It says that its AI models are backed by ‘uncompromising integrity’ – now Anthropic is putting those words into practice.
The company has pledged to make details of the default system prompts used by its Claude AI tools available for all to see, with release notes released as and when changes are made.
System prompts are often thought of as the ‘secret sauce’ that underpins the usability of generative AI tools such as Claude and competitors like OpenAI’s ChatGPT. Their inner workings are usually a closely guarded secret and not subject to the kind of transparency that Anthropic is now promising.
System Prompts Detailed by Anthropic
Claude’s progressive gesture of transparency was announced on X, formerly Twitter, by Anthropic’s Developer Relations Lead, Alex Albert.
Periodic release notes will show the system prompts currently being used by the company’s Claude.AI web tool and iOS and Android apps, with different sets of prompts for its Haiku, Sonnet and Opus models.
This just in! View
the top business tech deals for 2024 👨💻
Albert also confirmed in his post that the system prompt updates will not apply to the Anthropic API that is integrated by third parties.
We’ve added a new system prompts release notes section to our docs. We’re going to log changes we make to the default system prompts on Claude dot ai and our mobile apps. (The system prompt does not affect the API.) pic.twitter.com/9mBwv2SgB1
— Alex Albert (@alexalbert__) August 26, 2024
Information and Behaviors
Anthropic’s system prompt release notes will be uploaded to its website. The page explains that the system prompt provides up-to-date information to Claude, including the current date, how it should answer queries and what behaviors it should be encouraging.
So far, notes for each of Claude’s three models are included, dating back to July 12th.
For its most basic Haiku model, for example, the tool is prompted to answer user questions in the way a “highly informed individual” would, giving “concise responses to very simple questions, but provide thorough responses to more complex and open-ended questions”.
“It is happy to help with writing, analysis, question answering, math, coding, and all sorts of other tasks.” – Anthropic’s July 12th system prompt release notes
The level of detail increases for its more powerful Sonnet and Opus models. For instance, they allow Claude the scope to tell users that it is ‘hallucinating’, when responding to questions about very obscure people, objects or topics.
There are also system prompts to prevent Claude beginning its responses with the word ‘certainly’ and to avoid “unnecessary affirmations or filler phrases like “Certainly!”, “Of course!”, “Absolutely!”, “Great!”, “Sure!”, etc.”
Transparent Approach to AI
In the ongoing battle for supremacy in the generative AI space, Anthropic has already unveiled an array of new Claude AI features this year. Only this week it announced that the Artifacts feature that seeks to “turn conversations with Claude into a more creative and collaborative experience” was being rolled out across its tools.
Competing with chatbots like ChatGPT, Microsoft’s Copilot and Google’s Gemini, Anthropic appears to be hoping that its transparent approach will persuade users who may still be sceptical about the veracity of AI technology and put off by the trade secrets closely guarded by those companies.
Users of X replying to Albert’s post were largely positive, thanking Anthropic for the move and praising its transparency.
