Anthropic has decided to release the system prompts for its AI models, Claude 3.5 Opus, Sonnet, and Haiku. The initiative, aimed at enhancing transparency, makes the prompts available through the Claude app on iOS and Android, as well as online.
Role of System Prompts
System prompts form the core directives that dictate the behavior and responses of AI models. The instructions influence the tone, mood, and ethical guidelines for the AI's interactions. For example, a prompt may urge the AI to stay polite yet avoid excessive apologies, or to acknowledge its limitations in knowledge. The main intention is to minimize the AI producing harmful or unseemly content.
Anthropic's move to make these prompts public highlights its dedication to ethical AI. Alex Albert, Anthropic's head of developer relations, has noted that the prompts will be updated regularly, with changes documented in release notes. The strategy may set a benchmark for other AI firms, encouraging similar transparency.
Prompt Details
Dated July 12, the prompts provide detailed instructions for the Claude models. For instance, the models are instructed not to open URLs or view videos and are barred from performing facial recognition tasks. Claude 3.5 Opus, in particular, is told to act as if “completely face blind,” avoiding identifying individuals from images.
The prompts also touch on the personality of the models. Claude is characterized as “very smart and intellectually curious,” and is encouraged to engage in diverse discussions. When it comes to controversial topics, the AI is directed to maintain neutrality and provide “clear information” and “careful thoughts.” Additionally, the AI should avoid using definitive terms like “certainly” and “absolutely.”
Impact on the AI Sector
Anthropic's disclosure of system prompts is a significant step in an industry often critiqued for its opacity. By opening up about its guidelines, Anthropic aims to highlight the need for human oversight in AI operations. The move might influence other AI providers to pursue greater transparency.
However, despite Anthropic's drive for transparency, the company is not without its controversies. Earlier this month, I reported on a US lawsuit filed against Anthropic by a group of journalists and authors. The case, filed in California, accuses the company of basing its operations on the unapproved use of numerous copyrighted books.
Andrea Bartz, Charles Graeber, and Kirk Wallace Johnson contend that Anthropic‘s methods are depriving authors of their deserved earnings.
