I Wrote It Myself — So Why Did the AI Detector Flag It? A Q&A on How to Actually Check Your Text

AI detectors don't read text. They model probability distributions over token sequences and flag output that looks statistically too predictable. Understanding that distinction — measurement versus judgment — is the key to diagnosing why text fails a detector check, and how to fix it.

It's also why Priya, a second-year public health master's student at a UK university, found herself in a surreal situation last semester: her dissertation supervisor asked "Did you write this?" after flagging a chapter she had written entirely herself. The culprit wasn't AI. It was formal academic prose doing exactly what it's supposed to do.

## The Underlying Measurement: What Detectors Actually Compute

When a detector scores a passage, it's calculating how expected each word choice is given the surrounding context. Language models like ChatGPT are trained to maximize coherence, which produces output that is statistically smooth — each token is a high-probability continuation of what came before. That smoothness is the signal detectors exploit.

The two core metrics are **perplexity** and **burstiness**. Perplexity measures how "surprised" a model would be by each word choice; low perplexity means the text was predictable, which reads as suspicious. Burstiness measures variance in sentence length — human writers naturally alternate between long, complex constructions and short punchy ones, while AI-generated text tends to settle into a uniform rhythm. Flat burstiness is another flag.

The practical problem: formal registers — academic papers, technical reports, policy documents — score poorly on both dimensions by design. They follow tight genre conventions that produce exactly the kind of smooth, predictable prose detectors are trained to catch. This creates systematic [AI detection false positives](/blog/false-positives-ai-detection) that hit hardest on ESL writers, people writing in highly structured genres, and anyone whose style is consistently clear. There's no clean binary threshold. It's probability, and probability has error rates.

## Running a Diagnostic: How to Check Your Own Text

The right approach before submitting anything is to run the text through the same class of detector your institution or client is likely using. The [free AI detector](/detect) on WriteMask returns not just an aggregate score but a sentence-level breakdown showing which specific sentences are driving the result — and that distinction matters enormously.

An aggregate score of 74% sounds alarming. But if three sentences are responsible for pulling the entire passage over threshold, the actual editing surface is small. The rest of the document is probably fine. Without sentence-level output, you're debugging with no stack trace.

When reviewing results, look for three specific signals:

  - **Perplexity score per sentence** — sentences where every word was the "obvious" next choice are the primary offenders
  - **Burstiness across the passage** — a sequence of similarly-lengthed sentences with parallel structure will flatten this metric
  - **Flagged line locations** — clustering matters; a section written under time pressure or in a more formal voice than the rest will often show up as a contiguous block of flags

For a deeper read on the detection pipeline, [how AI detectors work](/blog/how-ai-detectors-work-2026) covers the technical internals — but you don't need to internalize all of it. Knowing that detectors target predictability rather than copied content is enough context to make edits that actually move the score.

## When You Wrote It and It Still Flagged

Priya's situation is the common case, not the edge case. When text you authored gets flagged, the fix has two parallel tracks: documentation and revision.

On documentation: gather everything that proves a process — draft versions, outline notes, timestamps, research annotations. [Proving an essay is human-written](/blog/how-to-prove-my-essay-is-not-ai-written) almost always comes down to a paper trail no AI would have generated. Supervisors and reviewers are more receptive to "here's my working history" than any technical argument about false positives.

On revision: [WriteMask](/dashboard) handles the statistical side by introducing the variation patterns that detectors interpret as signals of human authorship — sentence rhythm variance, less predictable transition choices, controlled shifts in sentence length. The 93% pass rate reflects how reliably it disrupts the specific patterns detectors are tuned to detect, without touching your argument or meaning. Priya found the revised chapter actually read more fluidly than the original; less uniform rhythm often makes prose easier to follow.

Clumsy humanization makes writing worse. A tool operating on sentence-level detection data targets the perplexity and burstiness metrics directly, rather than randomly paraphrasing content. The edits are surgical.

## Pre-Submission Checklist

The failure mode to avoid is treating this as a one-number problem. Running a document through a detector and staring at a single percentage tells you almost nothing actionable. The correct workflow: check at the sentence level, identify which lines are actually triggering the score, revise those specifically, and do it early enough to have time before the deadline.

Priya spent 40 minutes on targeted revisions after a sentence-level check and resubmitted confidently. The issue was never the whole chapter — it was a handful of sentences written in a particularly compressed, formal register. Finding those sentences is the hard part. Fixing them, once identified, rarely requires rewriting more than a paragraph or two.