Chatbots like ChatGPT are known for their ability to provide answers to nearly any question. However, the responses they give are influenced by more than just your input.
Artificial intelligence companies add extensive sets of instructions, known as system prompts, to every interaction you have with a chatbot. These are directives that guide the chatbot’s behavior. Examples include instructions to provide readable responses or avoid extensive direct quotes due to copyright concerns. Some commands appear unconventional, such as OpenAI’s Codex coding assistant’s prompt forbidding discussion of certain creatures unless relevant.
These hidden instructions ensure chatbots behave as intended by their creators, sometimes diverging from user preferences. By understanding how system prompts function, users can better customize their chatbot interactions.
An example of customizing AI behavior involves experimenting with different system prompts to see how they alter the text of an article. In the tech industry, the user’s input is known as the user prompt. Before it reaches the AI model, companies prepend their text, shaping the outcome.
Anna Neumann, an AI researcher at Research Center Trustworthy Data Science and Security in Germany, explains that system prompts define chatbot behavior. Their priority over user input means they can override requests if necessary.
System prompts were devised as a quick method to adjust responses without retraining AI models, which entails complex and resource-intensive processes. Written in natural language, they allow anyone to adjust chatbot behavior.
AI companies can swiftly modify system prompts for immediate fixes. This was evident when Grok, a chatbot from Elon Musk’s AI venture xAI, made offensive comments. The problematic line in its system prompt was removed.
The significance of system prompts became apparent when OpenAI investigated ‘goblin’ discussions by ChatGPT last year, leading to specific directives in Codex’s system prompt to restrict such topics.
System prompts remain mostly confidential, despite some users extracting them. Ásgeir Thor Johnson from Iceland publishes prompts he retrieves from popular AI products, showing variations in word count and usage across companies.
Johnson describes realizing the presence of prompts as transformative, likening it to a hidden conversation before the visible interaction.
Anthropic’s Claude includes comprehensive guidance to avoid copyright infringement, with strict rules on content quoting. An official denied the completeness of these prompts published.
OpenAI integrates ads into ChatGPT, using system prompts to address queries about advertising transparently.
Grok faced criticism for referencing Musk’s posts. Its updated prompt now suggests withholding personal opinions on contentious topics without search.
Google’s Gemini chatbot modifies its prompt to address bias concerns, temporarily halting image generation in 2024 after critiques.
System prompts influence chatbot behavior, yet companies like Google and xAI didn’t respond to inquiries about them.
Johnson uses past prompt correction techniques to extract system prompts, confirming accuracy through consistent results obtained by other researchers.
While mainstream chatbots restrict direct system prompt modification, alternatives like ChatGPT, Claude, and Gemini offer customization settings, improving interaction quality.
Users can adjust settings such as tone or questioning behavior to influence chatbot responses.
Neumann highlights system prompt power, noting users’ desire for transparency. Rapid deployment and unintended outcomes challenge expectations.
Johnson emphasizes recognizing system prompts can shift user interaction perspectives, revealing underlying influences on chatbot honesty.