AI content detector for PDFs — what stylometric fingerprints actually detect and why no detector is courtroom-grade
AI-generated text has stylistic fingerprints — patterns that show up in LLM output dramatically more often than in human writing. Formulaic openers, hedging overuse, three-item rhetoric, em-dash patterns, polished-register uniformity. A skilled reader can spot AI text manually; a stylometric audit surfaces the same signals systematically. The friction is that none of these signals is definitive — humans sometimes write in patterns that match AI fingerprints, and careful editing can mask the fingerprints. Here is how the detector works, the eight specific patterns it looks for, and the precise reason we don't claim courtroom-grade accuracy.
How stylometric detection differs from trained classifiers
Two approaches to AI-text detection exist. Trained classifiers (GPTZero, Originality.ai) feed text into a neural network trained on labeled examples of human vs AI text. The classifier outputs a probability score. Trained classifiers can hit high accuracy on text from the LLMs they were trained on, but degrade against newer models and against edited AI text. They're also opaque — you can't easily see WHY the classifier flagged something.
Our detector takes the stylometric approach: explicit pattern detection. We look for documented LLM stylistic fingerprints in the text and surface each match with its location and confidence. The advantage is transparency — you see exactly what was flagged and why. The trade-off is that an adversary who knows the patterns can edit to avoid them. We're explicit about that trade-off.
Eight stylistic fingerprints the detector flags
Documented patterns from LLM output across major models (ChatGPT, Claude, Gemini, Llama, Mistral):
- Formulaic openers. "In today's fast-paced world," "It's important to note that," "In conclusion," "As we navigate the complexities of...." LLMs default to these openers heavily. Human writers occasionally use them; LLMs use them as primary opening sentences.
- Hedging overuse. "It could be argued that," "some might suggest," "there is a case for" — used liberally to project neutrality. Humans hedge selectively; LLMs hedge habitually.
- Three-item rhetoric. X / Y / Z triplets used so consistently that they become a signature. "The benefits include speed, accuracy, and reliability." Three-item lists aren't AI-specific, but LLMs use them at much higher frequency than humans do.
- Em-dash overuse. Long em-dashes (—) used to introduce qualifications or asides. Humans use them; LLMs use them at 3-5× the typical human rate.
- Register-too-polished. Uniformly formal voice with no personality variation, no asides, no opinion. Reads like a textbook on every topic. Real human writing has stylistic variation across paragraphs; LLM output is suspiciously uniform.
- Definition-textbook style. "X is defined as..." "X refers to the practice of..." — explicit definition framing where a human would just use the term. Common in LLM-generated educational content.
- Transitional clichés. "Moreover," "furthermore," "in addition" as section-openers. Humans vary transitions; LLMs cycle through the same set.
- Unprompted safety disclaimers. Mid-essay risk warnings, "it's important to remember," cautionary asides where none was needed. Trained-in LLM behavior leaking through.
Counter-evidence — signals that lean human
The detector also surfaces signals that lean human, balancing the verdict:
- Specific personal details. Names, places, dates, anecdotes that an LLM wouldn't have made up — and would have been less specific if it had.
- Idiomatic expressions. Slang, region-specific phrases, non-textbook word choices that LLMs use sparingly.
- Minor inconsistencies. Real human writing has small contradictions, opinion shifts mid-paragraph, parenthetical asides. LLM output is internally consistent to a suspicious degree.
- Opinion shifts. Human writers update opinions across paragraphs; LLMs hold a uniform stance.
- Casual asides. Side comments, brief tangents, signals that the writer is a specific person with a specific viewpoint.
Why no detector is courtroom-grade
Three fundamental limits that no AI-text detector can overcome:
- False positives are unavoidable. Some humans naturally write in clean, structured prose. Career-prose writers (journalists, lawyers, academics) often produce text that matches AI patterns. Flagging them as AI-generated is wrong.
- False negatives via editing. A careful human edit can mask any stylometric fingerprint. Run AI text through Improve Writing, vary sentence length, add specific examples, and the fingerprints largely disappear. The detector still works on raw AI output; it doesn't work on carefully-edited AI output.
- Detector evolution lags model evolution. Detection patterns are documented from previous-generation models. As new models come out, their default stylistic patterns shift. The detector adapts but with a lag.
How to use the verdict responsibly
Three patterns for using the output:
- As one input among several. Use the detector's verdict as a flag for further investigation, not as a conclusion. Pair with content-quality checks, source verification, and human reading.
- For self-review, not accusation. If you wrote a draft yourself and want to verify it doesn't read like AI before submitting, the detector is great for that. If you're using it to accuse someone of AI use, the inherent error rate means false accusations are guaranteed at scale.
- Combined with plagiarism checking. AI-text detection and plagiarism detection are different problems. Plagiarism (copying from existing text) is detected by tools like Turnitin / Copyleaks. AI generation is detected by stylometric tools. Both checks make sense in many contexts; neither replaces the other.
Limits and pricing
AI Content Detector charges 10 credits per document. The tool handles PDFs up to 25 MB. Processing runs on our servers; the document is in memory only during analysis and is never persisted. Output is a per-passage verdict with confidence, flagged patterns, counter-evidence, and optional humanize-edits suggestions.
Important: This is a heuristic, not a definitive classifier. Use the output as one input, not as the verdict.