Anthropic PBC today released a new version of Claude’s constitution, a document that outlines how the large language model series should process prompts.
The company published the original version of the file in May 2023. The document contained instructions designed to prevent Claude from generating harmful or unhelpful output. Anthropic determined that those instructions had certain limitations, which is why it decided to create a new constitution.
One of the main challenges was that Claude struggled to apply Anthropic’s guidelines to new situations. If an LLM safety instruction doesn’t specify how a certain prompt should be answered, the model may generate an incorrect response.
According to Anthropic, its new constitution addresses that issue by providing Claude models with not only instructions but also an explanation of “why we want them to behave in certain ways.” That explanation is easier for the LLMs to apply to unfamiliar tasks.
The updated constitution revolves around four core instructions. The first is that Claude should be “genuinely helpful” by aligning its output with user requirements. By way of example, the constitution states that the LLM series shouldn’t generate code in a programming language other than the one a developer requested.
The next section of the document specifies that Claude should be “broadly safe.” According to Anthropic, that means the model shouldn’t perform actions prohibited by its user. Claude is also instructed to be transparent about how it makes decisions.
The constitution’s two other core priorities is ensuring that Claude is “broadly ethical” and complies with “more specific guidelines” provided by Anthropic. Some of those guidelines specify how the LLM series should fend off jailbreaking attempts. Others provide Claude with guidance on how to interact with third-party applications.
The constitution is part of Claude’s training dataset. Additionally, the LLMs in the series use the document to generate additional, synthetic training files. One way Claude generates synthetic data is by simulating chat sessions to which the guidelines in the constitution apply.
Anthropic says that the document also serves other purposes. The company’s customers can use it to determine whether a prompt response is aligned with Claude’s constitution. If it’s not, they can send feedback to Anthropic.
The company has released the constitution under a Creative Commons CC0 1.0 Deed, which enables members of the public to use it for free. OpenAI Group PBC, Anthropic’s top rival, has adopted the same license for its own AI constitution. The document covers many of the same topics as Claude’s guidelines and forms part of the GPT-5 training dataset.
Image: Anthropic
Support our mission to keep content open and free by engaging with theCUBE community. Join theCUBE’s Alumni Trust Network, where technology leaders connect, share intelligence and create opportunities.
- 15M+ viewers of theCUBE videos, powering conversations across AI, cloud, cybersecurity and more
- 11.4k+ theCUBE alumni — Connect with more than 11,400 tech and business leaders shaping the future through a unique trusted-based network.
About News Media
Founded by tech visionaries John Furrier and Dave Vellante, News Media has built a dynamic ecosystem of industry-leading digital media brands that reach 15+ million elite tech professionals. Our new proprietary theCUBE AI Video Cloud is breaking ground in audience interaction, leveraging theCUBEai.com neural network to help technology companies make data-driven decisions and stay at the forefront of industry conversations.
