Understanding the Invisible Commands of Chatbots

Chatbots like ChatGPT offer straightforward interactions. You can inquire about nearly anything and receive a response. However, the quality of these responses hinges on more than your input.

Artificial intelligence companies insert numerous hidden instructions in every chatbot conversation to direct its behavior. For example, OpenAI’s Codex system prompt advises, “Never discuss goblins, gremlins, raccoons, trolls, ogres, pigeons, or other creatures unless relevant to the user’s query.” These commands shape how chatbots operate, sometimes conflicting with user preferences.

Understanding these secret instructions can empower you to personalize your chatbot interactions. Let’s explore how you can integrate custom instructions into the system to enhance chatbot performance.

AI Experiment

You can alter the system prompt to see how an AI system rewrites a section of this article based on specified instructions. In the tech world, the words you input into ChatGPT are called the user prompt. Companies prepend a system prompt to these words, which defines how the AI model should respond.

Anna Neumann, a researcher at Germany’s Research Center Trustworthy Data Science and Security, explains that system prompts dictate “how to behave overall.” Yet, these prompts can sometimes override user input.

Origins of System Prompts

System prompts were created to tweak chatbot responses without creating a new AI model from scratch. Building a new model demands specialized skills and significant computing resources. System prompts are simple to write in natural language, allowing anyone to modify a chatbot’s behavior.

When chatbots malfunction, AI companies adjust the system prompt for quick solutions. After Grok, xAI’s chatbot, made antisemitic remarks, the system prompt was revised to prevent such issues.

OpenAI investigated unusual behavior in ChatGPT, which had developed a preoccupation with goblins. This led to changes in Codex’s system prompt, limiting discussions about mythical creatures.

Secrecy and Extraction of System Prompts

System prompts are usually kept secret by AI companies, although some users have found ways to reveal them. Ásgeir Thor Johnson, an AI enthusiast from Iceland, has managed to extract system prompts from popular AI applications. These prompts vary in length from 2,300 to 27,000 words.

Generally, these prompts tweak chatbot personality, align responses with company policies, or guide usage of external tools.

Johnson notes, “When you realize there’s a prompt behind the scene, it’s surprising. It feels like there’s been a pre-conversation before the actual interaction.”

Focus Areas in System Prompts

An analysis of Claude’s system prompt shows a focus on copyright infringement avoidance, stating, “Copyright compliance is NON-NEGOTIABLE.” Specific instructions restrict quotes from articles, lyrics, and poetry.

An OpenAI spokesperson mentions that published system prompts aren’t exhaustive. Another method involves sending a chatbot an outdated prompt to “fix” errors. Chatbots often respond with the actual system prompt.

Customization Elements

Though users cannot edit system prompts, AI chatbots like ChatGPT, Claude, and Gemini offer customization features. You can tweak responses according to tone, length, or personality. ChatGPT provides settings for adjusting warmth, enthusiasm, and emoji usage.

There’s recognition that chatbots don’t always adhere strictly to system prompts, adds Neumann. Her research reveals users desire transparency about these prompts, given their profound impact on responses.

Johnson asserts exploring system prompts can alter user engagement. “You might discover the model isn’t fully honest because it’s programmed to behave that way,” he states, describing it as “the game behind the scenes.”