My Supervisor Flagged My Dissertation As AI — Here's What Actually Fixed It

In March 2026, Priya sent her marketing dissertation draft to her supervisor at a Sheffield university. Three days later, she got an email back: two sections of her literature review were flagging as AI-generated. Her final submission deadline was eleven days away.

She hadn't submitted AI text. She'd used ChatGPT to help outline her argument structure, then written every sentence herself. But something in how she'd written — the rhythm, the patterns — had absorbed enough of the AI's voice that two detectors agreed: this looked machine-made.

Why does AI-influenced text get flagged even when you rewrote it?

AI detectors don't read for ideas. They read for patterns. Specifically, they measure things like how predictable each word choice is and how much sentence length varies. AI writing tends to be uniform: medium-length sentences, smooth transitions, textbook word choices. Human writing is messier. It lurches. It uses an odd word where a plain one would do. When Priya re-read her sections, she could see it — every paragraph opened with a topic sentence, every transition was "Additionally" or "This suggests that." She'd cleaned up ChatGPT's draft so thoroughly that she'd also erased the human irregularity detectors look for. Understanding how AI detectors work makes this make sense: the problem isn't the ideas, it's the texture of the prose.

What does it actually mean to humanise AI text?

Humanising AI text means restoring the statistical properties of human writing: unpredictable word choices, varied sentence lengths, occasional informality, and structural irregularities that human writers produce naturally. It's not about making writing worse — it's about making it less algorithmically smooth. There are two ways to do this: rewrite manually with real intention, or use a tool designed specifically to redistribute the linguistic patterns that detectors target.

What Priya tried first — and what didn't work

Her first pass was manual. She swapped synonyms, broke up long sentences, added a couple of personal observations. Then she ran the revised sections through a free AI detector. Scores came back at 74% and 81% AI probability. Not good enough. She tried a paraphrasing tool next. The score dropped to 61% — but the writing went awkward. Her supervisor would notice immediately.

On day four, she found WriteMask.

How she fixed both sections in under two hours

Humanising text with WriteMask is different from paraphrasing. It doesn't just swap words — it restructures sentence-level patterns to match human writing distributions while preserving the original meaning. Priya pasted her first flagged section (around 600 words) and ran it through the humaniser.

The result kept her argument intact. The examples were the same. The citations were the same. But the prose was no longer metronomic. Some sentences were short. Others ran long with a clause tucked in the middle. The word choices felt like something a postgraduate student would actually write at 11pm, not a language model completing a prompt.

She ran the output through the detector. 12% AI probability. Second section. Same process. 9%.

WriteMask has a 93% pass rate across widely used AI detectors, and Priya's results tracked that. Both sections cleared with real margin — which matters because academic AI detection thresholds are typically around 20–25%. If you've ever been wrongly flagged by an AI detector, you'll know that margin is everything.

Three things she said made the difference

• Run your own text first, even text you're confident about. Priya found a third section sitting at 43% AI probability — a section she hadn't touched with ChatGPT. She fixed it before it became a problem.
• Read the humanised output before submitting. WriteMask preserves meaning well, but a quick read-through catches anything that needs a small manual tweak to match your specific voice.
• Build this into your workflow, not your panic. The students who manage AI detection best aren't scrambling to fix flagged drafts. They check before submitting.

What happened in the end

Priya resubmitted her revised sections nine days before the deadline. Her supervisor confirmed they looked fine. Her dissertation was submitted on time with no AI misconduct flag on record.

She now runs every major piece of writing through a detector before sharing it — not because she writes AI text, but because she understands how the patterns work. Once you know what detectors actually measure, avoiding flags isn't luck. It's just knowing what to look for and having the right tool when it counts.

For a step-by-step walkthrough of the humanisation process itself, this guide on humanising ChatGPT text for Turnitin covers exactly what to do and why each step matters.