How to Check Your Paper for AI: What Professors Really See

One in five students already uses ChatGPT for essays and dissertations — and lecturers know it. But how exactly do they detect AI-generated text? What methods are universities deploying, and what gives a paper away most reliably? This article explains the reality behind AI detection at universities today.

How widespread is AI use in academic papers really?

An internal survey of around 1,250 students in summer 2024 found that one in five reported having used ChatGPT for coursework and dissertations. The actual figure is likely higher — many students are unsure whether their use was permitted in the first place.

At the same time, many students are aware that AI use is handled differently across institutions. Policies vary widely: some universities allow AI as a tool with a declaration requirement, others prohibit it entirely. Students who don't know their own institution's rules are taking a serious risk.

How do lecturers detect AI-generated text?

Lecturers and examiners use two fundamentally different approaches — and often both at once.

Method 1: Technical AI detectors

Universities are increasingly deploying specialised software. Tools like Turnitin, GPTZero and Copyleaks analyse statistical patterns in text and return a probability score for AI generation.

Detection rates for unmodified ChatGPT text sit at around 74%. Manually revising or paraphrasing the text reduces the technical detection probability — but does not eliminate it entirely.

Method 2: Human expertise — and why it matters more

Experienced lecturers know their students' writing styles. A sudden, pronounced change in the quality of expression is for many of them the strongest signal of all. No algorithm can replicate this kind of contextual knowledge — a supervisor who has worked with a student for two semesters will immediately notice if Chapter 3 of a dissertation reads in a completely different style from Chapters 1 and 5.

The five most common detection signals

These are the signals most frequently flagged by both lecturers and AI detectors as indicators of AI-generated writing:

• Uniform sentence structure: AI produces text with a noticeably regular sentence length and rhythm. Short interjections, rhetorical questions and deliberate stylistic breaks are almost entirely absent. Human writers vary far more — sometimes long, sometimes short, sometimes fragmentary.
• Formulaic phrasing: Certain phrases are characteristic of AI output: "It is important to note that…", "In summary, it can be said that…", "These foundations form the basis for…". Each one in isolation is unremarkable — repeated throughout a paper they form a recognisable pattern.
• Invented or incorrect references: AI systems hallucinate sources — they fabricate author names, page numbers and publication years that sound plausible but don't exist. Examiners who spot-check citations catch this immediately.
• Style shifts between chapters: The so-called "sandwich pattern" — a human-sounding introduction and conclusion with an AI-generated body — is one of the most common patterns in student work. The score difference between chapters makes it immediately visible.
• Shallow content despite polished language: AI describes, explains and summarises — but rarely develops a coherent, original line of argument. The text sounds academically competent, but lacks genuine intellectual substance.

What happens when AI use is detected?

The consequences depend on the institution and the extent of undeclared AI use. The range is wide:

Why "just paraphrasing it" isn't enough

Many students try to "humanise" AI-generated text by manually revising it or running it through paraphrasing tools. This reduces the technical detection rate — but doesn't eliminate it, and it does little against the expert eye of a supervisor who knows your writing.

The real problem runs deeper: AI text that has only been superficially reworked retains its structural characteristics — the even sentence rhythm, the lack of argumentative depth, the generic way sources are integrated. These patterns are much harder to disguise than individual word choices.

Why students check their own work

Not every student who uses an AI detector has actually used AI to write their paper. Many check their work for a different reason: they want to make sure that text they've polished with AI tools — for grammar or style, for instance — won't be flagged as AI-generated.

What ThesisScan does differently

Most AI detectors analyse a document as a whole and return a single score. The problem: an academic paper is not a uniform text. Introduction, theory, methodology and conclusion can each have very different AI profiles.

ThesisScan analyses text chapter by chapter — making the sandwich pattern visible in a way that a single overall score would conceal. The report gives you a separate score for each chapter, highlights the most suspicious passages with a specific explanation, and — on the Standard plan — provides suggested rewrites for flagged phrasing.