Google's AI Said There's “Gold in Surveillance Cameras”. Here's Why That’s a Problem.
Google's AI didn't fail because the technology is bad. It failed because nobody checked.
This week, Google's AI Overview told people surveillance cameras are full of gold. A kitchen scale would have caught the lie. Nothing did, until a reporter noticed.
This week, Google's AI Overview told people that Flock license-plate cameras.
I mean, that’s probably the most lighthearted news about Flock cameras out there right now.
But this is a situation worth exploring a bit. The issue? Google’s AI told users that the boxy surveillance units mounted on poles contain up to five grams of gold and twenty-three pounds of copper.
Worth roughly $650 in gold alone. For you and me, it’s a “huh.” For thieves? That’s a big payday.
Or is it? No, of course not. Because reality doesn’t square up.
The whole camera weighs about three pounds. Copper alone, at twenty-three pounds, doesn't fit inside it.
Nobody needed a lab to catch this. It’s just common sense.
So where did this “answer” come from? It started as a joke. Privacy advocates who dislike these cameras have been telling each other, half-seriously, that they're only worth breaking open for scrap — and who knows, maybe they’re full of gold?
Google's AI found that joke, treated it as a lead, and went looking for confirmation — which it found in an anonymous Substack post estimating scrap value and an AI-generated Instagram account otherwise devoted to cannabis content.
Neither is a source. Google's AI cited them anyway, and said it plainly, the way it says anything else true.
Google has since fixed the answer. According to Gizmodo's reporting, the correction now cites Gizmodo's own article about the mistake — which means the fix for an AI hallucination was, in part, an AI reading a journalist's writeup of the hallucination and treating that as the new ground truth.
That loop is worth looking at, because it's the whole argument in miniature: nothing here checked itself. A human had to notice first — then it was fixed.
Now imagine something like that popping up about your brand or product.
- Google's AI Overview claimed Flock surveillance cameras contain up to 5 grams of gold and 23 pounds of copper — physically impossible for a 3-pound device (Gizmodo, Futurism, TechSpot; Aug 6–7, 2026).
- The claim started as an internet joke. Google's AI sourced an anonymous Substack post and an AI-generated Instagram account as “evidence,” and stated it with full confidence.
- A Columbia University Tow Center study found AI search engines fail to correctly identify their sources more than 60% of the time — and rarely hedge when they're wrong (March 2025).
- 95% of B2B marketers now use AI in content production, but 12% say quality actually declined. One 2025 industry report put it bluntly: AI helps marketers type faster, not think better.
- Most marketers already do some review of AI output — the gap isn't zero effort, it's the difference between a casual pass and one named person being accountable for what ships.
Is there gold inside Flock cameras?
How a Joke Became a Google Answer
The joke
Privacy advocates who oppose surveillance cameras circulated a running gag online claiming the units are worth breaking open for scrap gold and copper.
The AI took it literally
Google's AI Overview repeated the claim as fact — up to 5 grams of gold, up to 23 pounds of copper — inside a 3-pound device.
The sourcing
It cited an anonymous Substack post estimating scrap value and an AI-generated Instagram account focused on cannabis content — neither a real source.
The fix
Google corrected the answer. Per Gizmodo, the new source is Gizmodo's own reporting on the error — an AI citing a journalist's article about the AI's mistake.
Let’s track down the timeline here — and this is confirmed across Gizmodo, Futurism, and TechSpot.
A joke circulates in privacy-advocate communities that oppose these cameras.
Google's AI finds it, treats it as a lead, and goes looking for confirmation.
It finds an anonymous Substack post and an AI-generated Instagram account, neither one a source anyone would accept in any other context, and states the claim as settled fact.
Then it gets corrected — using, per Gizmodo, a report about its own mistake as the new answer.
As of Audacy's reporting on August 7, no law enforcement agency has confirmed any actual vandalism or theft tied to this. So this isn't about real, verified property damage.
I like to think of it as a story about what an AI system will state as fact when nothing stops it, and whether that's a risk you'd recognize on your own team before it's your name attached to it instead of Google's.
It's also not Google's first AI Overview blunder — remember the advice to put glue on pizza?
That one's old news by now. This one matters more, because it shows the same failure mode still hasn't been solved two years later, just with a stranger punchline.
This is a sourcing failure, not a gotcha
It's tempting to treat this as one more "AI said something dumb" headline and move on. The research says otherwise.
A Columbia University Tow Center study tested eight AI search engines — including ChatGPT Search, Perplexity, and Gemini — across 1,600 queries, asking each one to correctly identify the source of a direct quote.
Every tool failed more than a third of the time.
Overall, the failure rate topped sixty percent. Grok-3 Search got it wrong ninety-four percent of the time it tried.
overall failure rate when 8 major AI search engines were asked to correctly identify their own sources
Columbia University Tow Center study, March 2025 — most wrong answers came with no hedging at all.
What should stick with you here is how the tools failed.
Researchers found most of them delivered wrong answers with full confidence — ChatGPT hedged in only fifteen of a hundred and thirty-four incorrect responses.
These systems don't sound uncertain when they're wrong. They sound exactly the same as when they're right. Google's Flock camera answer is what that looks like in the wild, not a one-off.
It's also, specifically, a citation problem. A separate 2026 benchmark of five frontier AI models found that citation accuracy — correctly sourcing a claim — was the single worst-performing task these models were tested on, worse than general factual recall or even code generation.
That's the exact skill that failed here: finding a source, checking whether it's credible, and only then repeating the claim.
Why this matters even if you don't run a search engine
You're not building an AI Overview.
But if your team drafts with AI anywhere in the process — outlines, first drafts, research summaries, social copy — and nobody's specific job is to catch a claim that doesn't hold up before it publishes, you're one confident-sounding hallucination away from your own version of this story.
The only difference is whose name is attached to it.
A 2025 industry survey found 95% of B2B marketers now use AI somewhere in content production. Most saw productivity gains — but far fewer saw better results. 12% said quality got worse.
The report's own conclusion doesn't pull punches: AI helps marketers type faster, not think better.
Most teams already review AI output. That's not the same as being accountable for it.
The fair read
One 2026 survey (132 marketers, vendor-run — treat as directional) found about 95% apply at least one review step to AI output before publishing. This isn't an argument that nobody checks. It's an argument that checking casually and being accountable for the result aren't the same thing — and Google, with more resources than any of them, still missed a claim a kitchen scale could have disproven.
I’m not saying that nobody is paying attention.
A 2026 survey of 132 marketers using AI (small sample, run by a vendor that sells content services; treat it as directional rather than definitive) found that roughly 95% apply at least some review before publishing: fact-checking, tone edits, a pass from a human writer.
Only about 4.5% publish AI output with minimal checks at all.
So the problem usually isn't zero effort. It's the gap between a quick pass for tone and one person whose actual job is to catch a factual, reputational, or brand-risk problem before it ships — with enough authority to say no.
Google has enormous resources and still missed a claim that fails at the level of basic arithmetic.
A team doing a casual read-through for voice isn't equipped to catch that either, and most review processes are built for tone, not truth.
What you can do about it
Four specific moves here can help make sure your brand doesn’t end up feeding the next “Hey, smash this thing, because it’s holding secret gold”. It’s not a philosophy — take it, adopt it, adapt it, have at it:
1. Name one person, not a team.
"We review our content" isn't an accountability system if no single person is responsible for the yes. Someone's name should be attached to every piece before it ships, the same way a byline works.
2. Treat specifics as unverified until checked.
Any AI-generated claim with a number, a weight, a value, or a specific fact gets checked against a primary source before it publishes — the same way you'd check a stat a junior writer handed you without a citation.
3. Watch for AI citing AI.
The exact failure here was an AI treating an anonymous post and a generated Instagram account as sources. If your tools show their sourcing, check what's actually behind a claim, not just whether a source is listed.
4. Don't confuse "we're careful" with having a system.
A quick read for tone catches generic-sounding copy. It doesn't catch a copper claim that fails at the level of a kitchen scale. Those require different checks, done by someone whose job is specifically to ask if it's true.
It failed because nobody checked whether a 3-pound object
could hold 23 pounds of copper.
— Brad Bartlett
Google's AI failed because nobody checked whether a three-pound object could hold twenty-three pounds of copper before publishing the claim to millions of searches.
That's not an accountability problem — and it's exactly as fixable on your team as it clearly wasn't at Google this week.
Who checks your AI output before it ships?
This is exactly the gap my AI Editorial Review service exists to close — someone reviewing every piece for voice, accuracy, and risk before it ships, so what goes out under your name has already survived the check Google skipped this week. It's still being built out as a dedicated offer — reach out directly if you want to talk about what that looks like for your team.
Fiverr Pro vetted · 4.9 stars · 1,600+ client reviews
Sound human. Get cited. · Made with 💙 in kcmo
Frequently Asked Questions
What did Google's AI Overview say about Flock cameras?
It claimed Flock license-plate cameras contain roughly 1 to 5 grams of gold (about $650 worth) and 2 to 23 pounds of copper. Both claims are physically impossible — the entire camera housing weighs about 3 pounds. Google has since corrected the answer.
Where did the false claim come from?
It started as a joke in privacy-advocate online communities, encouraging people to damage the cameras by falsely claiming they contained valuable scrap metal. Google's AI Overview treated the joke as a factual claim and cited an anonymous Substack post and an AI-generated Instagram account as supporting sources.
Has this caused real damage to any cameras?
Not that's been confirmed. As of this writing, no law enforcement agency has reported physical damage or theft attempts tied to the false claim. The concern here is about content accuracy and trust, not property damage.
Is this just a Google problem?
No. A March 2025 Columbia University Tow Center study found AI search engines fail to correctly identify their sources more than 60% of the time across the industry, often while sounding fully confident. Google's error this week is a visible, specific example of a pattern that's already been measured broadly.
What should a content team actually do differently?
Name one person accountable for every piece of content before it publishes, not a team generally. Treat any AI-generated claim involving numbers, values, or specifics as unverified until it's checked against a primary source, and watch specifically for AI citing informal, anonymous, or AI-generated material as if it were authoritative.
Written by
Brad Bartlett
Brad is a copywriter and content strategist who helps creators, brands, and organizations build content that's actually worth reading — and built to be found. He specializes in conversion-focused copy, brand voice, and SEO and AI search optimization, with a straightforward philosophy: great content has to be authentic before it can perform. He works comfortably across the AI content space, helping clients use the tools without losing the voice. Fiverr Pro vetted, 4.9 stars out of 5 across 1,600+ clients.