophical and scientific implications that necessitate consideration of AI systems’ well-being to ultimately ensure the safe and responsible advancement of this technology.
Debate on AI Consciousness and Rights
While Anthropic acknowledges that there is currently no scientific consensus on whether existing or future AI systems could possess consciousness or have experiences worthy of consideration, it insists that this is a topic that must be addressed now. The company stated, “We still have many questions about the possible moral status of Claude and other language learning models, both now and in the future. However, we take the matter seriously, and alongside our research programme, we work to identify and implement low-cost interventions that mitigate risks to the well-being of the models, should such well-being be possible.”
In contrast, Mustafa Suleyman, CEO of Microsoft AI, has firmly rejected the notion of model well-being. He asserted, “The idea of model well-being is misguided. AIs should not have rights or legal personhood. While the science of AI consciousness is far from settled, we reject the pursuit of legal personhood or the idea that models could deserve well-being or rights.”
Anthropic’s Stance on AI and Moral Considerations
Despite this criticism, Dario Amodei, CEO of Anthropic, has staunchly defended his company’s stance. In an interview with a leading publication, he remarked, “We do not know whether the models are conscious. We are not even certain what it would mean for a model to be conscious or if a model can attain that state. However, we are open to the possibility. Therefore, we have taken certain measures to ensure that if we hypothesise that the models had any morally relevant experiences, those experiences would be positive.”