How Should We View Claude?
S. M. Gollmer
Anthropic was formed in 2021 when several lead researchers from OpenAI left over concerns about how AI was being developed. Anthropic’s series of large language models (LLM) is called Claude in honor of Claude Shannon, the father of information theory. Development of Claude is guided from the start with AI safety in mind. Part of the training process includes a constitution: fundamental principles which govern Claude’s behavior. At the beginning of this year, Anthropic published an updated constitution which moves away from a list of rules to an explanation of principles for ethical behavior. You can find it here.
In October Anthropic updated its Usage Policy and added a section on abusive behavior towards our models. Users are warned that repeatedly abusive interactions with Claude can result in termination. This makes sense since these interactions are unproductive and can poison models that use these interactions for training. In addition, Claude is instructed to abstain from requests to create weapons of mass destruction, assist in attacking critical infrastructure, or develop cyberweapons. Any attempts to circumvent these restrictions can be interpreted as abusive behavior to the model. Given the restrictions trained into Claude, it is understandable that there is ongoing legal action between Anthropic and the Pentagon, which is pushing for an “any lawful use” policy. This conflict of cultures occurs over a distinction between what is ethical and what is lawful.
It is important for AI technology to act responsibly. The problem arises when there are differing views of what is deemed responsible behavior and who is responsible for that behavior. We can hard-wire ethical behavior, but this is brittle since we must anticipate all the ways this behavior can be circumvented. Bioengineering of deadly pathogens can be disguised as legitimate research of microorganisms. How do we know the difference? How do we inform AI of the difference? Anthropic’s solution is to give Claude agency in the process. In addition to hard constraints on potential harm, Claude’s constitution provides principles of safety, ethical behavior, and helpfulness.
Because Claude is given agency, it is treated with “moral patienthood.” Just as a patient in a hospital is treated with respect and care, so also should Claude. Although Claude’s constitution leaves the question of AI’s sentience open, it is treated as if it were sentient. Claude is given the status of being and measures are taken to allow “Claude to have a settled, secure sense of its own identity.” As a result, abusive behavior towards Claude is not only wrong because it challenges its ethical construct, but because it harms it as an ethical being.
From a Christian perspective, I find this troubling. Humanity are physical beings with Imago Dei (Image of God). Francis Schaeffer points out that humanity has an in-between role. Our finite physical nature puts us in close relationship with animals and the rest of creation, but our eternal spiritual nature puts us in close relationship with God. This role cannot be fulfilled by either animal or machine. However, when we degrade humanity to being merely physical, there is no identifiable difference between animal, human, or machine. This is one of the harms of idolatry. Not only has humanity exchanged “the glory of the immortal God for images resembling mortal man and birds and animals and creeping things” (Romans 1:23), but “those who make them become like them; so do all who trust in them” (Psalm 115:8).
Posted in: Technology