How AI text detection works, and where it fails

Every few months a new language model writes more fluently than the last one. Detectors have to keep up, and the people who rely on them need to know what a score actually means. This article explains how modern AI text detection works, where it breaks down, and how to use it responsibly.

What a detector looks for

Language models generate text one token at a time, each time choosing among the most likely next words. Over a whole paragraph that leaves statistical fingerprints a person rarely produces:

• Predictability. Machine text tends to pick the expected word. A detector can measure how surprised a reference model is by each word; human writing surprises it more often.
• Uniform rhythm. People mix long and short sentences, change register, digress and correct themselves. Models keep a steadier rhythm.
• Stock phrasing. Openings like "In today's fast-paced world" and closings like "Ultimately, it is a multifaceted endeavor" are far more common in model output than in human writing.

No single signal is reliable on its own, so a good detector combines many of them.

How LookAI combines signals

LookAI runs an ensemble: several neural classifiers trained on human and AI text, plus measurements of predictability from reference language models. A calibration layer turns their outputs into one score from 0 to 100, set separately for each supported language because the same writing habits look different in Spanish, Japanese or Arabic.

The calibration is the important part. Raw classifier outputs are not probabilities, and a model that is 90% sure in English can be badly overconfident in Croatian. We fit the calibration on texts the models have never seen, and we set the thresholds so that fewer than 1 in 100 human texts in our tests reach the AI band.

Sentence-level detail

A single document score hides a common case: a human essay with one AI-written paragraph pasted in. LookAI also scores every sentence in the context of its neighbours, and highlights a passage only when at least two consecutive sentences, or about 40 words, read as machine-written. Single sentences are too short to judge on their own, and highlighting them would mostly show noise.

Where detection fails

Detectors are statistical tools, and they have known weak spots:

1. Short texts. Below about 80 words there is simply not enough signal. LookAI marks those results as low confidence.
2. Formal, formulaic human writing. Legal clauses, scientific abstracts and news briefs are predictable by design, so they can look machine-like.
3. Heavy editing. A person who rewrites a model's draft thoroughly has, in a real sense, written a new text. Detectors catch light humanizer passes much more often than deep rewrites.
4. Writers working in a second language. Simpler vocabulary and textbook phrasing can raise scores. Calibrating per language helps, but it is not a complete fix.

How to read a result

LookAI returns three classes rather than a bare number:

• human when the score is below 30 with medium or high confidence,
• ai when it is 75 or above with medium or high confidence,
• inconclusive for everything in between, or whenever confidence is low.

The inconclusive band is deliberate. It is where honest uncertainty lives, and it is where a person should look at the text, the writing history and the context before drawing any conclusion.

A detector result is evidence for a person to weigh, not a verdict to act on automatically.

That principle is also in our Terms of Service: results must never be the only basis for a decision that affects someone's grades, job or reputation.

The short version

AI detection works by measuring how predictable and how uniform a text is, combined across several models and calibrated per language. It is accurate enough to flag text worth reviewing at scale, and not accurate enough to replace judgement. Use the confidence level, look at the highlighted passages, and keep a human in the loop.