When you run a dozen AI-detectors at once, do false alarms and missed fakes both get worse — or does fixing one just create the other?

This explores what happens to mistakes when you use a whole set of AI-text detectors, one for each model version, instead of a single one. The short answer from the corpus: the two kinds of error don't build up together. They move in opposite directions, and how you combine the detectors decides which one you get. In a test covering rewrites from 23 LLM versions, running a separate detector for every version and flagging a text if any of them fires caught 98.6% of AI rewrites. It also wrongly flagged 11.9% of human-written abstracts, roughly one in eight. Training one detector on all 23 versions together turned this around: false flags stayed near 1%, but it missed 33.8% of rewrites from the newest version Can one detector catch all LLM versions at once?.

The arithmetic behind the false flags is the useful part. If you flag a text whenever any detector fires, each detector's small false-alarm rate adds to the total. Suppose each of 23 detectors wrongly flagged a human abstract only about half a percent of the time, and their mistakes were independent. Together they would flag about 12% of human abstracts, close to what the study saw. (That calculation is an illustration, not the study's own analysis.) Every detector you add for broader coverage is another chance to accuse a human writer. Pooling the training data does the reverse: it averages the versions' signals into one blurred picture, and the newest version's quirks get washed out. So the misses pile up exactly where they matter most.

Those misses also come from somewhere specific: they cluster at the moments when a new model is released. Detectors trained on earlier versions catch over 99% of rewrites made just before a new release and only 3.8% just after, because each new version picks words differently Why do AI detectors suddenly fail when models update?. So a multi-detector system doesn't lose accuracy slowly over time. It works well until a release, then fails abruptly, and the only fix is to add another detector for the new version. That fix pushes the false-alarm count up again.

It also matters whether the detectors' mistakes are independent. In theory, repeating independent checks can drive classification error toward zero, but only when each check's errors are independent and the two classes really behave differently Can repeated quiet probes separate decoys from genuine objects?. Detector errors often fail that condition. Fake-news detectors, for example, flag truthful AI-written text as fake because they mistake AI writing style for deception Why do fake news detectors flag AI-generated truthful content?. When detectors share a bias like that, the same human writers keep getting flagged. Agreement-based checks have the matching weakness on the other side: they catch answers that vary between samples but miss mistakes repeated identically, because agreement looks like confidence even when every sample is wrong Can agreement across samples reveal when models are wrong?. Agentic evaluation systems show a similar pattern, where one faulty memory module spread its errors through an otherwise strong pipeline Can agents evaluate AI outputs more reliably than language models?.

The takeaway: building an ensemble doesn't remove the precision/recall trade-off. It only lets you choose where to place it. A flag-if-any-detector-fires setup protects against missed AI text at the expense of human writers. A single pooled detector protects human writers at the expense of catching the newest models. The corpus doesn't yet have a study of designs that sit between these, such as weighted voting or detectors that check how recent a text's style is, so that remains an open question.