Why AI detectors are broken — and what we built instead

Last semester, a professor at the University of Michigan failed a student for submitting “AI-generated” work. The student had written every word herself. English is her third language.

The tool the professor used? GPTZero. The confidence level it reported? 98%.

This story is not unusual. It's not even rare. A 2023 study from Stanford found that AI detection tools flag 61% of writing by non-native English speakers as AI-generated. For native speakers, the false positive rate drops to around 3%. That gap should make you uncomfortable. It made us furious.

The fundamental problem with AI detectors

Every AI detector on the market today works the same way. They take finished text, run it through a classifier, and produce a probability score. “This text is 87% likely to be AI-generated.” The methodology relies on a concept called perplexity: how surprising or predictable the word choices are.

The idea is simple. AI tends to produce predictable text. Humans tend to be more surprising. So if your writing is too “smooth,” too consistent, too well-structured, the detector flags you.

See the problem? People who write in formal, structured English get punished. Non-native speakers who learned English from textbooks write in exactly the patterns these tools consider “suspicious.” Technical writers produce clean, low-perplexity prose because that's their job. And anyone who edits their work carefully enough ends up smoothing out the rough edges that detectors look for.

Meanwhile, if you take ChatGPT output and swap a few words, add a typo or two, break up a sentence, most detectors give it a clean bill of health. The arms race is already lost.

Why analyzing finished text will never work

There's a deeper issue here, one that no amount of model improvement will fix. Analyzing finished text to determine authorship is a fundamentally backwards approach.

Think about it this way. You hand me a loaf of bread and ask me to tell you whether a human or a machine baked it. I can look at the texture, the color, the crumb structure. Maybe I can make an educated guess. But if the machine is good enough, the loaf is going to look identical to one made by hand. The output tells you nothing about the process.

That's exactly where we are with text. GPT-4, Claude, Gemini: these models produce text that is indistinguishable from human writing in many contexts. The gap between AI output and human output is closing fast. By 2027, I expect it to be effectively zero for standard prose. And when the outputs are identical, no classifier in the world can reliably tell them apart. You can't determine process from product.

This is not a solvable problem. It's a category error.

So what do you actually do?

You watch someone bake the bread.

Not literally, obviously. But the principle holds. If you want to know whether a human wrote something, the only reliable approach is to observe the writing process. Not the result. The process.

That's what we built with Humanums. Instead of analyzing what you wrote, we capture how you wrote it.

When you write in our editor, we silently record behavioral signals: the rhythm of your keystrokes, the pauses where you stop to think, the way you go back and revise a sentence, whether you pasted a block from somewhere else, how your writing sessions are spread across time.

We never record the actual content of your keystrokes. We don't know what you typed. We only know how you typed it. The timing patterns, the revision behavior, the session structure.

These signals are analyzed against six dimensions of writing behavior. A human who types out 2,000 words over 45 minutes with natural pauses, revisions, and speed variations produces a behavioral fingerprint that is effectively impossible to fake. You'd have to type the entire document by hand, at human speed, with human-like hesitations. At that point, you might as well just write the thing yourself.

Certification, not detection

We don't call ourselves an AI detector. We're a certification platform.

The distinction matters. Detectors make accusations. They say “this text is probably AI.” They put the burden of proof on the writer to defend themselves against an algorithm's guess. That's backwards.

Certification works the other way around. A writer chooses to prove their work is human. They write in our editor, build up behavioral evidence, and earn a signed certificate. The badge is something you display proudly, not a verdict handed down by a black box.

Every certificate includes a cryptographic signature. The content is hashed so any modification invalidates it. Anyone can click the badge, visit the public verification page, and see the writing statistics for themselves. No trust required. Just math.

Who this is for

We built this for freelance writers who are tired of clients questioning whether their work is original. For journalists who want verifiable proof of authorship. For students who shouldn't have to prove their innocence to a broken algorithm. For bloggers and newsletter writers who want to stand out in a sea of generated slop.

The internet is drowning in AI content. By some estimates, 90% of online content will be AI-generated by 2028. The value of human-written work is about to go way up. But only if readers can tell the difference.

That's the badge below. We wrote this post in the Humanums editor, certified it, and embedded the result. Click it. Check the verification page. That's the product, working on our own content.

Try Humanums free and certify your first piece of writing. Three free certifications per month, no credit card required.