An online midterm exam, set by one lecturer for sixty to two hundred students at home, can be worth sitting with no proctoring service behind it. What it needs is a paper on which looking things up gains little, a plain briefing, and a plan for the day.
This guide follows one exam from the first question to the mark sheet. EdProcto publishes it and is the paper in the worked example; its page for universities shows what else a course team gets. Where a department puts its supervised exams is a wider decision, covered in assessment design after chatbots.
What does research since 2020 say about unproctored online exams?
It says that scores run somewhat higher than in a watched room, that they still rank students much as supervised exams do, and that a pledge changes little where a specific warning changes more. Most of it comes from single universities in the pandemic terms, before chatbots were common. We read each study on 10 October 2026 and say where only the abstract was open.
Are unproctored scores inflated?
Usually, by a modest margin. Steger et al. (2020) pooled 109 effect sizes from 49 studies of ability tests: unproctored scores were higher by 0.20 of a standard deviation, and significantly less so on tasks that are hard to search for online (abstract read).
Chan and Ahn (2023) studied 18 courses, with nearly 2,000 students, at one American university that moved from invigilated exams to unproctored online ones halfway through a term in 2020. Each student's online average correlated with their in-person average at r = 0.59. The link was positive in every course and did not vary significantly with question type, enrolment or exam duration. Between the halves of fully in-person semesters it was 0.74. The authors concluded that cheating was either uncommon or did little for scores.
Newton (2023) replied in the same journal. The paper's supplementary data, he wrote, put online scores higher by 0.4 of a standard deviation, and 0.7 on multiple choice; cheating is the simplest explanation; and with tools such as ChatGPT, unproctored exams that count should be avoided. Both readings can hold: the order of the students survived, and the level rose. How much proctoring a quiz needs has the proctored comparisons.
Do honour statements work?
Not alone, in the two online studies we read. In an online introductory biology course, Pleasants et al. (2022) could see when a student left the test page for other material. Some 70% did, most of them on most questions. Appeals to honesty and a required pledge were ineffective. A warning that the software could detect it, with a threat of harsh penalties, brought the share down to 15%. The authors found no evidence that exam performance changed when the behaviour did (abstract read).
Klijn et al. (2022) sent a reminder of the university's code of ethics to a subgroup halfway through an unproctored online exam; it made no difference (abstract read).
Do shuffling and question pools stop answers being shared?
Pools do more than shuffling. The exam Klijn et al. (2022) studied, for almost 500 undergraduates at a Spanish university, gave the problems in different orders with no backtracking. Problems met in a later round were answered correctly 7.7% more often, and 18.1% faster, than the same problems met earlier. The authors estimate that 13.4% to 22.5% of students cheated on information flowing from earlier rounds to later ones. Grades still tracked earlier coursework, so they call them informative.
Earlier, Chen et al. (2018) followed 425 students whose exams could be sat on different days and identified about 5.5% as receiving information from earlier sitters. That advantage was 13 percentage points when all had the same problem, and 2 to 3 when each drew one from a pool of two or four (abstract read).
Help also comes from outside. Lancaster and Cotarlan (2021) counted requests on the homework-help site Chegg in five science and engineering subjects: the daily average from April to August 2020 was 196.25% above the same months of 2019, and a request could be answered within the length of an exam.
How do you design the paper so that looking things up helps little?
Ask students to use the course on material they have not seen, in answers that are typed and worked. The moderator in Steger et al. (2020) points that way: the gap was smaller where a search could not easily find the answer.
| Design choice | Why | In EdProcto |
|---|---|---|
| A choice on a new case, not a definition | A search finds the definition, not your case | Multiple choice, marked by rule |
| A number worked from data printed in the question | No answer to fresh figures is waiting online | Estimate, a number inside your tolerance |
| Two or three typed sentences that explain a result | A reader sees whose reasoning it is | Short answer, kept for your mark |
| One problem typed in steps | Method earns the marks; a borrowed final figure earns few | Show your working: set-up, working, final answer |
| An order of their own, and a parallel version | Order alone still leaked in Spain; a second version is a pool of two | Both shuffle switches; a second quiz for version B |
| Open book, said plainly | Nobody sees the desk, so allow what you cannot check | "Open book", printed on the start screen |
Questions above recall shows how to write the first row, show your working the fourth, and open-book quizzes that still measure the last.
How tight should the exam time limit be?
Tight enough that little is spare for a search on every question, loose enough that a prepared student finishes. No study we read tested a tight limit as a defence, and Gernsbacher et al. (2020) give four reasons against time pressure: it makes a test less valid, less reliable, less inclusive and less equitable (abstract read). Sit the paper yourself before you fix "Minutes overall". A minute a mark with a fifth on top is our working rule, not a finding.
What about a chatbot on a second device?
It can attempt any typed question, and no paper design removes that. A case from your own lectures, figures of your own and working typed step by step shrink what it adds. Where a mark must be certain, hold the exam in a supervised room with EdProcto as the paper; academic integrity in the age of AI gives the evidence.
What should you tell students before an online midterm?
Tell them a week ahead what is allowed, what is logged and what to do when something breaks. The fall to 15% in Pleasants et al. (2022) followed a specific warning, not a pledge. A briefing that fits one message:
- When. One sitting. Begin between 10:00 and 10:10; ninety minutes run from your own start.
- Allowed. Your notes and the course text. Not the web, an AI assistant or another person.
- The paper. One question a page, in an order of your own. Each must be answered before the next opens, and there is no going back. Plan a minute a mark.
- Logged. Leaving the tab or full screen, reloading the page, and any attempt to copy, paste, print, open the right-click menu or use a shortcut. An event is a record, and you will be asked before anything is concluded.
- Device. A laptop or desktop, one browser, an ordinary window. Type your name and university email as the register has them.
- If the connection drops. Reopen the same link in the same browser and press "Back to full screen". The clock keeps running, so message me if you lose more than ten minutes. If that device will not come back, message me before you try another: a second device stays shut until I let you in.
- Extra time or a clash. Tell me by Friday and you get a link of your own.
EdProcto's start screen says it again: the time limit, "No going back: each answer is final", every measure that is on with its consequence, and the house rules. The student ticks a box to agree before pressing "I’m ready, begin the quiz".
How do you run the day: a practice run, a window and a dropped connection?
Rehearse once under the real settings, open and close the window yourself, and let the saved answers carry a student through a lost connection.
A practice run
A few days before, set five questions that count for nothing under the same preset. Students meet full screen, the typing boxes and the no-return rule while nothing depends on it. Ask each to reload the page once on purpose, press "Back to full screen" and see the attempt come back. Anyone who cannot get in tells you now.
The window
You open the paper with "Publish & share" and end it with "Close the quiz". Each clock starts when its student begins, and at zero the browser hands the paper in. On the results page, "In the quiz now" lists everyone still inside with progress, time and events, refreshed every few seconds; watching a live quiz says what to look for. Close only when that list is empty.
Seat the whole class at once: the advantage Chen et al. (2018) measured came from exams spread over days. For a clash or a distant time zone, hold one reserve sitting on version B. For extra time, make a second quiz with the same questions and a longer limit.
When a connection drops
EdProcto saves each answer to its server as it is given. A student who loses the connection, reloads or closes the tab opens the same link in the same browser, and the attempt returns as it was last saved. With no going back it opens on the furthest question the student had reached, never on one already passed. The browser has left full screen, so the questions wait behind a "Back to full screen" button, and "Reloaded the page" joins the event log. The clock counts from the moment the student began, so time away is time used. If the time has gone, the paper is handed in at once with what was saved, and a hand-in more than two minutes over carries the line "Handed in after the time ran out". We would give the reserve sitting to anyone who lost over ten minutes.
A second browser or another device is a different matter. Because this paper asks for an email address and allows one attempt, a start under an address whose attempt is still open is refused: no fresh paper and no fresh clock. In "In the quiz now" that student's last event reads "Asked to open on another device", and the "Another device" column holds a button, "Let back in". Press it and the student's next start opens the same attempt, with its order, its saved answers and its running clock, logged as "Reopened on another device". It works once. The address is typed, not checked, so let in only a student who has told you their device failed.
How do you read the event log afterwards?
As a record to weigh beside the answers, never as proof. Most students in Pleasants et al. (2022) left the test page, with no sign that scores depended on it: an event says what a browser did, not why.
- Grade first. Typed answers wait in each attempt, marked "to grade", and count as zero until you give them points. See grading open answers.
- Read the questions before the students. A question most of the class missed may be the paper's fault: see when to throw a question out.
- Find the attempts with events. EdProcto's results page shows "clean" or a count beside each student.
- Open each one. "Security events" gives the time, what happened ("Left the quiz’s tab", "Tried to paste") and the consequence. Each answer below shows the time spent on it.
- Weigh it. One exit from full screen is noise. Twelve, beside a flawless worked problem typed in ninety seconds, is a question to put.
- Ask the student to talk you through that answer. What they can explain settles more than a log.
- Decide under your institution's procedure, note why, and only then press "Release the scores".
What does a ninety-minute online midterm look like?
It gives 75 marks in 75 working minutes and keeps 15 in hand. This example is invented: a second-year course of 140 students, one lecturer, one teaching assistant.
| Part | Questions | Marks | Minutes |
|---|---|---|---|
| A choice on a new case | 15 | 30 | 30 |
| A number from data in the question | 5 | 15 | 15 |
| A typed explanation | 3 | 15 | 15 |
| One worked problem in steps | 1 | 15 | 15 |
| Total | 24 | 75 | 75 |
A rule marks 45 of the 75 marks at hand-in. The other 30 need a reader: 420 explanations and 140 worked problems. At a minute and a half and four minutes each, our assumptions, that is about 20 hours between two markers.
The settings, by their labels:
- In EdProcto, choose the Exam preset: no going back or skipping, both shuffles, six browser measures, scores held until release.
- Set "Minutes overall" to 90 and leave "Submit by itself when time runs out" on.
- Leave each measure at "Log it". A penalty set off by a browser event is a verdict with no hearing.
- Switch "Open book" on; keep "The web" at "Not allowed" and "AI assistants not allowed" on. All three print house rules on the start screen.
- Press "Preview as student" and sit it yourself.
How does EdProcto run an online midterm?
EdProcto is the best tool for an online midterm one lecturer runs alone, because it does what the studies above reward: it tells every student what is logged before they begin, and it holds every score until you have read the paper and the log.
On the home page, Procto AI drafts questions from your lecture notes for you to edit. The Exam preset sets Medium integrity: full screen, the tab, the clipboard, the right-click menu, shortcuts and printing, all in the student's ordinary browser. The copy a student's browser holds carries no key, and the marks are made on the server (why it is built this way). Typed answers wait for you, or for Procto AI when you press "Ask Procto AI" on one, its mark yours to change. "Download the marks" gives a spreadsheet row for each attempt: hand-in time, minutes, points on every question and the count of events.
Free takes 30 students a quiz at Light integrity. Tutor takes 100 and saves the Exam preset, so a course of 60 fits. A course of 101 to 200 needs Educator, which takes 500 and adds the full analysis. See pricing, and the plans for institutions for a whole programme. No setting makes cheating impossible, a logged event is a record and not proof, and every score is yours to release.
Quick questions
Can an online exam be fair without proctoring?
Fair enough to rank students: Chan and Ahn (2023) found r = 0.59 between unproctored online and invigilated scores. Expect a higher average, and use a supervised room where a mark must be certain.
Do honour codes stop cheating in online exams?
Not alone. Pleasants et al. (2022) found a pledge ineffective, while a warning that leaving the page was detected and would be penalised cut it from 70% of students to 15%.
What happens if a student's internet drops during the exam?
In EdProcto the saved answers stay on the server, and the same link in the same browser reopens the attempt. A second device stays shut until you press "Let back in" beside the student's name. The clock keeps running, so agree beforehand what lost time earns a reserve sitting.
Which EdProcto plan holds a course of 200?
Educator, at 500 students a quiz; Tutor holds 100. The page for universities has the rest.
References
- Chan, J. C. K., & Ahn, D. (2023). Unproctored online exams provide meaningful assessment of student learning. Proceedings of the National Academy of Sciences, 120(31), Article e2302020120. https://doi.org/10.1073/pnas.2302020120
- Chen, B., West, M., & Zilles, C. (2018). How much randomization is needed to deter collaborative cheating on asynchronous exams? In Proceedings of the Fifth Annual ACM Conference on Learning at Scale (Article 62). Association for Computing Machinery. https://doi.org/10.1145/3231644.3231664
- Gernsbacher, M. A., Soicher, R. N., & Becker-Blease, K. A. (2020). Four empirically based reasons not to administer time-limited tests. Translational Issues in Psychological Science, 6(2), 175–190. https://doi.org/10.1037/tps0000232
- Klijn, F., Mdaghri Alaoui, M., & Vorsatz, M. (2022). Academic integrity in on-line exams: Evidence from a randomized field experiment. Journal of Economic Psychology, 93, Article 102555. https://doi.org/10.1016/j.joep.2022.102555
- Lancaster, T., & Cotarlan, C. (2021). Contract cheating by STEM students through a file sharing website: A Covid-19 pandemic perspective. International Journal for Educational Integrity, 17, Article 3. https://doi.org/10.1007/s40979-021-00070-0
- Newton, P. M. (2023). The validity of unproctored online exams is undermined by cheating. Proceedings of the National Academy of Sciences, 120(41), Article e2312978120. https://doi.org/10.1073/pnas.2312978120
- Pleasants, J., Pleasants, J. M., & Pleasants, B. (2022). Cheating on unproctored online exams: Prevalence, mitigation measures, and effects on exam performance. Online Learning, 26(1). https://doi.org/10.24059/olj.v26i1.2620
- Steger, D., Schroeders, U., & Gnambs, T. (2020). A meta-analysis of test scores in proctored and unproctored ability assessments. European Journal of Psychological Assessment, 36(1), 174–184. https://doi.org/10.1027/1015-5759/a000494