Claude · The mechanism
How AI writing detection works, and why it is not proof
Once you see how it works, you understand it. Same for why it breaks.
AI writing detection is statistical scoring. The tools measure how predictable text looks. They use signals from the perplexity and burstiness families. They do not read watermarks or C2PA metadata. The accuracy record is poor. OpenAI discontinued its own classifier in 2023. It caught 26% of AI-written text. It mislabeled 9% of human text as AI. A 2023 Stanford study found detectors flagged more than half of genuine non-native English essays as AI-generated. University libraries call detectors problematic and not recommended as a sole indicator of misconduct. Two universities have disabled the detector feature in their integrity tools. Plain text has no C2PA container at all, so provenance checks on it rest on statistical scores. The first independent audit of C2PA found it should not yet be relied upon for high-stakes uses. A score is a probability, not proof.
Why this page exists
This is the reference page for the head term. Most AI writing detection searches are tool-focused, but the people looking for information want the mechanics and the accuracy record. The other hub pages link here instead of explaining detection again.
The tells that show up here
Not every AI tell appears everywhere. These are the ones that cluster on this surface, and what it is about the format that produces them.
- vague-attribution
- 'Studies show' without dates or sources is the exact habit this page must not have.
- not-x-its-y
- The core distinction (probability vs proof) risks the parallel tic.
- testament-family
- Explainer slop reaches for 'testament' when describing detector limits.
Before and after
A generic AI draft on the left. On the right, the real output of running that draft through underslop, checked so it adds no name or number the draft did not already carry. The voice is The Steady Hand.
AI detection represents a groundbreaking leap forward in the fight for authenticity! It's not just about algorithms; it's a testament to the ever-evolving battle between creation and verification. Research consistently shows that modern detectors achieve near-perfect accuracy, serving as a cornerstone of trust in today's fast-paced digital world!
AI detection is a practical tool for checking authenticity. It scores how predictable text looks. The tools keep changing as creation and verification change. Some detectors are accurate on their own tests. Their record on real writing is mixed. A score is a signal, not a verdict.
How to do it by hand
- Keep the receipt Save the drafts, the notes, and the version history now. If a score is ever raised, you’ll want the record of how the text was made.
- Check the tool's limits Ask what the tool can and cannot do. Its score is a probability. The record shows a 26% true-positive rate and a 9% false-positive rate for the classifier OpenAI itself withdrew.
- Look at the actual text patterns Read the sentences that got flagged. The score rests on predictability, so careful, formulaic, or second-language writing still reads as predictable, and it stays human.
- Talk to the person who decides A flag is not a verdict. Institutions now use scores as one indicator, not the only one. The person deciding your case may not know the tool's record.
Questions
How do AI detectors work?
They score text by how statistically predictable it looks, using signals like perplexity and burstiness. They don't check watermarks or C2PA metadata.
Are AI detectors accurate?
The record says no. OpenAI dropped its own classifier in 2023. It caught only 26% of AI text and wrongly flagged 9% of writing by people. A 2023 Stanford study found the detectors marked more than half of genuine essays from non-native English speakers as AI.
Can a detector prove text is AI-generated?
No. University libraries call detectors "problematic and not recommended as a sole indicator" of misconduct. Universities that reviewed them closely turned the feature off rather than trust it.
Do detectors read C2PA metadata?
No. Detector products work on statistics and do not read C2PA. Plain text has no C2PA container at all, so there is nothing for them to read. The first independent audit of the standard found it should not yet be relied on for high-stakes uses.
Why do some universities disable AI detectors?
Because of the record. The University of Waterloo stopped its AI detection in September 2025. Curtin University turned off its AI writing detection on January 1, 2026. Libraries advise against using one score alone.
Claude marks are real and the coverage is still settling. underslop is the edit, not a remover: it edits an AI draft so it reads like you wrote it, and it makes no promise about what any detector will report. Your text is held for 0 seconds.