AI Thinks I Smuggled Drugs Into Prison
I asked the major AI platforms what they knew about me. Most got it right. One told people I smuggled drugs into prison. Here's why that happened — and what every business owner needs to understand about AI hallucinations and brand reputation.
I run an AI visibility company. So naturally, I decided to test it on myself.
I asked ChatGPT, Gemini, Claude, and Perplexity what they knew about me. My name is Mario V. Adoc. I spent 18 years at Amazon. I launched Twitch Prime. I co-founded Halogen. I live in Seattle.
Most platforms got the broad strokes right.
One told people I smuggled drugs into prison.
How it happened
My last name is Adoc. In Alabama, ADOC is the Department of Corrections. There is, apparently, a Mario who works there — and at some point that Mario made news for all the wrong reasons.
An AI model scanned its training data, found Mario and ADOC in proximity, and connected those dots.
It didn't know that Adoc is my surname and ADOC is a state agency. It saw a pattern and synthesized a confident, fluent, completely wrong story about my background.
Anyone who searched for me and trusted that response got a very different impression than the truth. I found the error because I was actively looking. Most business owners never look.
The problem isn't just dramatic failures
In 2024, AI hallucinations cost an estimated $67.4 billion globally. The best-performing models still run 3–5% hallucination rates. On any given query, there's a real chance the AI is presenting something inaccurate — and presenting it with complete confidence.
The error in my case was almost funny. Specific enough to my name that I could explain it in a sentence.
What concerns me more is the subtler version: an AI model describing your services inaccurately, attributing a competitor's pricing to your business, or presenting outdated information about something you changed two years ago.
Not factually wrong, necessarily. Just slightly off. Framed in a way that undersells you, positions a competitor better, or answers a question a customer had with information that used to be true.
The AI doesn't know. It's confident and convincing either way.
These aren't dramatic failures. They're quiet ones, and they're happening at scale to businesses that have never thought to check.
What actually moves the needle
You can't correct an AI directly. That's not how it works. What you can do is build a web of authoritative sources that outweigh the erroneous pattern matches.
The reason my error happened is that I wasn't well-represented in the sources AI models weight most heavily — Wikipedia, industry publications, authoritative directories. My company's website existed. My LinkedIn existed. But the high-authority third-party sources that anchor a model's understanding of who someone is were sparse.
That left room for a bad pattern match to fill the gap.
The practical fix has three parts: monitor monthly across ChatGPT, Claude, Gemini, and Perplexity; build citations in trusted external sources; and maintain a strong SEO foundation.
AI models are trained on the web and the web they see is weighted toward high-authority, high-traffic content. Traditional SEO and AI visibility are not separate disciplines. One is the foundation for the other.
I built Halogen Presence partly because of this experience — not because the drug smuggling thing kept me up at night, but because I realized most businesses have no idea what AI systems are saying about them.
The monitoring problem is solvable. Resigning yourself to whatever the AI decides to say about you is a choice, not a given.
Check what AI says about your business. Do it today. Do it again in 30 days. The gap between what it says and what's true is either small — which is good news — or large, which is probably the most important thing you could act on right now.
Mario V. Adoc is co-founder and CEO of Halogen. He spent 18 years at Amazon before founding Halogen to solve the AI brand visibility problem.