How AI Detection
Actually Works

AI content detection is the process of analyzing text to determine whether it was written by a human or generated by a large language model (LLM) such as ChatGPT, Claude, or Gemini. Unlike plagiarism detection — which compares text against a database of known sources — AI detection analyzes the statistical and linguistic properties of the text itself.

The core insight is simple: AI models and humans produce text in systematically different ways. When a language model generates text, it is constantly predicting the next most likely word given everything that came before. The result is output that is grammatically clean, structurally uniform, and statistically predictable. Human writing, by contrast, is messier, more varied, and harder to predict — even when it is technically well-written.

AI detectors exploit this difference by measuring how “AI-like” the statistical patterns in a piece of text are. The two primary signals are perplexity (how predictable the text is) and burstiness (how varied the sentence structure is).

Perplexity is a statistical measure borrowed from information theory. In the context of AI detection, it answers the question: how surprised is a language model when it encounters this text?

To understand why this matters, consider how LLMs work. A model like GPT is trained to predict the next word in a sequence. When it generates text, it selects high-probability word sequences — the words that “make sense” given the context. This produces fluent, grammatically correct output, but it also means the text follows predictable patterns.

When a detection model reads AI-generated text, it assigns low perplexity scores — meaning the text is highly predictable. When it reads human writing, it assigns higher perplexity scores, because humans make more unexpected word choices, introduce tangents, use colloquialisms, and generally deviate from the most statistically likely phrasing.

Perplexity in practice

Here’s how perplexity scores distribute across different types of writing. Lower scores indicate higher predictability (more likely AI-generated).

Notice that ESL (English as a Second Language) writing and edited AI text overlap in the medium range — this is one of the core reasons AI detectors produce false positives on non-native speakers. The uniformity that characterizes their writing style is statistically similar to lightly edited AI output.

Burstiness measures how much variation exists in sentence length and complexity throughout a text. The concept is borrowed from linguistics and network science, where “bursty” phenomena are those that cluster and vary rather than distributing uniformly.

Human writing is naturally bursty. In a single paragraph, a human writer might produce a long, complex, subordinate-clause-heavy sentence, followed by a blunt three-word fragment. The rhythm shifts. Emphasis is created by contrast. This natural variation is hard for AI to replicate — at least without specific prompting.

AI models default to producing text with consistent sentence structure and length. The variation in a 500-word GPT response tends to fall within a narrow range. Sentences are grammatically complete, similarly structured, and rarely burst into fragments or unexpectedly complex run-ons.

Why burstiness catches AI that perplexity misses

Some AI prompting techniques specifically address perplexity. A user can instruct an AI to “write in a more casual, varied style” and reduce perplexity significantly. But sentence-level burstiness is harder to prompt away — the model still tends to produce sentences within a consistent structural range, even when the vocabulary becomes more informal.

This makes burstiness a complementary signal to perplexity. When both measures are low, detection confidence increases significantly.

Beyond perplexity and burstiness, different AI models leave distinct stylistic fingerprints in the text they generate. These signatures emerge from differences in training data, reinforcement learning from human feedback (RLHF), and the specific objectives each model is optimized for.

Detection engines trained against reference outputs from multiple models can identify these patterns and flag which model is most likely responsible for the text — not just whether AI was involved at all.

Model signatures are probabilistic, not deterministic. As models are updated and improved, their fingerprints shift — which is why detection engines need to be continuously updated against new model versions and release families.

The AI probability score returned by this tool is a composite of all three signals — perplexity, burstiness, and model signature matching — combined using a weighted classifier trained on a labeled dataset of human-written and AI-generated text.

No AI detector is accurate 100% of the time. Understanding where these tools fail helps you use them responsibly — particularly in high-stakes contexts like academic integrity or employment screening.