When AI text detectors should say “not enough text”

A short passage, a different test

Can an AI detector reliably judge a ten-token snippet using a threshold chosen for much longer passages? An arXiv preprint by Ritesh Janga and Justin Zhan examines why that can go wrong—and proposes withholding judgments at lengths where detection performs poorly.

Tokens are the chunks into which language models divide text, sometimes whole words and sometimes word fragments. With fewer tokens, a detector has less evidence. But the consequences are not simply lower accuracy: depending on the detector, a threshold calibrated on long text can either flag too many human passages or miss more machine-generated ones.

The study combines mathematical analysis, simulations and computational experiments with three existing detectors. It does not introduce a new detector or establish a universal minimum length for identifying AI writing.

Why one threshold can mislead

A score-based detector assigns a number to a passage and compares it with a cutoff. Its false-positive rate is the fraction of human passages incorrectly flagged; its true-positive rate is the fraction of machine-generated passages caught.

The authors derive how these rates change with length under an approximately normal statistical model that accounts for dependence between tokens. If the average human score stays stable, shorter passages have more variable scores. More human passages can then cross a cutoff chosen for longer text. Under those assumptions, a cutoff producing 1% false positives at one length produces about 12.2% at one-quarter that length.

That is a theoretical result, not a universal measurement. Human scores can also shift systematically with passage length, pushing false positives in the opposite direction. The real-detector experiments show both behaviors.

Calibrating against comparable passages

The proposed procedure compares a passage with held-out human examples in the same length bin. It uses the passage’s score ranking to calculate a conformal p-value: how unusual that score is relative to the human reference examples, not the probability that AI wrote the passage.

Under an assumption called exchangeability—roughly, that calibration examples and future human passages are comparable draws from the same distribution—the procedure bounds the false-positive probability within each bin. It does not require normally distributed scores.

Crucially, that guarantee averages over possible calibration samples. It does not promise that every particular calibration set will deliver exactly the target error rate.

The researchers tested GPT-2-based log-likelihood and log-rank scores, plus Binoculars, which compares two language models. Human passages came from Wikipedia, AG News and Amazon Reviews; machine passages came from Llama-3.1-8B-Instruct and Mistral-7B-Instruct-v0.3. Passages were truncated to six lengths between 10 and 200 GPT-2 tokens, with source documents kept separate across data splits.

Across 300 random document-level resplits, length-specific calibration produced mean false-positive rates between 4.93% and 5.02% in all 18 detector-length combinations, against a 5% target. Individual splits varied.

By contrast, a Binoculars cutoff calibrated at 200 tokens falsely flagged 29.4% of human ten-token passages in the reported comparison. The equivalent log-likelihood cutoff flagged only 1.0%—but that conservatism came with poor detection power.

Controlling errors does not create evidence

After calibration, none of the tested detector-length combinations reached an 80% true-positive rate. The best pooled result was Binoculars at 200 tokens, catching 59.1% of machine passages in the reported test split. At ten tokens, it caught 10.5%.

The authors therefore add a bin-level abstention rule. Using separate machine-text validation examples, they estimate detection power. If even its upper confidence bound falls below a chosen floor, the entire length bin returns “insufficient evidence at this length.”

At an 80% floor, every tested bin abstained. At 50%, only 200-token Binoculars results were released; at 30%, all three detectors’ 200-token bins were released. Passing this rule means inadequate power was not established by that criterion—not that performance above the floor was certified.

Limits beyond length

The pooled results conceal substantial differences between text sources. At 50 tokens, the same calibrated log-likelihood cutoff falsely flagged 13.2% of Wikipedia passages but 0.8% of AG News passages. A pooled guarantee is not a guarantee for every subgroup.

All experiments used English text. Non-native writers, mixed human–AI authorship, adversarial editing and lengths above 200 tokens were not evaluated. The corpus balance also changed at 200 tokens because few news articles were long enough, complicating comparisons across lengths.

The preprint supports a narrower conclusion than “short AI text is undetectable”: these detectors, under these conditions, often lacked useful power. Length-aware calibration can manage one source of error. Abstention makes the remaining weakness visible rather than turning it into a confident accusation.