What is new in AI detection: verified or not
We went through recent 2026 work. This page keeps only what we could find in a primary source (a paper abstract or the text of the law). Authors' numbers remain the authors' numbers: we did not reproduce them.
What is confirmed
- Benchmarks are closer to real life. DetectRL-X (ACL 2026) tests detectors on 8 languages and 6 domains, with text from 4 commercial models. Texts are not only generated but also polished, expanded, condensed, then paraphrased and perturbed. It measures the effect of domain, model, attack and length. These are exactly the cases where detectors are weakest.
- Rhythm of surprisal as a feature. The authors of DivEye measure how the unpredictability of words fluctuates across a text. By their account the method beats zero-shot detectors by up to 33.2% and adds up to 18.7% to other detectors, and is robust to paraphrasing. This is the authors' claim; the abstract does not mention multilingual tests.
- More than two classes. The GPTZero paper describes a hierarchical scheme with finer human and AI categories. The abstract gives no accuracy figures, and this is a vendor description.
- Shuffling words. The preprint Luminol-AIDetect compares the perplexity of a text with that of its shuffled copy. The authors report 18 languages and 11 attack types, but the paper is still under review.
- Marking by law. Article 50(2) of the EU AI Act requires outputs of systems that generate synthetic audio, image, video or text to be marked in a machine-readable format. Exceptions: an assistive function for standard editing, and cases where the system does not substantially alter the input. According to the article text it applies from 2 August 2026; for systems already on the market the marking deadline is 2 December 2026 per the implementation timeline.
What we left out, and why
- Figures such as "Binoculars gets about 55% on Claude text" and "a 12.7-point loss under paraphrase": we saw them only in a search engine's summary and did not check the primary sources.
- Claims about which vendors already watermark text, and plans of individual companies: sources contradict each other and we found no primary source.
- A possible postponement under the "Digital Omnibus" package: we found no confirmation.
- The idea that shuffling or rhythm features will help specifically for Ukrainian or Portuguese: the papers do not say so; it is only our hypothesis to test.
What this means for checking a text
Nothing here cancels the earlier conclusion: there is no single reliable test, and detectors fail most on mixed texts and paraphrase. In our own preliminary measurement, human text edited by a model was almost never recognised: about 0.5% were flagged at a 1% false-positive rate. That is why we show an estimate with uncertainty and say when the signal is not enough, instead of a "percentage of AI involvement".
Why the marking rules matter, and why they do not solve it
A legal duty to mark generated text sounds like the end of guessing. It isn't. A mark that a system adds to its own output tells you something only when it survives copying, translation and editing, and only for systems that apply it. Text from a model that does not mark its output, or text a person rewrote by hand, carries no mark at all. Marks are also a different thing from what our tool looks for: we detect literal invisible characters (259 kinds) and typography, and we cannot see statistical watermarks hidden in the choice of words.
So for now two kinds of evidence exist side by side. One is the style of the text, which is weak and can be fooled in both directions. The other is a declared mark, which is strong when present and silent when absent. Neither replaces the human evidence: drafts, notes, version history and a conversation with the author.
What to do next
Read how AI text detection works today for the full list of methods and their limits, and why modern AI detectors get it wrong for the failure modes. When you need to check a real text, use the checker and treat the result as a hint about where to look.