A company selling rewriting tools claims to have developed a model capable of defeating the most accurate AI-text detectors on the market, with no loss of quality. We have no way of verifying this claim, which comes from an interested party. It nonetheless deserves attention, because it illustrates a problem we have been documenting for months.
We explained in our article on recognising generated text why this race was lost before it began. Here is the commercial illustration of that reasoning.
Why the asymmetry works against detection
The mechanism is simple and relentless.
A detector looks for a statistical signal: an unusual regularity in word choice, excessive homogeneity, a distribution that does not look like human writing.
To evade it, there is no need to reproduce human writing. It is enough to degrade the signal — introduce variation, break up the regularity, mix registers.
Destroying a signal is always easier than preserving it. That is why we wrote that the cat starts with a structural handicap, and why a watermark announced by a major lab was circumvented in four hours.
The real problem is not circumvention
It bears repeating, because it is the point that discussions systematically miss: the main danger of detectors is not that they miss cheats, it is that they accuse innocent people.
False positives hit identifiable profiles: those whose first language is not English, those who write in a highly structured way, those who use a limited vocabulary.
We noted this in relation to a court finding that non-native English speakers use AI twice as much — not to cheat, but to offset a disadvantage. A detection tool applied to that population would mechanically produce mass accusations against those who needed it most.
An institution that bases a sanction on a detection score exposes itself to two problems at once. First, it will not catch those using an evasion tool, that is, precisely the most determined. Second, it will punish good-faith students whom statistical chance has singled out. The system therefore catches the opposite of what it targets. That is why we have always recommended shifting the question: not "was this text written by an AI", but "can this person defend what they submitted". The second question has an answer, and it takes five minutes.
The approach that works
We repeat it because it is the only one that is not a losing race.
Prove the authentic rather than unmask the fake. This is the principle behind the C2PA provenance standard: signed content proves something, unsigned content proves nothing. The asymmetry finally works in the right direction, because a signature is hard to forge and easy to verify.
Make the publisher accountable. This is the choice made by the European transparency framework: rather than technically tracking every piece of content, the person who distributes it is held responsible. A rule that cannot be verified technically can remain effective if it exposes those who break it.
What to take away
Whether this commercial claim is accurate or exaggerated makes little difference. The reasoning holds regardless: a statistical signal is easier to degrade than to preserve, and that asymmetry will not reverse.
What must change is how detectors are used. As one clue among others, within a body of evidence, they have a modest usefulness. As proof, they produce targeted injustices without catching those they target.
And there remains the question this circumvention industry raises by implication: if a market exists for making a text undetectable, then another market exists for detecting it. Two industries now live off each other, and neither produces anything useful.