Falsely Accused of Using AI? A Practical Defence Guide

You are not imagining it

Being told that software believes you cheated — when you did not — is genuinely distressing. It is also more common than most institutions admit. Understanding why it happens is the first step to answering it.

Why honest writing gets flagged

Detectors do not identify AI. They identify predictability. Any writing that is smooth, evenly paced, and conventional in vocabulary will score as machine-like, whoever wrote it. The people most likely to write that way include:

• Second-language writers, who lean on reliable structures and a narrower, safer vocabulary.
• Students trained in rigid academic formats — five-paragraph essays, IMRaD lab reports, standard legal or medical templates.
• Careful editors. Ironically, polishing away every awkward sentence removes exactly the irregularity detectors read as human.
• Neurodivergent writers whose natural style is highly systematic.

None of that is cheating. It is just prose that happens to sit in a statistically inconvenient place.

What to do in the first 24 hours

1. Do not rewrite anything

The instinct to "fix" the document is understandable and harmful. Your existing file, with its metadata and revision history intact, is evidence. Leave it exactly as it is.

2. Collect your process trail

Gather everything that shows the work developing over time:

• Version history (Google Docs: File → Version history; Word: File → Info → Version History)
• Draft files with their original timestamps
• Handwritten notes, outlines, mind maps — photograph them
• Browser history or library records showing your research
• Messages where you discussed the assignment with anyone

3. Ask what the score actually was

Request the specific figure and the tool used. A 40% "AI likelihood" on a 400-word essay is a very different claim from 95% on 3,000 words — and many institutions will not have considered that distinction.

4. Respond in writing, calmly

Keep it factual. State that you wrote the work, that you understand the tool produces probabilistic estimates rather than proof, that you can provide process evidence, and that you are willing to discuss the content of the work in person. Offer, do not demand.

The strongest argument you have

Ask to discuss your own argument. Nobody who did not write a piece can defend its reasoning, explain why they cut a paragraph, or say what they nearly argued instead. A ten-minute conversation is far more probative than any percentage, and most reasonable assessors know it.

Points worth raising politely

• Detector vendors themselves publish confidence caveats — ask whether the institution has reviewed them.
• Peer-reviewed research documents elevated false-positive rates for non-native English writers.
• Several universities have disabled AI detection outright over reliability concerns.
• Ask whether the policy permits a finding based on a single automated signal with no corroboration.

Protecting yourself going forward

• Write in a tool that keeps version history, and never disable it.
• Keep an outline file for every substantial assignment.
• If your institution permits assisted work, disclose it in a footnote every time.
• Run a style check on your own drafts so a flat passage is never a surprise — you will see which sentences read as mechanical before anyone else does.

A word on fairness

We build detection software, and we think the current norm — treating a probability as a verdict — is indefensible. That is why Sahihly reports a confidence rating alongside every score, highlights the specific sentences driving the result, and states plainly in its output that scores are guidance for improving writing, not evidence of misconduct. If a tool cannot tell you why, it should not be used to decide anything about a person.