How AI Text Detection Works — And Its Blind Spots (2026)

Perplexity

Perplexity measures how "surprised" a language model is by each word in a sequence. Given the words that came before, how likely is the next word? If the model would have easily predicted it, perplexity is low. If the word is unexpected, perplexity is high.

AI language models produce low-perplexity text because they're optimized to generate the most probable continuation. They reliably pick the statistically expected word. Human writers are less predictable — they use unexpected metaphors, make unusual connections, and sometimes choose the interesting word over the obvious one.

This is why perplexity is a useful signal. But it's not a perfect one. Formal, technical, or academic writing naturally has low perplexity because it follows conventions. A dry research paper written by a human can score as AI. A creative AI output with unusual prompting can score as human.

Burstiness

Burstiness measures variance in sentence length. Human writing is bursty — we naturally alternate between short punchy sentences and longer explanatory ones. This creates rhythm.

AI writing is less bursty. Language models optimize locally (each sentence), which tends toward sentences of similar length and complexity throughout a passage.

GPTZero was the first major detector to explicitly incorporate burstiness, and it remains one of the most reliable signals because it's hard to fake without actually restructuring the writing.

Trained classifiers (Originality.ai, Copyleaks)

More sophisticated detectors don't just measure perplexity — they use trained models that have learned to distinguish AI text from human text across many examples. These models pick up on patterns that perplexity doesn't capture: syntactic preferences, specific phrase patterns, structural tendencies.

Originality.ai v3 uses this approach and is significantly harder to bypass with simple methods because it's not just measuring one signal.

N-gram patterns

Some detectors measure the frequency of specific word sequences (n-grams) that are overrepresented in AI output. "It is important to note that", "Furthermore, it is worth", "delve into the intricacies" — these phrase patterns appear in AI text at rates far above human writing.

Semantic coherence

AI text is often "too coherent" — every sentence follows logically from the previous one, every paragraph connects cleanly. Human writing digresses, makes unexpected jumps, and has sections that feel slightly off-topic before coming back. Some detectors measure this.

Syntactic patterns

Passive voice frequency, subordinate clause structure, the position of the main verb — these patterns differ between AI and human writing in measurable ways.

Detectors work well on:

Unmodified AI output. Direct ChatGPT or Claude output, without any editing, typically scores >90% AI on major detectors. The signals are strong and consistent.

Long-form content. Detectors have more signal to work with in longer texts. A 1,000-word essay is much more reliably classified than a 100-word paragraph.

Standard prompting patterns. When AI is prompted with standard instructions ("write an essay about X"), the output follows predictable patterns. Detectors trained on this data work well.

Consistent use of specific AI tools. Each AI model has somewhat distinctive stylistic tendencies. GPT-4 has specific patterns. Claude has others. Detectors can sometimes identify not just "AI" but which AI was used.

Short texts. With fewer than 250 words, accuracy drops significantly for all major detectors. There's not enough signal. False positive and false negative rates both increase.

Domain-specific formal writing. Legal documents, medical reports, technical specifications — these read as low-perplexity, low-burstiness text regardless of who wrote them. Human-written formal documents frequently score as AI.

ESL (English as Second Language) writing. Research has consistently found that ESL human writing is incorrectly flagged as AI at higher rates than native English writing. This is a significant and well-documented problem with real-world consequences.

Heavily edited AI output. When AI text is substantially rewritten at the structural level — not just paraphrased, but restructured, with rhythm changed and phrasing signature removed — detection rates drop sharply.

Inconsistent results. Running the same text through the same detector multiple times sometimes produces different scores. This inconsistency is underacknowledged by detector providers.

The false positive problem is particularly important in academic contexts. Detectors are regularly used to flag suspected AI use in student work — but false positive rates of 5-15% mean that in a class of 100 students, several genuine human essays may be incorrectly flagged. The consequences of a false positive accusation can be serious.

Detectors and humanizers are in an ongoing arms race. As humanizers improve, detectors update their models. As detectors improve, humanizers adapt. This is unlikely to reach a stable equilibrium.

The practical implication: no humanizing method is permanently reliable. What achieves 2% detection today might achieve 20% in 18 months as detectors retrain.

The more robust approach is to focus on text quality rather than detector manipulation. Text that a human reader genuinely finds natural and engaging tends to pass detectors too — because the features that make text feel human are the same features that detectors are trained to recognize.

This is why the most effective humanization tools don't optimize against detectors specifically — they produce better, more natural writing, which happens to pass detectors as a byproduct.