How AI detection systems work — and their limitations.

Important note: This article explains AI detection at a general conceptual level based on publicly known methodology. This is not insider knowledge of any specific product. Detection systems vary considerably in their approach. This page is not intended to help users circumvent detection — it is here to help writers understand why AI text reads differently and what genuine improvement looks like.

The core principle: statistical predictability

Most AI detection systems are based on some variation of the same core observation: language models choose words based on statistical probability. Given a sequence of words, an LLM predicts and selects the next word based on what is statistically most likely given its training data.

This means that AI-generated text tends to be statistically predictable — each word choice is, to some extent, the expected choice. Human writing, by contrast, introduces unexpected words, unusual phrasings, personal references, and genuine opinions that are less statistically predictable at the token level.

Perplexity as a signal

Perplexity is a measure of how surprised a language model is by a sequence of text. High perplexity means the text was unexpected (less predictable). Low perplexity means the text was very predictable.

AI-generated text tends to have low perplexity when evaluated by a similar language model — because it was generated by exactly the kind of statistical process the detector is measuring. Human text tends to have higher perplexity because humans make unexpected word choices.

However, perplexity alone is not a reliable detector. Technical writing, formal academic writing, and highly structured genres are inherently low-perplexity — they use predictable vocabulary because that is appropriate for the genre. A simple perplexity-based detector would incorrectly flag much legitimate human writing.

Burstiness as a signal

As discussed in the sentence structure article, human writing shows variation in sentence length (“burstiness”). AI output tends to be less bursty — more uniform in sentence lengths. Modern detection systems incorporate structural variation analysis alongside vocabulary-level signals.

Why detectors have significant false-positive rates

Several published academic and industry analyses have noted that current AI detectors have meaningful false-positive rates — texts written by humans being flagged as AI-generated. This is an inherent challenge because:

• Non-native English speakers often write in more formal, predictable patterns that resemble AI output.
• Certain genres (legal writing, academic papers, technical documentation) are inherently low-perplexity.
• Any writer who reads a lot naturally absorbs common phrases that appear frequently in training data.
• Detection models can drift as AI models evolve, making older detectors less accurate against newer outputs.

What phrase-substitution tools can and cannot do

Tools like this one operate at the surface level: they replace specific words and phrases. This can reduce some pattern-level signals (like repeated transition words) but does not change the underlying token probability distribution of the text.

Genuinely improving AI-assisted text to be both high-quality and authentically human requires human judgment, personal perspective, and genuine editing — not just phrase substitution. The tool here is a useful step in that process, but it is not sufficient on its own.