Paste the same paragraph into three AI detectors and you might get three different answers: 90% AI, 40% AI, "likely human." That isn't a bug in one tool. It's built into how detection works. Here's what's going on under the hood, and how to read a score without overreacting.
What detectors actually measure
Most AI detectors don't "recognize" AI the way a person recognizes handwriting. They measure statistical patterns, mainly two:
- Predictability (often called perplexity). Language models write by choosing likely next words. Text where every word is very predictable looks machine-made. People make more surprising choices.
- Variation (often called burstiness). People write unevenly: a long sentence, then a short one, then a fragment. AI output tends to be uniform, with similar sentence lengths and the same structure over and over.
Many detectors also use a classifier trained on examples of human and AI text, which picks up habits like stock phrases ("it is important to note"), balanced three-item lists, and template paragraphs.
The output is a probability estimate, not proof. "80% AI" means the text resembles what the model learned to call AI. It doesn't mean 80% of it was written by a machine.
Why scores vary so much
Different tools, different training. Each detector was trained on different data, with different models and thresholds. They disagree in the same way two weather apps do.
Short text is unreliable. With only a few sentences there isn't enough signal, so scores on short passages swing wildly.
Some human writing looks like AI. Formal, formulaic writing scores as "AI" more often: legal text, technical documentation, textbook-style explanations, and writing by non-native English speakers who stick to safe, common phrasing. False positives are real, and they hit some writers harder than others.
Some AI writing looks human. Lightly edited AI text, or text prompted toward a casual style, often passes.
Genre matters. Sales copy is full of predictable hype phrases, so it tends to score high whoever wrote it.
Models keep changing. Detectors are always chasing newer models, so a score from one month may not match the next.
Honest limits
- No detector is proof. A score should start a conversation, not settle one. Even detector makers generally warn against using scores alone for serious decisions.
- No humanizer can guarantee a pass. Anyone promising "100% undetectable on every detector" is overselling. The text changes, the detectors change, and nobody outside a company like Turnitin can test against its system.
- A low score doesn't mean good writing. You can game a detector and still produce something nobody wants to read.
Checking before and after
If you edit AI-assisted drafts, it helps to see whether your changes actually shifted the patterns detectors look for. FreeHumanizerAI (https://freehumanizerai.com) has a built-in AI detector that shows a before and after score next to the rewrite, so you can compare without copying text between sites.
The site is also open about its limits. Its published ZeroGPT tests show an AI-written essay going from 100% to 0% AI after its Ultra mode, and a business email scoring 0%, while a product description still scored 70%. It notes that results depend on the text and that no humanizer can promise 0% on every detector. That kind of mixed result is what you should expect from any tool.
How to use a score sensibly
- Use longer samples. A few paragraphs give a more stable reading than one.
- Check more than one detector if it matters, and look at the overall trend rather than one number.
- Look at which sentences get flagged. Those are usually the generic ones, and they're worth rewriting anyway.
- Fix the writing, not the score. Add specifics, vary sentences, and cut filler. Readers reward the same changes that move scores.
- Follow the rules where you are. Schools and publishers have AI policies. Disclose AI help where it's required. A low score doesn't change what the policy says.
The bottom line
AI detectors estimate how statistically "machine-like" a text is. They're useful as a rough signal and unreliable as a verdict. Read the score as a hint about where your writing is generic, then make those parts more specific.
Top comments (0)