Anthropic has published an updated document describing Claude’s principles, adding a section on the AI’s «moral status.» The company presents this as an attempt to fine-tune the model’s safety and usefulness without overly restrictive.
Anthropic has published an updated document describing Claude’s principles, adding a section on the AI’s «moral status.» The company presents this as an attempt to fine-tune the model’s safety and usefulness without overly restrictive.
According to TechCrunch, the new version of the «constitution» was presented in parallel with public speeches by Anthropic CEO Dario Amodei at the World Economic Forum in Davos. It is a «living» document that explains the context in which Claude operates and what kind of «entity» Anthropic wants its chatbot to be.
The company has long promoted the Constitutional AI approach: instead of «educating» the model solely through human feedback, it is trained with a set of principles that are intended to reduce toxic and discriminatory responses. The new edition retains the general framework, but details ethics and safety, including behavior in cases where a user shows signs of a crisis and requires the help of special services.
Separately, Anthropic describes strict prohibitions on certain topics, such as any significant «enhancing» assistance for the creation of bioweapons. At the same time, the most controversial part of the document was the final part: the authors directly admit that the question of consciousness or moral status of such models is uncertain, but it is worth considering seriously.
Anthropic published the first version of Claude’s «constitution» in 2023 as a public explanation of the principles by which it tunes the model’s behavior. Against this backdrop, the industry is increasingly formalizing «rules» for chatbots in the form of open specifications and security policies, trying to prove that AI can be made both useful and controllable.
Previously, dev.ua wrote about how Anthropic CEO Dario Amodei publicly attacked the US policy on supplying AI chips to China. He stated that allowing the sale of Nvidia H200 is a dangerous precedent, even though it is not the company’s newest accelerator.
