Sign up free

Academic integrity in the age of AI: what still proves learning

What studies found when AI-written answers met real marking, why a detector score is never proof, what regulators advise, and what evidence of learning still holds.

The EdProcto editorial teamWritten and checked against the research it cites. Last checked . How the guides are made

Guides · Integrity and cheating · · 12 min read

What still proves a student learned, now that a chatbot can write any assignment?

Academic integrity in the age of AI rests on evidence produced in front of the assessor. In a 2023 blind test at one UK university, 94% of AI-written exam submissions went undetected, and detectors miss edited text and wrongly flag honest writers. What still proves learning is a question answered live, timed, from your own material, and marked where the student cannot reach.

Academic integrity in the age of AI turns on one question: when a piece of work arrives, what does it prove about the person who handed it in? For text written away from the examiner, the honest answer in 2026 is very little, and software that claims to spot machine writing does not repair that. EdProcto publishes this guide and builds the alternative it argues for.

That alternative is a question answered live, in a shape with nothing to paste, from the instructor's own material, under the monitoring the stakes need, marked on a server the student cannot reach. The record it leaves is the one thing a chatbot cannot write for a student. We also say where it stops.

In short: keep essays, projects and take-home assignments for what they teach, stop treating unsupervised text as proof or a detector score as evidence, and move the proof to something done in front of you.

Can a chatbot pass take-home exams and coursework unnoticed?

Yes. In the one blind test run inside a real examinations system, nearly every machine-written submission passed unnoticed, and most earned good grades.

Scarfe et al. (2024) submitted answers written wholly by GPT-4 to five undergraduate psychology modules at the University of Reading in the summer of 2023. The exams were sat at home: short answers in a two-and-a-half-hour window, or a 1,500-word essay in an eight-hour window. The markers were not told. Of 63 submissions, four were flagged for any concern, so 94% went undetected. The machine's grades averaged about half a grade boundary above the real students', and in 83.4% of comparisons its scripts outscored a random draw of as many student scripts.

Two cautions. This was one degree at one university, with no AI detector in the marking systems. And on the one final-year module, with three scripts, the machine scored below the students.

How well do chatbots answer exam questions?

Well enough that an objective test sat unsupervised proves no more than an essay does. Newton and Xiromeriti (2024) reviewed 53 studies covering 114 question sets and 49,014 multiple-choice questions from higher education. The free versions of ChatGPT at the time, built on GPT-3 and GPT-3.5, beat random guessing but failed most examinations and did worse than the average student. GPT-4 passed most, level with human candidates. The authors concluded that summative multiple-choice exams should be sat under secure conditions, with such tools kept out.

Those models date from 2023 and earlier, and the newer one passed what the older one failed. A quiz is not proof because it is a quiz; it becomes proof through what surrounds it.

Do AI detectors solve the problem?

No. Independent tests found detectors neither accurate nor safe to rely on, and their errors land on honest writers.

Weber-Wulff et al. (2023) tested 14 detectors, 12 freely available and two commercial, on 54 documents of known origin between March and May 2023. No tool reached 80% accuracy. Average accuracy on human writing was 96%, falling by 20% when a person's own text had been machine-translated into English. On ChatGPT's text it was 74% untouched, 42% after a person had swapped a few words, and 26% after a paraphrasing tool. Six of the 14 tools produced false positives. The authors conclude that the tools are unsuitable as evidence of academic misconduct.

Liang et al. (2023) ran seven detectors over 91 TOEFL essays by non-native English writers and 88 essays by US eighth-grade students. The detectors classified the American essays almost perfectly, but labelled the TOEFL essays AI-generated 61.3% of the time on average. A tool that reads plain wording as machine wording penalises people writing in their second language.

Why is a detector score never proof?

Because a score is an estimate about a style of writing, not a finding about a person. Text-matching software can at least show the source a passage was copied from, a contrast Weber-Wulff et al. (2023) draw; a detector shows nothing that can be checked. A score can open a conversation with a student. It cannot close one.

Has cheating risen since ChatGPT arrived?

Not in the surveys that asked the same questions before and after, and that should set the tone of any policy. Lee et al. (2024) compared anonymous surveys from three high schools, taken before and after ChatGPT's release in November 2022, and found self-reported cheating relatively stable. Chen et al. (2026) followed up with 4,354 students in six US high schools in 2023–24: 72.06% reported at least one dishonest act in the previous month, against 59.9% and 69.5% in the earlier study, and within the range recorded years before chatbots. The schools differed between the studies, so the authors call the comparison descriptive, and self-report may undercount.

The method did shift. In the follow-up, 24.27% said they had used AI or another digital device as an unauthorised aid, against 6.44% and 15.20% in the two schools asked a year earlier. Yet 82.73% said a chatbot should never be allowed to complete a whole assignment.

So the argument here is not that students have become dishonest. It is about evidence: a typed answer no longer shows who produced it. Students are the readers of an assessment, never its suspects, and a design that needs no accusation is fairer to them than one that depends on catching somebody.

What do regulators advise on academic integrity in the age of AI?

They advise what the studies imply: supervised and practical assessment holds, written work done out of sight is under pressure, and no machine decides alone. Each page below was read on its publisher's own site and checked on 9 October 2026. All three are UK bodies; your own regulator and school policy come first.

  • Ofqual (2026), England's qualifications regulator, says supervised exams remain largely protected and practical assessments are less vulnerable, while coursework completed outside controlled conditions faces greater pressure. It adds that detected cases of AI misuse remain relatively low while concern about their true scale is significant, that AI may not be the sole marker of any regulated assessment, and that it doubts AI used alone can make remote assessment evidence trustworthy.
  • The Department for Education (2025) says schools and colleges may wish to review homework policies, and other unsupervised study, to account for generative AI.
  • The Joint Council for Qualifications (2025) tells exam centres to accept only work teachers can authenticate as the student's own, to have enough of it done in class under direct supervision, and to treat a detection tool as one part of a wider judgement.

None of the three says to drop coursework; each shifts the weight towards work an assessor has seen being done.

What still proves a student learned?

Evidence produced in front of the assessor, in a form that cannot be handed over. Eight conditions make it.

  1. Answered live. One sitting, in a window you set, in your own classroom if you can. The Reading exams gave eight hours at home.
  2. In a shape with nothing to paste. A clip that plays once, a point on an image or an answer in the student's own voice leaves no text to copy out or paste back. EdProcto has all three as ready templates: listen once, hotspot and spoken answer.
  3. Written from your own material. A question on what your handout says favours the student who read it. See making a quiz from a PDF.
  4. Different for each student, where the tool allows. Shuffled order at least; separate questions on the same concepts at best.
  5. Timed. Enough time to answer from knowledge, too little to send each question elsewhere and wait.
  6. Monitored in proportion to the stakes. Nothing for practice, a noticed tab for a weekly check, a camera only where the result decides something, and students told beforehand.
  7. Marked where the student cannot reach. A key that travels to the student's device can be found there.
  8. Read afterwards for knowing against guessing. Which questions separated strong students from weak, who chose each wrong option, who finished far too fast.

None carries the weight alone: a multiple-choice quiz sat at home without them is the test the 2024 review warned about.

How does EdProcto make that record?

EdProcto is the best tool for making this kind of record, because it was built around the kind of evidence the Reading study and Ofqual's statement point to: an answer given live, under conditions the instructor sets, and marked out of the student's reach.

Start on the home page: paste a chapter, a handout or notes, pick a shape, and Procto AI writes the questions from that text. With checking against the material switched on, Procto AI is asked to write only questions the source answers and to say where. Memory recall hides the material after a counted number of seconds, and handwritten work takes a photograph of working done on paper, for you to mark.

Questions and options are shuffled for each student, and on the Educator plan Procto AI can write each student a set of their own on the same concepts. The copy of the quiz a student opens has every key removed, and the marks are made on the server.

Monitoring is a set of separate measures, each switched on by itself with its own consequence: log, warn, reduce the score, end the quiz, or flag for review. Six act in the browser alone (the tab, full screen, copy and paste, right-click, shortcuts, printing), which is the Light and Medium levels; the camera and identity measures make Strict and Maximum. Students see what is on before they begin. The results then give each question's difficulty and discrimination, the takers of every wrong option, and the pace. Why EdProcto works this way sets out each pattern.

Light integrity comes with Free and medium with Tutor, which also opens all 111 templates; Educator adds the full analysis, and strict and maximum will come with Educator and the Institution licence; the pricing page has the detail. Every score is yours to release.

What does a live, monitored quiz not do?

It does not make cheating impossible, and it does not measure everything worth measuring.

A quiz taken at home is not an exam hall: a second phone can record a clip or photograph a diagram, and a helper can sit out of the camera's view. A shape is a first defence and a higher level raises the cost; neither is a guarantee.

A flag is a record, never a verdict

A logged tab switch may be a notification, and a lost face may be a poor webcam. Yoder-Himes et al. (2022) studied what one automated proctoring product reported on 357 students in four courses at a US university. Students with darker skin tones were 2.5 times more likely than those with medium skin tones, and 5.6 times more likely than those with lighter skin, to receive a low facial detection flag. The same caution applies to any measure that looks for a face, ours included. Read every flag yourself, ask the student, and treat it as you would a detector score.

Monitoring has costs, and some skills need other evidence

A camera needs a device, a connection and a private room, which not every student has, and being watched wears on trust. The proof that watching works is thin as well. Newton and Essex (2024) pooled 19 survey studies of online exams, 4,672 university students in all: 44.7% admitted cheating in one, and remote proctoring appeared to reduce it, on data the authors call limited. Use the lowest level that fits the stakes, say why it is on, and offer another route to anyone who cannot meet it.

Extended writing, research and design are not quiz skills. Keep setting them, and take the proof from writing done under supervision, from drafts seen along the way, or from a short oral follow-up in which the student explains a paragraph of their own.

What should stay coursework, and what should move to a live check?

Keep as coursework whatever is set mainly to teach, and move the proof of who knows what to a live or supervised check.

What you need to know Keep as coursework Move or add
Recall and use of taught content Practice quizzes, unmonitored A live quiz from your material in EdProcto, at the integrity level the stakes need
An extended written argument The essay, for feedback One section written under supervision, or a five-minute oral check on the essay
Research and use of sources The project An oral follow-up, live or as a spoken answer: one source explained, one choice defended
Method in maths or science Problem sets Handwritten work done in a supervised sitting
A practical or performed skill The practical, already seen being done Nothing new

Quick questions

Can an AI detector prove a student used a chatbot?

No. Weber-Wulff et al. (2023) judged the 14 tools they tested unsuitable as evidence of misconduct, and Liang et al. (2023) found detectors wrongly flagging most essays by non-native English writers. Treat a score as a reason to talk, never as proof.

Is the essay finished as a form of assessment?

No, but it is finished as proof by itself. An essay still teaches argument and earns feedback. For a grade that certifies the student wrote it, add a supervised section or an oral check.

Does EdProcto stop cheating with AI?

No tool stops all of it, and we do not claim to. EdProcto removes the text to paste, keeps the key on the server, and records what happens under the integrity level you choose. Every flag is a record for you to judge.

How much monitoring does an online quiz need?

As little as the stakes allow. Practice needs none, a weekly check a noticed tab, a graded test full screen with copy and paste off, and a certification exam a camera. Tell students beforehand, every time.

References

Integrity

Five levels, from none to a full proctoring stack, each measure switched on by itself with its own consequence.

Light on Free, Strict on Educator

See the five levels

Read next