How AI text detectors work, and where they go wrong

October 28, 2026 · 6 min read · ZeroClaude team

Every few weeks a new language model makes headlines, and every few weeks someone asks the same question: can you actually tell when a machine wrote something? The honest answer is "often, but not always". To understand why, it helps to know what a detector is really looking at.

Models write the likely next word

A language model produces text one token at a time, and at each step it leans towards words that are statistically likely in that context. People do something similar, but much less consistently: we pick an odd word because it sounds right to us, we start a sentence and change our mind halfway, we repeat ourselves when we are tired. Detectors try to measure that difference in predictability.

The signals a detector combines

• Predictability. How surprising each word is to a reference model. Very low surprise across a whole passage is typical of generated text.
• Burstiness. Human writing mixes short and long sentences and changes rhythm. Generated text tends to keep an even pace.
• Stock phrasing. Transitions such as "furthermore", "in conclusion" or "it is important to note" appear far more often in model output.
• Specificity. Personal details, names, numbers from lived experience and small inconsistencies are more common in human writing.

ZeroClaude scores every sentence with a language model trained to spot these cues and blends that judgement with statistical measures such as sentence-length variation. The overall percentage is the share of your text that ends up highlighted, so the number and the highlighting always agree.

Why a score is a probability, not a verdict

Detectors are trained on examples, and examples never cover everything. A short paragraph simply does not contain enough signal; a heavily edited AI draft can read as human; a careful human writer can sound machine-like. That is why we show a percentage and highlight sentences instead of printing a yes-or-no stamp.

Writing that often confuses detectors

• Formal templates: abstracts, legal clauses, product descriptions and cover letters follow fixed patterns that look generated.
• Non-native writing: simpler vocabulary and textbook grammar can raise the score unfairly.
• Very short texts: under about 150 words, any detector becomes unreliable.
• Lists and definitions: factual, neutral sentences with no personal voice.

How to read a result

Treat the score as the start of a conversation. Look at which sentences are highlighted and ask whether they fit the rest of the text. If the stakes are high, ask for drafts, notes or version history. And remember that no tool, including ours, can prove who wrote a text.

Keep reading