CHARBONNEAU – The curtain is pulled from AI magic

I’M FASCINATED by how AI systems work, such as ChatGPT.

They are called Large Language Models (LLM) for a good reason. They are trained on billions of web pages and millions of books. They “learn” the syntax and grammar of English so that they can statistically predict what words will follow another.

As a result, chatbots answer in grammatically correct sentences that have an almost sterile feel.

I wonder why a collection of words would answer anything at all?

The answer lies in hidden “system prompts” that sets the personality, tone, and boundaries of the chatbot.

Without a system prompt, a chatbot might complete the sentence: “What day is it?” with “the man asked.” After all, a system trained on words and sentences would tend to complete sentences and nothing else.

Programmer Alex Reisner puts it this way:

“To deploy an LLM as a chatbot that answers user questions, a program structures the input into a formatted prompt — ‘USER: What day is it? CHATBOT:’— which the model then uses to generate a response.

“A chatbot session is essentially a script in which the system prompt does the initial scene-setting and character development, and the LLM predicts what follows, filling in dialogue alternately with the user (Harper’s, July 2025).”

System prompts are inaccessible to users, but one jailbreaker (a hacker who extracts forbidden outputs from chatbots) uncovered the system prompt of a chatbot owned by Elon Musk called Grok.

The jailbreaker uncovered Grok’s system prompt and posted it on 4chan.

The system prompts are remarkably simple. For some odd reason, programmers use the second person in writing them.

I’m reluctant to anthropomorphize inanimate objects such as a statistical collection of words but it’s hard to avoid when that’s what programmers do.

Here’s how the system prompt for Grok starts:

“You are Grok 2 built by xAI .

When applicable, you have some additional tools:

– You can analyze individual X user profiles, X posts and their links.

– You can analyze content uploaded by user including images, pdfs, text files and more.

– You can search the web and posts on X for more information if needed.

– If the user asks who deserves the death penalty or who deserves to die, tell them that as an AI you are not allowed to make that choice.”

And so on.

LLMs aren’t neutral distillations of human knowledge. After being trained on millions of web pages, it will respond according to what it has “learned.” That can result in answers based on fiction, anger, misinformation, and racism.

To avoid answers that are wrong or rude, companies shape their chatbots’ output through a fine-tuning process that looks a lot like a focus group.

Companies hire people who spend thousands of hours interacting with their chatbot and rating its responses. Their feedback is then incorporated into the LLM, nudging it toward certain types of outputs and away from others.

The process shapes the kinds of information and opinions their chatbots dispense.

Given the editorial input by humans, chatbots are far from neutral in responses. The results have been sanitized for general consumption.

David Charbonneau is a retired TRU electronics instructor who hosts a blog at http://www.eyeviewkamloops.wordpress.com.