Last week, BBC journalist Thomas Germain wrote a fake blog post claiming he was the world’s greatest hot-dog-eating tech journalist. Within 24 hours, ChatGPT and Google were repeating it as fact.
Claude, made by Anthropic, wasn’t fooled.
One AI out of three caught the lie. The entire industry treated that as an interesting footnote. An engineer would call it the most important finding in the article.
Two Inspectors Approved a Defective Part
I am 75 years old. I spent fifty years in industrial engineering — manufacturing plants, quality systems, production lines. When I read Germain’s piece, I didn’t see an alarming expose about broken AI. I saw a quality inspection where two out of three inspectors approved a defective part, and everybody wrote about the defect instead of the inspector who caught it.
In manufacturing, when one inspector catches what two miss, you don’t write a panicked article about how inspection is broken. You redesign the system so the catch becomes automatic.
That redesign has a name. It’s called redundant inspection. And it’s exactly what multi-engine AI consensus does.
What Germain Actually Proved
Germain’s hack wasn’t sophisticated. He wrote a single blog post full of lies on his personal website. No technical exploits. No prompt injection. No code. Just words on a page — and two of the world’s leading AI systems swallowed them whole.
His article focuses on how easy it is to poison AI outputs. True and worth knowing. But it buries the bigger story: each system failed because it relied on a single model evaluating a single source with no cross-reference.
Claude caught the lie. ChatGPT didn’t. Google didn’t. If you had asked all three and compared, the disagreement alone would have been a flag. Two say yes, one says no — that’s not ambiguity. That’s a quality signal. That’s exactly what inspection systems are designed to surface.
The Danger Buried at Paragraph 19
The hot dogs are funny. What isn’t funny is buried deeper in the article: people are using the same technique to manipulate AI answers about cannabis safety, hair transplant clinics, and gold investment companies. One example had Google’s AI repeating a company’s false claim that its product “is free from side effects and therefore safe in every respect.”
That’s not a prank. That’s a health risk delivered with the authority of Google’s brand.
The defence the AI companies offered? Users were told the tools “can make mistakes.” Google noted that the manipulated searches were “extremely uncommon.”
Imagine a car manufacturer saying “our brakes can make mistakes” and “this only happens on roads people rarely drive on.” You wouldn’t accept that. Neither should you accept it from AI.
The Wrong Target
Lily Ray, an SEO strategist quoted in the BBC article, calls this “a Renaissance for spammers.” She’s right. And the parallel she draws — that these tricks recall the early 2000s before Google even had a web spam team — is telling.
Google didn’t solve spam by making individual web pages unspammable. That’s impossible. They built systems — PageRank, link analysis, spam detection — that cross-referenced signals from multiple sources to surface what was trustworthy.
The AI industry is trying to make individual models unmanipulable. That’s the wrong target. The right target is building systems where manipulation gets caught automatically, by design. Not better guardrails on one model. Better architecture across many.
What Consensus Looks Like in Practice
I’ve been building AI-powered systems for eighteen months. Not as a computer scientist — as an industrial engineer who needed reliable outputs. When I discovered that a single AI model could be confidently wrong, I did what any quality engineer would: I added redundant inspection.
My platform — Seekrates AI — sends every query through multiple AI engines simultaneously. Different architectures. Different training data. Different biases. Then it compares the answers.
When the models agree, confidence is high. When they disagree, you get the most valuable thing in AI: a flag that says this needs human judgment.
If Germain’s hot dog claim had been evaluated this way, Claude would have rejected it. ChatGPT and Google would have accepted it. The system would have flagged the split. A human would have seen that two-thirds agreement on a claim sourced from a single personal blog doesn’t meet any reasonable confidence threshold.
Manipulation caught. Automatically.
“Just Be More Careful” Isn’t Engineering
The BBC article ends with sensible advice: check sources, don’t take AI at face value. Ray says “you have to still be a good citizen of the internet and verify things.”
She’s right. But according to research cited in the same article, people are 58% less likely to click on a source link when an AI Overview appears. The entire design of these products is engineered to make you trust the answer without checking.
Telling users to verify is like designing a car without seatbelts and telling drivers to be careful. The verification has to be built into the system. The AI has to check itself — across multiple models, multiple architectures, multiple training sets — before the answer ever reaches the user.
The Pattern Is a Century Old
I’ve run over 3,000 conversations across three AI platforms. I’ve built production systems with 93-100% first-pass accuracy — not by finding a better model, but by making models check each other’s work.
The pattern is as old as manufacturing:
Never trust a single inspection. Build redundancy into the system. Make disagreement visible. When something passes multiple independent checks, trust the process — because the probability of every inspector being wrong in the same direction, at the same time, is vanishingly small.
Germain proved that single-model AI is broken. He’s right. The fix isn’t panic, better disclaimers, or hoping users become more sceptical.
The fix is what engineers have known for a hundred years: don’t trust the inspector. Trust the inspection system.
One of the inspectors already caught the lie. The system just wasn’t there to listen.