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

Priya is a second-year public health master's student at a UK university. Last semester, her dissertation supervisor pulled her aside after a chapter submission and said four words that sent her spiraling: "Did you write this?"

She had. Every word. But something about the phrasing — clean, structured, academic — had triggered her supervisor's AI detector. Priya spent a week trying to understand what "AI-generated" even means in the context of her own writing. We sat down with her and an AI detection specialist to get some real answers.

What Does It Actually Mean to Check If Text Is AI Generated?

Q: When someone says they want to check if text is AI generated, what are they actually measuring?

A: They're measuring statistical patterns. AI models like ChatGPT tend to produce text that is highly predictable — each word flows into the next in ways that feel smooth but are actually quite uniform. Detectors assign a probability score based on how "expected" each word choice is. High predictability means the tool suspects AI. Low predictability reads as more human.

The catch? Formal, structured writing — academic papers, technical reports, policy documents — naturally scores high on predictability. It's not AI. It's just following genre conventions. That's exactly what happened to Priya.

Q: So even human-written text can look like AI to a detector?

A: Absolutely. This is the core problem, and it's more common than most people know. AI detection false positives hit hardest for ESL writers, people who write in formal registers, or anyone whose style happens to be clear and consistent. There's no clean threshold separating "definitely human" from "definitely AI." It's all probability, and probability makes mistakes.

How Do You Actually Check Your Own Text Before Submitting?

Q: Priya wants to check her own writing before her supervisor does. What's the right approach?

A: Run it through the same kind of detector your institution or client is likely using. The free AI detector on WriteMask gives you a realistic read and — this part matters — it flags which specific sentences are triggering the highest AI probability scores, not just a single number to panic about.

That sentence-level breakdown changes everything. A passage that scores 74% AI overall might have three specific sentences driving the whole result. Those are the ones to rework. The rest of your writing is probably fine.

Q: What should I actually look for when I check my text?

A: Three things, specifically:

• Perplexity — how predictable the word choices are. Low perplexity means the detector thinks each word was too "obvious" a follow-on from the last.
• Burstiness — whether your sentence lengths vary. Human writing tends to mix long, complex sentences with short, punchy ones. AI flattens this out into a steady rhythm.
• Sentence-level flags — which lines are actually triggering the score. A good detector highlights these so you know exactly where to edit.

Q: Understanding how AI detectors work sounds technical. Do I need to understand all of this?

A: Not deeply. But knowing that detectors look for predictability — not plagiarism, not copied text — helps you understand why formal writing flags even when it's 100% original. Once you know that, you can make targeted edits: vary your sentence rhythm, swap out predictable transition phrases, let a bit of your own voice show. You don't have to rewrite everything from scratch.

What If Your Text Checks as AI Even Though You Wrote It?

Q: That's Priya's exact situation. She wrote it. It flagged. Now what?

A: Two steps. First, document your process. Drafts, outline notes, timestamps — how to prove your essay is human almost always comes down to showing a paper trail that no AI would have left. Then, actually fix the flagged text so it stops triggering the detector.

This is where a tool like WriteMask does something that manual editing can't easily replicate at scale. It humanizes text by introducing the kind of variation detectors look for as a signal of human authorship. The 93% pass rate reflects how consistently it disrupts those statistical patterns — without changing your meaning or argument. You're not making the writing worse. You're making it less predictable in the specific ways detectors care about.

Q: Does humanizing text risk making it sound worse?

A: If done clumsily, yes. But a tool built around sentence-level detection data targets the predictability patterns, not your word choices. Priya found her chapter actually read slightly better after — the rhythm was less uniform, which made the argument easier to follow. That's often how it goes.

The One Thing to Do Before You Submit Anything

Q: Quick summary — what's the most important step?

A: Check early and check specifically. Don't just run the whole document and stare at one big percentage. Use a detector that shows sentence-level results, fix what's actually flagging, and give yourself time to revise before the deadline. Ten minutes of checking before you submit beats a difficult conversation with a supervisor afterward.

For Priya, running her chapter through a sentence-level check and spending 40 minutes on targeted revisions dropped her score enough to resubmit confidently. The problem almost never requires rewriting everything — just finding the few sentences that are doing most of the damage.