Are AI humanizers really undetectable?

When a humanizer claims to be 'undetectable', it almost always means: passes a specific list of public detectors at the time of marketing copy. The list is sometimes shown on the landing page (GPTZero, Turnitin, Originality, Winston). The claim becomes outdated as soon as any of those detectors updates its model.

What it does not mean: undetectable to a human reader. Undetectable to a forensic analysis. Undetectable to detectors that haven't been benchmarked against. Undetectable forever.

The strongest evasion is against the most generic surface features of LLM output: even sentence length, uniform vocabulary, generic transitions. Humanizers introduce variability in these dimensions effectively. A detector that relies primarily on those signals will see lower scores after humanization.

Several signals are much harder for humanizers to erase. Repeated semantic structure — the way an LLM tends to make the same kind of argument in the same order — survives most paraphrasing. Generic factual hedging ('various studies have shown') survives unless the humanizer specifically rewrites for specificity. Distinctive LLM phrasings around contrast ('however', 'on the other hand', 'in contrast') resurface even after aggressive rewriting.

Detectors that look at these deeper structural patterns — rather than just surface statistics — are more robust against humanization. The trade-off is a higher false-positive rate on formal human writing that legitimately shares those patterns.

Humanization and detection are an adversarial arms race. Each side updates against the other. Any specific 'undetectable' claim is true for a window and then degrades. The same is true of detectors: they catch up, then fall behind, then catch up again.

The practical implication is that no one should make a high-stakes decision — academic integrity, employment, defamation — based solely on whether a humanizer passed or failed a particular detector on a particular day. The signal is moving too fast to bear that much weight.

Detection plus context, always. The score is a starting point. The conversation, the drafting history, the consistency with the writer's known voice — those are the evidence. A confident 'undetectable' tool marketing claim is a reason to do the conversational check, not to skip it.