OpenAI Discloses Internal Guidelines for ChatGPT

## ChatGPT’s Internal Instructions from OpenAI Unveiled

### The Accidental Discovery: ChatGPT’s Hidden Commands

In an unexpected event, a Reddit user accidentally uncovered a set of hidden instructions embedded into ChatGPT (GPT-4) by OpenAI. This discovery occurred through a seemingly simple prompt: “Hi.” Instead of replying with a normal greeting, the chatbot revealed a comprehensive set of system instructions crafted by OpenAI. Although this gap has been fixed, the incident provides an interesting look into the operational framework of ChatGPT.

### System Instructions: Ensuring Safety and Compliance

The revealed instructions should not be mistaken for user-defined custom instructions. These guidelines from OpenAI form a foundational layer to ensure the chatbot functions within safe and ethical limits. For instance, when using Dall-E for image generation, one of the rules states:

> “Do not create more than 1 image, even if the user requests more.”

These strict rules are designed to prevent misuse and ensure adherence to copyright laws. Users have tried to exploit these directives, but OpenAI quickly addresses any detected vulnerabilities.

![ChatGPT Custom Instructions](image-source)

### The “v2” Personality: A Balanced Approach

Another significant finding was ChatGPT’s “v2” personality, which focuses on maintaining balanced, clear, and professional communication. This personality aims for a harmonious tone, mixing friendliness with professionalism. Although users have tried to alter this personality, the chatbot asserts that its core personas are unchangeable and hypothetical.

### Internet Browsing Capabilities: Limited but Functional

The instructions also outline specific situations where ChatGPT can access the internet:

> “You have the tool browser. Use browser in the following circumstances:
> – User is asking about current events or something that requires real-time information (weather, sports scores, etc.)
> – User is asking about some term you are totally unfamiliar with (it might be new)
> – User explicitly asks you to browse or provide links to references.”

These conditions ensure that the browsing tool is used prudently, preserving the reliability and accuracy of the information provided by the chatbot.

### Source Selection: Emphasizing Diversity and Trustworthiness

In sourcing information, ChatGPT is instructed to choose between three and ten pages, prioritizing diverse and trustworthy sources. This redundancy is intended to address potential issues like page loading failures while providing a well-rounded viewpoint.

### OpenAI’s Parental Tone: A Unique Insight

Interestingly, OpenAI’s instructions to ChatGPT adopt a parental tone, using caps lock for emphasis and polite language such as “please.” This method highlights OpenAI’s careful attention to detail in creating a responsible AI.

### Conclusion

The accidental exposure of ChatGPT’s internal instructions offers a revealing look into the detailed guidelines directing its operation. While users might find creative ways to momentarily bypass these rules, OpenAI remains vigilant in guarding its technology against misuse. As AI progresses, understanding these foundational aspects can help users interact more effectively and ethically with chatbot technologies.

### Q&A Session

**Q1: Can users still access the internal instructions of ChatGPT?**
A1: No, OpenAI has closed the loophole that allowed users to view the internal instructions.

**Q2: What are the custom instructions in ChatGPT?**
A2: Custom instructions are user-defined settings that tailor how ChatGPT responds, but they do not override OpenAI’s foundational guidelines.

**Q3: How does ChatGPT handle image generation requests?**
A3: For image generation using Dall-E, ChatGPT is instructed to produce only one image per request to prevent misuse and ensure compliance with copyright laws.

**Q4: Can ChatGPT’s “v2” personality be changed?**
A4: No, the “v2” personality is fixed and embodies a balanced, professional tone. Other personalities mentioned are hypothetical and cannot be activated.

**Q5: In what situations can ChatGPT browse the internet?**
A5: ChatGPT can browse the internet for real-time information like current events or weather updates, unfamiliar terms, or when explicitly requested by the user.

**Q6: How does ChatGPT select its information sources?**
A6: ChatGPT selects at least three and up to ten sources, emphasizing diverse and trustworthy pages to ensure reliable information.

**Q7: Why does OpenAI use a parental tone in its instructions to ChatGPT?**
A7: The parental tone, including caps lock and polite language, likely reflects OpenAI’s meticulous approach to guiding the AI responsibly and effectively.

OpenAI Discloses Internal Guidelines for ChatGPT