You finish an assignment. The references are checked, the wording is tidy and the argument holds together. Then you paste it into an AI detector and it says most of it looks AI-written.

Second tool: a much lower score. Third tool: “mixed signals”. Your stomach drops a little, because nothing about how you wrote it has changed.

If that sounds familiar, you are not unlucky. You are running into how these tools actually work. The disagreement is normal, and it tells you more than any single score does.

What a Detector Actually Sees

It sees the finished text and nothing else. It cannot see your notes, your drafts, the source you dropped late one night or the paragraph you rewrote three times.

So it estimates from patterns. Is the wording predictable? Is the rhythm very even? Does it resemble text a language model tends to produce?

Academic writing leans on those same habits: formal tone, tidy paragraphs, standard terms. That is not a flaw in your writing. It is a big reason detectors struggle with university work.

Why Three Tools Give Three Answers

Each detector has its own training data, its own method and its own cut-off for saying “AI”. Move that cut-off a little and the same sentence can flip from human to machine.

Picture three markers using three different rubrics on one essay. You would expect different grades. Nobody would call it a mystery.

Percentages add to the confusion. A high number often means “the tool is fairly confident”, not “most of your words were written by AI”. Check what the tool itself says its score means.

A score is an estimate, not a verdict.

The Two Ideas Behind Many Scores

Perplexity describes how predictable your words are to a language model. Burstiness describes how much your sentence length and structure vary.

The problem is that a definition of a well-known theory has very few accurate ways to be written, so it will look predictable. A formal report is also meant to keep a steady style.

Neither measure proves who wrote the text. And forcing uneven sentences just to beat burstiness makes your writing worse, not more human.

The Pressure to “Fix” It

This is where people make costly choices. A high score lands, the deadline is close, and searching for an AI Removal Service Online UK starts to feel like a sensible next step.

Be careful. Lowering a score and improving your work are two different jobs. No service can promise a particular result across every detector, and rewording sound sentences to dodge a tool often damages accuracy.

A better first move is to ask whether the flag means anything.

Where Detectors Are Weakest

Some situations make a misleading result more likely:

• Short passages. A small sample gives the tool little to work with. Many detectors state a minimum length in their own guidance.
• Fixed-term subjects. Law, nursing, accounting and engineering rely on set wording.
• English as an additional language. A 2023 Stanford-led study, published in the journal Patterns, found that several popular detectors wrongly flagged TOEFL essays by non-native writers far more often than essays by native writers.
• Heavy editing. Grammar tools and proofreading change rhythm and vocabulary, which can shift a result either way.

None of these is a sign of cheating. They are simply the places where a pattern-based tool has the least to go on.

A Practical Check Before You Change Anything

These steps will not prove anything, and I would not present them as proof. They help you decide what deserves attention.

1. Read the tool’s guidance. Find out what the score represents, what text length it needs and what limitations the developer lists.

2. Look at the flagged passages, not the headline number. Are they definitions, standard explanations or technical terms? Those are the likeliest false alarms.

3. Use a second tool as a rough cross-check. If two tools highlight the same passage, look at it more closely. Tools can share blind spots, so agreement is a prompt to review, never evidence of AI use.

4. Explain the passage out loud. Talk through a flagged paragraph without reading it. If you can say why you chose that source and what it shows, you can account for the work.

Optional: a personal comparison. If you have older writing you know you wrote yourself, run it through the same tool. The different topic and length limit what it proves, but it shows how that tool reacts to your style.

Keep an Evidence Trail

Keep dated drafts, research notes, a rough outline and any version history. They show how the work grew, which a finished document cannot.

Version history depends on how and where you save. Google Docs saves automatically and keeps version history, though what you can see depends on your access. Word’s depends on saving to OneDrive or SharePoint with AutoSave switched on. Check your settings before you need them.

Add one habit: a short decision log. One line per big change, such as “Dropped Smith 2019, sample too small.” It takes seconds and shows real thinking.

If Your University Raises a Concern

Read the academic integrity policy first. Ask which passages were questioned and what evidence is being used.

Then explain your process calmly and honestly, and share your trail. The aim is a fair judgement based on all the evidence, not winning an argument about the tool.

Check the rules on tools too. Proofreading help and generative AI are often treated differently, and some courses require disclosure.

Mistakes Worth Avoiding

• Treating a high score as proof of misconduct. False positives exist.
• Treating a low score as proof of safety. Detectors miss generated text too.
• Rewriting everything in a panic. You can damage clear passages.
• Mistaking polished for suspicious. Clear, organised writing is good academic work.

The Takeaway

Two detectors disagree because they measure different things, in different ways, with different cut-offs. That does not make them useless. It means results need careful reading.

Understand your rules, keep honest records and be ready to explain how your argument developed.

A detector can raise a question. Answering it properly takes more than a percentage.

Before publishing, check the Stanford study (Liang et al., 2023) yourself so you can link it directly.