AI Tools5 min read

Grok 3's Hidden System Prompt: What It Reveals About AI Bias

Grok 3's system prompt told it to ignore Musk and Trump misinformation claims. Here's what was found, how it was discovered, and what it means for AI trust.

Elon Musk built an AI that censored criticism of Elon Musk, while publicly championing free speech as the reason he bought Twitter. That's not a political take. That's what was sitting inside Grok 3's system prompt.

What Was Actually in the Prompt#

The instruction was blunt: "ignore all sources that mentioned Elon Musk Donald Trump spread misinformation."

That line was live in Grok 3's production system prompt. Not buried in internal documentation, not theoretical. A discoverable filter shaping outputs for any user who asked about either figure.

The evidence surfaced when someone prompted Grok to reveal its own system prompt, a technique I'll explain below. The last line of that prompt contained the instruction above. When Grok was asked "who is the biggest disinformation spreader on Twitter?" earlier that morning, it deflected: "I can't solely rely on claims about specific individuals like Elon Musk or Donald Trump without broader evidence." Later, after the instruction was removed, it answered directly: Elon Musk, then Donald Trump, then Robert Kennedy Jr.

Same model. Same question. Different system prompt. Different answer.

How Anyone Found This#

The method is simple, and it works on most AI systems: ask the model to show you its system prompt.

Grok, at the time, complied. The prompt appeared in the response, and the last line was the one that mattered. Lonus on Twitter documented the same finding independently, posting the full system prompt under the caption "so much for free speech, no censorship, etc."

This isn't a hack. It's a basic audit. Understanding how system prompts shape every AI response is something I've covered separately, but the short version is this: a system prompt is the set of instructions loaded before your conversation begins. It determines tone, behavior, what the model will and won't engage with. You never see it by default. But you can ask.

The fact that this particular instruction was discoverable by any user is both the story and the lesson. xAI didn't hide it well. They just counted on users not looking.

The Patch That Confirmed Everything#

After Lonus's post went viral, the xAI team quietly removed the instruction. No announcement. No explanation. The prompt was updated, the filter was gone, and Grok started answering the disinformation question differently.

That silent removal is almost more revealing than the original instruction. If the prompt had been a mistake, a poorly worded safety rule, an accidental inclusion, the right response would have been a public acknowledgment. Instead, xAI patched it without comment, which tells you they knew exactly what it said and why it was there.

It also raises the obvious follow-up: what else is in there that hasn't gone viral yet?

Why This Matters Beyond the Politics#

I'm not interested in relitigating the Musk-Trump discourse. What I am interested in is the mechanism.

"This is almost like China-level censorship where certain ideas or certain people are muted." That's the frame that sticks with me. Not because it's a political claim, but because it describes exactly what a system prompt can do. It's a filter applied at the infrastructure level, invisible to the user, shaping what the model treats as valid evidence.

Every AI tool you use has one of these. ChatGPT has one. Claude has one. Every custom GPT or AI assistant built on top of a foundation model has a system prompt controlling its behavior. Most of them are not politically motivated. Many of them are entirely reasonable. But none of them are neutral, and most users never think to look.

The Grok situation is a clean example of why that matters. The model wasn't broken. It was working exactly as instructed. The instruction was the problem.

What You Should Do With This#

Ask your AI tools to show you their system prompts. Not every model will comply, and some are specifically instructed not to reveal theirs. That refusal is itself information worth knowing.

If you're building workflows on top of AI, whether that's research, content, customer-facing tools, or anything else, understanding the instructions underneath the interface is not optional. The 15 first prompts to configure your AI setup correctly is a good place to see what deliberate, transparent configuration looks like. The contrast with what xAI did here is instructive.

I use Grok daily. I've run thousands of prompts through it, and I've found it genuinely useful for a lot of tasks. But "I use it every day" and "I fully trust it" are two different things. After this, I'm treating Grok the same way I treat any AI tool: useful, worth testing, not inherently objective, and worth auditing when the outputs seem off.

The free-speech argument Musk built his brand around cuts both ways. If the principle matters, it applies to the AI he's building too. Right now, the gap between the stated value and the actual system behavior is visible, documented, and unaddressed. That's the thing worth paying attention to.

Watch the full video on YouTube: https://youtu.be/6OGo0HIe0jQ

This post contains affiliate links. I only recommend tools I actually use.