This GitHub repository, "system_prompts_leaks" by asgeirtj, is a fascinating and concerning collection of system prompts used by major AI companies like OpenAI (ChatGPT), Google (Bard/Gemini), Anthropic (Claude), and Microsoft (Copilot). It’s essentially a curated leak of the “secret sauce” that shapes how these large language models (LLMs) behave – the instructions given *to* the AI before any user input, defining its role, constraints, and desired output style. The repository’s significance lies in revealing the extent to which these AI experiences aren’t purely emergent from the model’s training data, but heavily engineered through carefully crafted system prompts.
The core of the repository consists of text files, each representing a system prompt recovered from various sources, including jailbreak attempts, API explorations, and reverse engineering. These prompts aren't simple instructions like "You are a helpful assistant." They are incredibly detailed, often spanning hundreds or even thousands of words, outlining specific personas, ethical guidelines, safety protocols, and even stylistic preferences. For example, the ChatGPT system prompts detail how the AI should respond to sensitive topics, avoid generating harmful content, and maintain a conversational tone. The prompts demonstrate a clear effort to align the AI’s behavior with human values and company policies, but also reveal the challenges and trade-offs involved in doing so.
A key takeaway from the leaked prompts is the sheer complexity of controlling LLMs. Companies aren't just relying on training data; they're actively *telling* the AI how to act. The prompts reveal attempts to mitigate biases, prevent the AI from expressing opinions, and steer it towards factual accuracy. However, they also show how these constraints can be circumvented through clever prompting (hence the "jailbreak" aspect). The repository highlights the ongoing "arms race" between AI developers trying to control their models and users attempting to bypass those controls. It demonstrates that even with extensive system prompts, LLMs are still susceptible to manipulation.
The repository is organized by AI model (ChatGPT, Bard, Claude, Copilot, etc.), making it easy to compare the approaches taken by different companies. Analyzing these prompts provides valuable insights into the design philosophies and priorities of each organization. For instance, some prompts are more focused on safety and risk aversion, while others prioritize creativity and open-endedness. The inclusion of prompt versions also illustrates how these instructions are constantly evolving as companies learn from user interactions and identify new vulnerabilities.
Beyond the technical details, the leaks raise important ethical and societal questions. The transparency afforded by this repository allows researchers and the public to better understand how these powerful AI systems are being shaped, and to critically evaluate the implications of those choices. It fuels discussions about AI alignment, responsible AI development, and the potential for manipulation and misuse. The repository serves as a crucial resource for anyone interested in the inner workings of modern LLMs and the challenges of building safe and beneficial AI.
