AI Detection Scores: Why the Percentages are Lying to You

A "90% AI-generated" score doesn't actually mean there is a 90% chance the text was written by an LLM. After digging into how these classifiers actually work, it's clear that most people are misinterpreting the probability distribution as a confidence interval.

The core issue is that these tools aren't "detecting" AI; they are measuring perplexity and burstiness. When a detector spits out a high percentage, it's essentially saying the text is highly predictable based on its training data. This creates a massive false positive problem for non-native English speakers or technical writers who use a standardized, formal style.

I ran a test with a few paragraphs of my own highly structured technical documentation—written entirely by hand—and one tool flagged it as 85% AI. The "evidence" was simply that my sentence length was too consistent.

If you're trying to build a reliable AI workflow or implement a content policy, relying on these scores is a recipe for disaster. Here is the reality of those numbers:

• Perplexity: High predictability = High AI score.
• Burstiness: Low variance in sentence structure = High AI score.
• False Positive Rate: Extremely high for academic or technical prose.

The only way to actually validate these scores is to run a side-by-side comparison with a known human-written baseline of the same genre. If the "human" baseline also scores 40% AI, then the detector's 90% score on your target text is actually only a 50% delta, which is far from definitive proof.

Stop treating these percentages as a "truth meter" and start treating them as a "predictability index."