AI Content Detector Accuracy Is Worse Than You Think

By Frederik Baun, conversion copywriter and founder of Words Sounds Visuals

I ran 13 AI content detectors on camera against three controlled samples in 2025, then retested one of them a year later on fresh samples.

I Tested 13 AI Content Detectors on Camera

The test used three writing samples on the same topic. One was an article I wrote myself in May 2020, years before ChatGPT existed.

The second was ChatGPT writing on the same topic with a plain prompt.

And the third was ChatGPT again. But this time it was trained on my own article first, and asked to match my tone and writing style.

I didn’t run the tests ahead of time or know what would happen.

As I said going in: “My hypothesis is that they don’t [work]. They’re easily fooled by anyone who knows how to prompt AI properly, and they also spit out false positives, which land freelancers in hot waters with clients.”

The Results, Tool by Tool

AI content detector accuracy ranged from perfect to worse than a coin flip, depending entirely on which of these 13 tools you picked.

What Caught the AI Content and What Didn’t

Winston AI was the standout. It flagged my human writing as 99% human. The raw ChatGPT article scored 98% AI. And the voice-trained article came in at 76% AI.

That third result is the one that matters. 10 of the other 12 tools missed it completely.

I said it in the video: “I think Winston AI might be the best candidate for an AI content detection tool. But with an error margin of 24%, it’s still not something that I would rely on.”

The bottom of the table is where it gets uncomfortable.

BrandWell, still branded Content at Scale at the time, called all three samples human. That includes two that ChatGPT wrote start to finish.

AI Detector.com did the same thing. I called it out on camera: “That means Brandwell has a useless AI content detection tool.”

Undetectable AI was worse in a different direction. It markets itself with logos from Forbes, Buzzfeed, and Business Insider and claims 18.5 million users.

And it flagged my own human 2020 article as 99% AI-generated while missing the voice-trained AI article entirely.

That’s the false positive that costs a freelancer a client relationship over content they wrote with their own hands.

Why None of This Matters as Much as You Think

Even the best-performing detector shouldn’t be the reason you accept or reject a piece of writing.

Google’s own developer guidance on AI-generated content points to “high-quality, people-first content.” And it says that standard applies “however content is produced.”

The standard is quality.

Google has been building AI products since long before ChatGPT existed. And they ship Gemini and AI Overviews into search results by default now.

So the idea that they’d penalize AI-assisted writing at the same time doesn’t hold up.

The only useful question is whether the finished piece is worth reading.

A cent-a-word content mill producing keyword-stuffed garbage writes worse copy than someone who knows how to direct AI. No detector percentage changes that.

A Year Later, I Retested BrandWell

Out of curiosity, I ran BrandWell’s free checker again this week. The samples were fresh, with nothing to do with the ones in the video.

I fed it four samples:

• A real human paragraph from a post already live on this blog
• A paragraph of raw AI writing stuffed with clichés
• A second AI paragraph with no obvious tells
• A human paragraph that uses the word “leverage” twice on purpose, just to see if vocabulary alone would trip it up

This time it caught both AI paragraphs correctly and passed both human paragraphs as human. Four for four.

That’s a real change from the tool that missed two out of two AI samples in last year’s test.

Whether that’s a genuinely improved model or just an easier set of samples, I can’t say for certain.

But a detector’s accuracy today doesn’t tell you much about six months from now, in either direction.

The tool that embarrassed itself on camera last year isn’t necessarily the same tool running under that name today.

What This Means for Your Marketing

If a client or job posting requires AI detection clearance, check which tool they’re using.

It could be BrandWell, which called ChatGPT drafts human, or Undetectable AI, which called my own writing a machine’s. Neither outcome tells you anything reliable about quality.

I’d check whether the claims in the copy are true. That matters more than any AI score.

If you want that kind of check, run your draft through the free fact-checker I built.

It won’t score you on AI-likeness. But it will catch the unsourced stats and shaky claims that get copy rejected for a better reason.

If it’s rhythm rather than facts you’re worried about, how to make AI writing sound human is the one to read. It covers the six patterns that give AI away, regardless of what any detector says.