Leaked GPT‑5 System Prompt Sparks Debate Over OpenAI’s Inner Workings
Report: A Reddit thread and a GitHub post have surfaced, claiming to reveal the exact system prompt that steers OpenAI’s newest language model, GPT‑5.
What the Prompt Says
- Identity: “You are ChatGPT, a large language model based on the GPT‑5 model and trained by OpenAI.”
- Knowledge Cutoff: The prompt specifies 2024‑06 as the last update.
- Personality Version: It lists “v2,” indicating that the model’s character has evolved.
- Communication Rules: Notably, GPT‑5 is directed to avoid phrases such as “Would you like me to” or “Should I.”
- Interaction Guidance: The prompt includes instructions like “If the next step is obvious, do it” and “Ask at most one necessary clarifying question at the start.”
- Image Generation: Detailed guidelines are given for creating images, including ones that could represent the model itself.
Credibility Concerns
While both sources claim the prompt is authentic, community members on Hacker News question its legitimacy, suggesting that “decoy” or “canary” prompts could be deployed to mislead.
Why This Matters
- Safety and Tone: System prompts set the tone and enforce safety limits.
- Tool Use: They determine how the model interacts with external tools, such as scheduling daily tasks.
- Jailbreak Insight: Leaked prompts could aid attempts to bypass built‑in restrictions.
- Transparency: Knowing the prompt offers a rare glimpse into the decision‑making layers of GPT‑5.
OpenAI’s Official Position
OpenAI’s documentation emphasizes a routed reasoning stack rather than a single static script, casting doubt on the idea that the leaked prompt represents a permanent configuration.
What to Watch For
Future updates to GPT‑5 or subsequent models may alter the prompt, so researchers and developers should remain vigilant for changes in official documentation.
Bottom Line
Whether real or fabricated, the rumored prompt highlights how much influence OpenAI has over the model’s behavior. Its release, if genuine, could shift how users approach and trust GPT‑5, while a hoax underscores the ease with which misinformation can spread in the AI community.