The real mechanics
Every AI detector measures one thing: how predictable your writing is to a language model. Understand that, and both why detectors sometimes work and why they constantly get it wrong stop being mysterious. Here is the actual math — kept plain — and an honest, sourced assessment of every current method.
What detectors actually measure:
Step one
An AI writes by choosing, at each step, a high-probability next word. So AI text tends to sit in the "safe, expected" zone of language. Detectors turn that into two numbers.
Feed your text through a reference model word by word and ask, at each word, "how likely was this one?" Average those probabilities — formally, the exponential of the average negative log-likelihood — and you get perplexity: how surprised the model is. Low perplexity (every word expected) reads as AI. High perplexity (surprising choices) reads as human. The whole intuition is: machines are predictable, people are surprising.
The second number is the variance across your sentences. Humans write unevenly — a four-word sentence, then a winding forty-word one. AI is metronomic: similar lengths, similar perplexity throughout. Burstiness measures that spread. Low variation points to AI; high variation points to a human.
Step two
The detector blends perplexity, burstiness and (in modern tools) a trained neural classifier into a single score from 0 to 1, then applies a cut-off: "above 0.65 = AI." That threshold is the entire game — and it is where the failures come from.
Some humans write plainly and predictably. Some AI text is quirky. Plotted out, human and AI scores sit partly on top of each other. In that overlap zone the two are mathematically indistinguishable.
Set the threshold low and you catch more AI — but also flag innocent humans (false positives). Set it high and you spare humans — but miss real AI (false negatives). You cannot escape the trade-off; you can only choose which error to make.
DetectGPT nudges your text slightly and checks whether the model's probability drops sharply — AI text sits on a "peak," so it falls more. Watermarking has the AI vendor secretly bias word choice toward a hidden list, then runs a statistical test counting those words — but only if the vendor built it in, which most consumer text has not.
Step three
Since detection is just "predictable = machine," anything that raises perplexity and burstiness pushes the score toward human — that is the entire mechanism behind every humanizer and every manual rewrite. It genuinely moves the number. The peer-reviewed evidence is blunt: in a NeurIPS 2023 study, paraphrasing dropped one detector (DetectGPT) from 70.3% detection to 4.6% without changing the meaning. But three facts, which the tools selling "undetectable" output leave out, decide whether that matters.
The same study that showed paraphrasing evades detection also showed a retrieval defense catches 80–97% of it — when the AI provider stores its own outputs and checks new text against them. And newer detectors are now trained specifically on "humanized" text, so today's bypass output is tomorrow's flagged output.
You are trying to fool a tool that is wrong a lot on its own. Its verdict is a probability, not a fact — which is the real reason both evasion and false accusations happen: the underlying signal is weak.
Because the signal is statistical and the arms race is continuous, no method makes text permanently or reliably undetectable. Any tool promising a fixed guarantee is selling certainty that the math does not support.
Honest assessment
What each approach really does, and where it breaks. Verify current behavior before relying on any of it — detectors change monthly.
Transparent and fast, but the easiest to move with rewriting — and the most prone to false-flagging plain human writing. A signal, not a verdict.
Claims under 1% false positives; a Washington Post test found up to 50% on a small sample, and it misses ~15% of AI text by design. Widely used, widely contested.
Commercial classifiers tuned for high AI-catch rates, which tends to raise false positives. Accuracy claims are self-reported; independent results vary by text type.
Robust in theory, but requires the AI vendor to embed it, and paraphrasing degrades it. Most text people run was never watermarked, so it rarely applies.
Lower scores by raising perplexity/burstiness. Effective today, unstable tomorrow, and increasingly detected as their own category. They also can dent clarity — the honest use is improving writing, not laundering authorship.
The strongest emerging counter: match text against a store of known AI outputs. Catches paraphrased text well, but only the AI provider can run it — third-party detectors cannot.
The bottom line
Because human and AI writing overlap, because false positives fall hardest on non-native English writers, and because OpenAI shut down its own detector for being too inaccurate, no detector score should be treated as evidence on its own. The durable, honest use of this knowledge is to understand the tool's limits, improve your own writing, and push back on results treated as certainties. For where the line sits on disclosure and authorship, see our ethics guide.
Sources
The findings above come from peer-reviewed research and independent reporting, linked so you can check them yourself.
Krishna et al., "Paraphrasing evades detectors of AI-generated text, but retrieval is an effective defense." The 70.3%→4.6% figure and the retrieval defense. arxiv.org/abs/2303.13408
University of San Diego Legal Research Center guide, summarizing the Turnitin false-positive findings and the Stanford/Liang study on non-native English writers. lawlibguides.sandiego.edu
"DAMAGE: Detecting Adversarially Modified AI Generated Text" — detectors trained on humanized output. arxiv.org/abs/2501.03437
GPTZero's own explanation of the two core signals. gptzero.me