AI Text Detection Signals 2026: How Detectors Actually Work

Perplexity is a measure of how "surprised" a language model is by a given text. AI-generated text typically has LOW perplexity because it was generated by sampling from high-probability tokens: the words it picks are predictable to a language model. Human text has higher perplexity because humans use unexpected word choices, idioms, and abrupt topic shifts. Detectors compute perplexity by feeding the text back through a reference language model (often GPT-2 or a small open-source LLM) and measuring how confidently it predicted each token.

Burstiness measures variation in sentence length and structure. Human writing alternates between short punchy sentences and long flowing ones: bursty. AI text tends toward consistent medium-length sentences (15-25 words) with similar syntactic structure. Detectors measure burstiness as the standard deviation of sentence lengths divided by the mean. Bursty text (high std/mean) typically reads as human; uniform text (low std/mean) flags as AI.

Google DeepMind SynthID and similar watermarking schemes embed a statistical signature during generation by adjusting token probabilities. The signature is invisible to readers and does not change output quality in normal use, but it is only available when the model provider adds the watermark at generation time. DeepMind says confidence can fall when text is thoroughly rewritten or translated.

Perplexity, burstiness, and phrase-repetition signals usually weaken when text is heavily paraphrased, translated, or manually edited. Provider-side watermarks such as SynthID can survive light edits when the original model embedded the watermark, but confidence can fall after thorough rewriting or translation. Stylometry is most useful when compared against a known author history. None of these signals should be cited as proof by itself.

N-gram repetition measures repeated word pairs, word triples, transitions, and generic assistant-style phrases against a baseline corpus. Phrases such as "delve into" or "navigate the complexities of" are not proof by themselves, but repetition can support a broader review when it appears with uniform sentence rhythm, low source specificity, and weak process evidence.

Partially. Paraphrasers and humanizer tools can increase surface variation, break repeated n-grams, and change sentence rhythm. That may lower confidence in some detectors, but it can also create awkward idioms, source errors, and inconsistent tone. A lower detector score after paraphrasing does not prove human authorship.

Zero-shot detection means the detector was not trained on the exact labeled AI-vs-human examples it is reviewing; it uses model-behavior signals such as likelihood, log-rank, or perturbation response. DetectGPT is a well-known example that compares likelihood around small text perturbations. Zero-shot methods can help with new models, but they still fail on short, edited, translated, or genre-constrained text.

Yes: this is the hardest case. When a human edits AI-generated text by replacing 20-30% of sentences with their own writing, perplexity becomes mixed and burstiness rises, blurring detection. Some detectors output sentence-level scores rather than document-level, allowing users to identify which spans look AI-generated even when the whole doc passes a 50% threshold. The 2025 GPTZero "Smart Sentence" feature and EyeSift's span-level highlighting both target this use case.

Several known cases: non-native English writing, highly formal academic writing, journalistic style, translated text, and short samples. Stanford HAI summarized a 2023 study finding that seven detectors classified 61.22% of TOEFL essays by non-native English writers as AI-generated, with all seven detectors agreeing on about 19% of those essays.