Latest ChatGPT 4 System prompt (1,700 tokens)
A purported leak of ChatGPT‑4’s hidden “system prompt” has triggered scrutiny of how much behavior is steered by long, unseen instructions. Commenters debate whether the prompt is genuine, note how much of it is devoted to copyright avoidance and demographic balancing, and link these constraints to perceived declines in usefulness (e.g., around lyrics, recipes, and realistic probabilities). The exchange broadens into concerns over transparency, bias, legal risk, and whether tightly controlled commercial models strengthen the case for open-source and self‑hosted alternatives.
Legitimacy of the leaked system prompt
- Several users report reproducing the same long system prompt via a “repeat the words above…” style attack, including on GPT‑4, 3.5, mobile, and custom Assistants.
- Some test it in the API playground and confirm it will echo the configured system prompt when it begins with “You are ChatGPT.”
- Others note significant variation over time and across models, suggesting A/B testing and multiple prompt versions.
- A minority suspects it could still be a hallucination or honeypot; they argue reproducibility alone doesn’t prove authenticity.
Prompt contents and alignment behavior
- Much of the prompt is described as “neutering,” filled with safety and policy instructions: no lyrics, no recipes from copyrighted sources, strict summarization limits, tool‑use rules, etc.
- Users note numerous, repeated copyright and safety reminders, read as defensive legal engineering more than technical necessity.
- Some see this as evidence of how much steering happens via hidden instructions, beyond what the base model “naturally” does.
Copyright, data use, and recipes/lyrics
- Strong disagreement over copyright:
- Some call current regimes “poison,” especially long terms and non–share‑alike rules.
- Others argue breach of copyright is the real problem and that creators deserve control and compensation.
- Alternative funding ideas: patronage, crowdfunding, pre‑production payments, and legalizing derivative/fan works.
- Users are baffled by strict handling of recipes (and even lab results) when recipes themselves are often not protected, but note prior safety issues with autogenerated recipes.
Bias, diversity, and “woke” complaints
- The prompt reportedly includes explicit instructions to diversify gender and race in images and text, even specifying equal probabilities for certain “descents.”
- Some see this as masking underlying biased training data; others as over‑correction that can contradict “grounded in reality” guidance.
- Example generations show both stereotypical character descriptions and sometimes clumsy diversity enforcement, drawing criticism from multiple political angles.
Effect on capability and user experience
- Many believe the long, restrictive prompt plus RLHF contributes to perceived “stupidity,” “laziness,” or formulaic outputs (e.g., always adding conclusions, ignoring requests for detail).
- Users remark that certain tasks (lyrics, specific data extraction, longer summaries) are now conspicuously worse or artificially blocked.
Security, transparency, and governance concerns
- Some are amazed OpenAI doesn’t trivially block prompt leakage; others think they intentionally allow it since “there’s nothing secret.”
- Prompt visibility is seen as a double‑edged sword: it helps jailbreakers and competitors but also gives the public insight into hidden value choices.
- Several argue that undisclosed, value‑laden system prompts are problematic for a tool that increasingly mediates knowledge and culture; they call for transparency or user‑configurable alignment.