Why AI Writing Lacks Rhythm (And How New Research Uses That to Catch It)

Quick answer

A September 2026 study from IBM Research and BITS Pilani (accepted to TMLR 2026) found that the strongest signal separating AI writing from human writing isn't how predictable a sentence is on average. It's how that predictability moves, whether it fluctuates in bursts the way human attention and creativity do, or stays locked in a flat, steady rhythm the way a likelihood-maximizing model does. The method, called DivEye, beat existing zero-shot detectors by up to 33.2% and made existing detectors up to 18.7% better when added as a supporting signal. The finding matters less as a product pitch and more as an answer to a question every honest detection company gets asked: why do single-signal checks keep failing, and what would actually fix that?

Perplexity told half the story

If you've read anything about how AI detectors work, you've run into "perplexity." It's a measure of how surprised a language model is by each word in a sequence. AI-generated text tends to score low, because the model that wrote it picked the statistically likely next word most of the time. Human writing tends to score higher, because people say unexpected things.

Perplexity is real and it's useful. It's also incomplete, and the DivEye researchers show exactly where it breaks down. Averaging surprisal across an entire passage collapses a lot of information into one number. Two texts can land on the same average surprisal score while looking completely different up close, one steady and uniform throughout, the other swinging between predictable stretches and sudden bursts of an unusual word choice or a structural left turn.

The researchers tested this directly. They collected 200 human-written essays and 200 GPT-4-Turbo essays on matched topics, ran both sets through a fixed evaluator model, and plotted the distribution of surprisal values for each. Human essays produced a wider spread with heavier tails; a passage would coast along predictably, then suddenly spike. GPT-4-Turbo's essays clustered tightly around lower surprisal with much less variance. Same underlying tool, same measurement, but the shape of the pattern over the text was the real signal, not the single-number average.

What "rhythm" actually means here

The paper's framing borrows a real linguistic idea: Uniform Information Density, the theory that efficient communicators try to spread information evenly across a sentence. Human writers approximate this, but imperfectly and unevenly, producing what researchers call burstiness: predictable stretches interrupted by unpredictable ones, tracking topic shifts, stylistic choices, and plain human inconsistency. Large language models, trained to maximize the likelihood of the text they generate, tend to satisfy that same efficiency principle too cleanly. The result is text that reads fluently but moves at a flatter, more uniform pace underneath.

DivEye captures this by going beyond the mean and variance of surprisal to compute higher-order statistics: skewness (how lopsided the distribution is, which flags occasional rare, surprising word choices typical of human writing) and kurtosis (how heavy the tails are, meaning how often a passage swings unusually far from its own average). It then tracks how surprisal changes moment to moment across the sequence, first-order shifts and second-order shifts in that rate of change, to detect whether unpredictability arrives in human-like bursts or stays evenly spread out the way a model's output does. Ninety-nine features in total, built from nothing but a passage's own token probabilities, no access to the model that wrote it required.

The numbers, and where the method actually held up

Tested across the MAGE and RAID benchmarks, two of the standard academic testbeds for AI-text detection, DivEye averaged 0.92 AUROC across in-distribution and out-of-distribution settings and beat prior zero-shot detection methods (including Binoculars and FastDetectGPT) by wide margins in the harder, out-of-distribution cases where a detector faces writing domains or models it has never seen. On the RAID leaderboard, which pools 8 writing domains, 12 models, and 4 decoding strategies, DivEye landed within 0.20 percentage points of AUROC of the best fine-tuned, purpose-trained detector on the board, while requiring no training data of its own.

The adversarial numbers are the ones worth sitting with. Against paraphrasing attacks, the most common way people try to strip AI fingerprints from generated text, DivEye outperformed the fine-tuned Longformer baseline by over 10 points of average accuracy and beat Binoculars by 11.2% on RAID's adversarial suite. The paper also tested the method on Chinese-language data (via the HC3 corpus) and found it held up across languages, a place where several detection approaches have historically struggled, as we've covered before.

None of this makes DivEye, or rhythm-based detection generally, infallible. The paper's own ethics section says as much: outputs should be treated as probabilistic signals, not verdicts, and any high-stakes use needs human review layered on top. That's the same honest caveat every credible detector in this space should be making, us included.

Why one signal was never going to be enough

The real headline isn't "here's a new detector." It's that the researchers built DivEye explicitly as a booster, a feature set meant to sit alongside existing detection methods rather than replace them, because it captures a genuinely different kind of signal than the fine-tuned classifiers and semantic-level checks most tools already use. Combining rhythm-based features with an existing detector's score, fed into a lightweight classifier, produced the biggest accuracy gains of the whole paper.

That's the architecture worth understanding, whether or not you ever touch this exact research: no single statistical signal, not perplexity alone, not burstiness alone, not any one company's proprietary score, should be asked to carry a detection decision by itself. A detection score is evidence, not a verdict, and the strongest detection systems stack multiple independent signals precisely because any one of them, examined in isolation, has known blind spots. That's the same principle behind why a single AI detector score should never decide a case on its own, academic misconduct or otherwise.

If you want to see what a multi-signal read looks like on a real passage, rather than a single perplexity number, run it through our free detector. You'll get a sentence-by-sentence breakdown instead of one score standing in for the whole picture.

FAQ

Is DivEye a product I can use right now? The researchers released it as an open academic project (code on GitHub, a demo on Hugging Face), not a commercial detection service. It's a research contribution meant to improve the field's methods, including feeding into tools like ours, not a competing consumer product.

Does this mean perplexity-based detection is obsolete? No. Perplexity remains a real and useful signal; the paper's point is that averaging it into one number loses information a text's full surprisal pattern carries. The strongest approach uses both.

Can this kind of detection be fooled by paraphrasing tools? The paper tested exactly that and found DivEye held up better than several existing zero-shot methods under paraphrasing attacks, though the authors don't claim any method is unbeatable. Some light paraphrasing has even been shown to increase detection confidence rather than lower it, which tracks with the idea that surface-level rewrites don't always erase the deeper statistical pattern.

Why does this matter for writers worried about false positives? A detection approach built on distributional patterns across an entire passage, rather than flagging individual suspicious phrases, is less likely to misfire on a human writer's one unusual sentence. That said, no detector eliminates false positives entirely, which is why a single score should never be the whole story.