Perplexity
Perplexity is a statistical measure that focuses on how predictable the words are in a piece of writing. Large language models predict the "next token" based on the pattern they have learned in the data. It will choose the most logically sound next word. The more predictable the words are according to patterns, the lower the perplexity of the writing.
Human writing is not as predictable. They write on a creative whim, choosing the words as their ideas present themselves in their mind. This can be surprising, chaotic, complex, and notoriously unpredictable. This is why human writing is of high perplexity.
An AI detector may interpret low-perplexity texts as AI-like and high-perplexity texts as human-like.
Burstiness
Burstiness is another statistical indicator that measures the change in sentence structure and length of any writing.
A detector will also check syntactic complexity and variation of vocabulary, with lower values of this leading to low burstiness.
If the length and complexity of the sentences vary, and the writing uses words that are both complex and diverse, then the burstiness is high.
AI models write with low burstiness, while human writing, especially high-quality work, will have high burstiness.
Especially when sentences alternate between different lengths and complexity, it is more likely to be human, as people don't follow pre-made patterns and tend to be more dynamic and spontaneous.
N-gram Analysis
Think of how text autocomplete works: AI works in a similar way. The models generate text by calculating the probability of the next word based on what came before.
Detectors use an engine that looks at the exact probability of every word choice. The detector carries out Top-k and Top-p evaluation, checking if the text consistently uses the top 10 or 20 most likely words that an AI model is likely to pick.
If the probability of the word choices is too perfectly smooth and rarely dips into highly unusual or rare words, the mathematical "curve" flags out as machine-generated text.
Lexical Diversity
Humans who write have a large vocabulary at their disposal to draw from when writing. But AI models, no matter how advanced, would rather stay clear and safe, and tend to use highly repeatable terms.
Detectors measure something called Textual Lexical Diversity, which calculates the ratio of unique words in comparison to total words. A lower MTLD index often signals the repetitive, controlled vocabulary typical of AI.
Type-Token ratio is another statistical rate, which measures the total number of unique words divided by the total number of words. AI texts feature a more uniform and narrow TTR, compared to human writing.
Syntactic and Formulaic Frequency
Most people who have interacted with AI would say one thing: the AI writing structure is very rigid. You will find overly structured sentences with perfectly balanced grammar across the written piece. Studies have shown that AI text heavily overuses fixed and predictable connector combinations.
Detectors note the mathematical frequency of logical connectors and also the paragraph structure, as AI models tend to generate writing with structurally consistent layouts.
The detection tools will flag texts if the paragraph lengths, sentence-to-paragraph distributions, and bullet-point formatting match a mathematically balanced grid.
Watermarking Signals
While watermarks are relatively new, they are expected to become one of the more vital clues of AI detection. These are invisible mathematical stamps left behind by the AI tool.
Some AI providers subtly bias model's word selection, making it favor certain word combination more prevalent. If a detector is built to recognize this specific bias, it can instantly confirm whether the text contains a cryptographic "watermark".
Semantic and Contextual Analysis
Semantic analysis evaluates how ideas connect with one another and whether the underlying thought process mimics that of a machine or humans.
Detectors can convert entire sentences and paragraphs into multi-dimensional mathematical coordinates.
AI text follows an uncommonly straight path through "vector space, revealing a predictable logical trajectory, while human writing tends to jump creatively between conceptual points.
Automated text also produces a lot of hallucination fluff, surrounding a central fact with repetitive and overly generalized content to create grammatical smoothness.
Detection tools tend to track the ratio between actual, verifiable facts and generic fillers. There is also a lack of logical cohesion in AI text, as within the generated writing, it can occasionally forget the exact logical boundaries it initially set and may add contradictory ideas. Some detectors can scan for semantic cohesion and catch AI writing that may not make sense.
Final Thoughts
AI detectors are still relatively new, in terms of reliability. As AI models become more advanced, AI detection tools are also evolving and becoming more sophisticated. As these tools become more advanced, the rate of false negatives and positives will also reduce.
Each detector uses a different combination of statistical analyses to determine whether the writing is AI-generated or human-written. Any tool that uses all these systems will be able to give you a more accurate result.