How common is cheating in online exams and did it increase during the COVID-19 pandemic? A Systematic Review

Academic misconduct is a threat to the validity and reliability of online assessment, and media reports suggest that misconduct spiked dramatically in higher education during the emergency shift to online exams caused by the COVID-19 pandemic. This study reviewed survey research to determine how common it is for university students to admit cheating in online exams, and how and why they do it. We also assessed whether these self-reports of cheating increased during the COVID-19 pandemic, along with an evaluation of the quality of the research evidence which addressed these questions. 25 samples were identi�ed from 19 Studies, including 4672 participants, going back to 2012. Online exam cheating was self-reported by a substantial minority (44.7%) of students in total. Pre-COVID this was 29.9%, but during COVID cheating jumped to 54.7%, although these samples were more heterogenous. Individual cheating was more common than group cheating, and the most common reason students reported for cheating was simply that there was an opportunity to do so. Remote proctoring appeared to reduce the occurrence of cheating, although data were limited. However there were a number of methodological features which reduce con�dence in the accuracy of all these �ndings. Most samples were collected using designs which makes it likely that online exam cheating is under-reported, for example using convenience sampling, a modest sample size and insu�cient information to calculate response rate. No studies considered whether samples were representative of their population. Future approaches to online assessment should consider how the basic validity of examinations can be maintained, considering the substantial numbers of students who appear to be willing to admit engaging in misconduct. Future research on academic misconduct would bene�t from using large representative samples, guaranteeing participants anonymity.


Introduction
Distance learning, also referred to as e-learning, blended learning or mobile learning (Zarzycka et al., 2021) is de ned as learning with the use of technology where there is a physical separation of students from the teachers during the active learning process, instruction and examination (Armstrong-Mensah et al., 2020).Assessment poses many challenges for distance learning.Summative assessments, including exams, are the basis for making decisions about the grading and progress of individual students, while aggregated results can inform educational policy such as curriculum or funding decisions (Shute & Kim, 2014).Thus, it is essential that online summative assessments can be conducted in a way that allows for their basic reliability and validity to be maintained.
Cheating (the gaining of an unfair advantage) poses an obvious threat to the validity of online examinations.Noorbebahani and colleagues recently reviewed online exam cheating in higher education.They found that students use a variety of methods to gain an unfair advantage, including accessing unauthorized materials such as notes and textbooks, using an additional device to go online, collaborating with others, and even outsourcing the exam to be taken by someone else.These can be driven by a variety of motivations, including a fear of failure, peer pressure, a perception that others are cheating, and the ease with which they can do it (Noorbehbahani et al., 2022).However, it remains unclear how many students are actually engaged in these cheating behaviours.Understanding the scale of cheating is an important pragmatic consideration when determining how, or even if, it could/should be addressed.There is an extensive literature on the incidence of other types of misconduct, but cheating in online exams has received less attention than other forms of misconduct such as plagiarism (Garg & Goel, 2022).
In response to the global COVID-19 pandemic, many institutions relied on distance learning to reduce in-person interaction and so to reduce the spread of the disease (Pokhrel & Chhetri, 2021).Universities shifted, very rapidly, to online assessment methods, with limited time to ensure that these methods were secure.There were subsequent media reports that academic misconduct was now 'endemic', with universities supposedly 'turning a blind eye' towards cheating (e.g.Henry, 2022;Knox, 2021).However, it is unclear whether this media anxiety is re ected in the real-world experience in universities.
Pandemic-induced 'lockdown learning' continued, in some form, for almost 2 years in many countries, causing a rapid, substantial and prolonged change in educational provision.We all, staff and students, had to learn a lot, very quickly, about distance learning.This triggered predictions that higher education would be permanently changed by the pandemic, with online/distance learning becoming much more common, even the norm (Barber et al., 2021;Dumulescu & Muţiu, 2021).One obvious potential change would be the widespread adoption of online assessment methods.Online exams offer students increased exibility, for example the opportunity to sit an exam in their own homes.This may also reduce some of the anxiety experienced during attending in-person exams in an exam hall, and potentially reduces the administrative cost to universities.However, students, like the media and university staff, reported concerns about the security of online assessments during the pandemic (Brown et al., 2022).One obvious response to these concerns is to use remote proctoring systems wherein students are monitored through webcams and use lockeddown browsers.However, the use of these systems has been controversial, with students feeling that they are 'under surveillance' (Lee & Fanguy, 2022).A recent court ruling in the USA found that the use of remote proctoring was unconstitutional (Bowman, 2022).This ruling seems unlikely to be the last; there is already a long history of legal battles between the proctoring companies and their critics (Corbyn, 2022), and it is still unclear whether these systems actually reduce misconduct.Alternatives have been offered in the literature, including guidance for how to prepare online exams in a way that reduces the opportunity for misconduct (Whisenhunt et al., 2022), although it is unclear whether this guidance is effective.
There is a large body of research literature which examines the prevalence of different types of academic dishonesty and misconduct.Much of this research is in the form of survey-based self-report studies.There are some obvious problems with using self-report as a measure of misconduct; it is a 'deviant' or 'undesirable' behaviour, and so those invited to participate in survey-based research have a disincentive to respond truthfully, if at all, especially if there is no guarantee of anonymity.There is also some evidence that certain demographic characteristics associated with an increased likelihood of engaging in academic misconduct are also predictive of a decreased likelihood of responding voluntarily to surveys, meaning that misconduct is likely under-reported when a non-representative sampling method is used (Newton, 2018).Some of these issues with quantifying academic misconduct can be partially addressed by the use of rigorous research methodology, for example using representative samples with a high response rate, and clear, unambiguous survey items (Bennett et al., 2011;Halbesleben & Whitman, 2013).Guarantees of anonymity are also essential for respondents to feel con dent about answering honestly, especially when the research is being undertaken by the very universities where participants are studying.A previous systematic review of academic misconduct found that self-report studies are often undertaken with small, convenience samples with low response rates (Newton, 2018).Similar ndings were reported when reviewing the reliability of research into the prevalence of belief in the Learning Styles neuromyth, suggesting that this is a wider concern within survey-based education research (Newton & Salvi, 2020).
However, self-report remains one of the most common ways that academic misconduct is estimated, perhaps in part because there are few other ways to meaningfully measure it.There is also a basic, intuitive objective validity to the method; asking students whether they have cheated is a simple and direct approach, when compared to other indirect approaches to quantifying misconduct, based on (for example) learner analytics, originality scores or grade discrepancies.There is some evidence that self-report correlates positively with actual behaviour (Gardner et al., 1988), and that data accuracy can be improved by using methods which incentivize truth-telling (Curtis et al., 2022).
Here we undertook a systematic search of the literature in order to identify research which studied the prevalence of academic dishonesty in summative online examinations in Higher Education.The research questions were thus 1.How common is self-report of cheating in online exams in Higher Education?(This was the primary research question, and studies were only included if they addressed this question).
2. Did cheating in online exams increase during the COVID-19 pandemic?
3. What are the most common forms of cheating?
4. What are student motivations for cheating?5. Does online proctoring reduce the incidence of self-reported online exam cheating?

Methods
Identifying samples.We used a methodology based on previous work in systematically reviewing survey-based research in education, misbelief and misconduct (Fanelli, 2009;Newton, 2018;Newton & Salvi, 2020).Searches were conducted in July and August 2022.Searches were rst undertaken using the ERIC education research database (eric.ed.gov) and then with Google Scholar.We used Google Scholar since it covers grey literature (Haddaway et al., 2015), including unpublished Masters and PhD theses (Jamali & Nabavi, 2015).The Google Scholar search interface is limited, and the search returns can include non-research documents search as citations, university policies and handbooks on academic integrity, and multiple versions of papers (Boeker et al., 2013).It is also not possible to exclude the results of one search from another.Thus it is not possible for us to report accurately the numbers of included papers returned from each term.
Search results were individually assessed against the inclusion/exclusion criteria, starting with the title, followed by the abstract and then the full text.If a study clearly did not meet the inclusion criteria based on the title then it was excluded.If the author was unsure, then the abstract was reviewed.If there was still uncertainty, then the full text was reviewed.
When a study met the inclusion criteria (see below), the speci c question used in that study to quantify online exam cheating was then itself also used as a search term.Thus the full list of search terms used was: "Cheating" AND "exam" AND "online", "pandemic" AND "cheating", "cheating in online exams", "cheating on online exams", "cheat on online exams", "cheat in online exams", academic integrity AND "online exams", academic integrity AND "online" AND "exams", "Academic integrity" AND "online exams" AND "incidence", "Academic integrity" AND "online " AND "exams" AND "incidence", "online exam" AND "cheating", "online exam" AND "academic misconduct", "online exam" AND "academic integrity", "online exam" AND "cheating" AND "COVID" AND "incidence", "online exam" AND "cheating" AND "COVID" AND "incidence", "frequency of cheating", "frequency" AND "cheating", "percentage" AND "cheating", "common" AND "cheating", "McCabe's Survey", "online" AND "cheating", "prevalence" AND "cheating" AND "exam", "online exams" AND "academic dishonesty" AND "survey", "academic dishonesty in online exams", "academic dishonesty in online examinations", "academic dishonesty in online learning", "ever cheated" AND "online exams", "digital exams", "digital examinations", "cheating in digital exams", "cheating in digital examinations", "cheat on digital exams", "cheat on digital examinations", "cheat in digital exams", "cheat in digital examinations", "cheat" AND "digital exams", "cheating in online classes" AND "survey", "Did/Do you as an online student open the coursebook during an online exam?", "Did/Do you as an online student have another person do your exam?", "Did/Do you as an online student obtain the exam questions from another student who did the test before you?", "Did/Do you as an online student help other students during the exam?", "Did/Do you as an online student send the answers to other students?","How often do you have someone else give you answers the answers during an online exam?", "Screen capture for online exam", "Rotate online exam taking", "Did/Do you as an online student open the coursebook during an online exam?", "Did/Do you as an online student have another person do your exam?", "Did/Do you as an online student obtain the exam questions from another student who did the test before you?", "Did/Do you as an online student help other students during the exam?", "Did/Do you as an online student send the answers to other students?","How often do you have someone else give you answers the answers during an online exam?", "Screen capture for online exam".
'Daisy chaining' was also used to identify relevant research from studies that had already been identi ed using the aforementioned literature searches, and recent reviews on the subject (Butler-Henderson & Crawford, 2020; Inclusion and Exclusion criteria.The following criteria were used to determine whether to include samples.Many studies included multiple datasets (e.g.samples comprising different groups of students, across different years).The criteria here were applied to individual datasets.

Inclusion criteria
Participants were asked whether they had ever cheated in an online exam (self-report).
Participants were students in Higher Education.
Reported both total sample size and percent of respondents answering yes to the relevant exam cheating questions, or su cient data to allow those metrics to be calculated.

English language publication
Published 2013-present, with data collected 2012-present.We wanted to evaluate a 10 year timeframe.In 2013, at the beginning of this time window, the average time needed to publish an academic paper was 12.2 months, ranging from 9 months (chemistry) to 18 months (Business) (Björk & Solomon, 2013).It would therefore be reasonable to conclude that a paper published in 2013 was most likely submitted in 2012.Thus we included papers whose publication date was 2013 onwards, unless the manuscript itself speci cally stated that the data were collected prior to 2012.

Exclusion criteria
Asking participants would they cheat in exams (e.g.(Morales-Martinez et al., 2019), or did not allow for a distinction between self-report of intent and actual cheating (e.g.(Ghias et al., 2014)) Phrasing of survey items in a way that does not allow for frequency of online exam cheating to be speci cally identi ed according to the criteria above.Wherever necessary, study authors were contacted to clarify.
Asking participants 'how often do others cheat in online exams'.
Asking participants about helping other students to cheat.
Cheating in formative assessments, or did not distinguish between formative/summative (e.g.quizzes/exams (e.g.(Alvarez, Homer, T. et al., 2022; Costley, 2019)) Estimates of cheating from learning analytics or other methods which did not include directly asking participants if they had cheated.
Published in a predatory journal (see below).
Predatory journal criteria.Predatory journals and publishers are de ned as "entities which prioritize self-interest at the expense of scholarship and are characterised by false or misleading information, deviation from best editorial and publication practices, a lack of transparency, and/or the use of aggressive and indiscriminate solicitation practices."(Grudniewicz et al., 2019).The inclusion of predatory journals in literature reviews may therefore have a negative impact on the data, ndings and conclusions.We followed established guidelines for the identi cation and exclusion of predatory journals from the ndings (Rice et al., 2021): Each study which met the inclusion criteria was checked for spelling, punctuation and grammar errors as well as logical inconsistencies.

Every included journal was checked against open access criteria;
o If the journal was listed on the Directory of Open Access Journals (DOAJ) database (DOAJ.org)then it was considered to be non-predatory.
o If the journal was not present in the DOAJ database, we looked for it in the Committee on Publication Ethics (COPE) database (publicationethics.org).
If the journal was listed on the COPE database then it was considered to be non-predatory.
Only one paper met these criteria, containing logical inconsistencies and not listed on either DOAJ or COPE.For completeness we also searched an informal list of predatory journals (https://beallslist.net) and the journal was listed there.Thus the study was excluded.
Metrics and Analysis.All data were extracted by both authors independently.Where the extracted data differed between authors then this was clari ed through discussion.Data extracted were, where possible, as follows: Author/date

Year of publication
Year study was undertaken.If this was a range (e.g.Nov 2016-Apr 2017) then the most recent year was used as the data point (e.g.2017 in the example).If it was not reported when the study was undertaken, then we recorded the year that the manuscript was submitted.If none of these data were available then the publication year was entered as the year that the study was undertaken.
Population size.The total number of participants in the population, from which the sample is drawn and supposed to represent.For example, if the study is surveying 'business students at University X', is it clear how many business students are currently at University X?
Number Sampled.The number of potential participants, from the population, who were asked to ll in the survey N. The number of survey respondents.
Cheated in online summative examinations.The number of participants who answered 'yes' to having cheated in online exams.Some studies recorded the frequency of cheating on a scale, for example a 1-5 Likert scale from 'always' to 'never'.In these cases, we collapsed all positive reports into a single number of participants who had ever cheated in online exams.Some studies did not ask for a total rate of cheating (i.e.cheating by any/all methods) and so, for analysis purposes the method with the highest rate of cheating was used (see Results).
Group/individual cheating.Where appropriate, the frequency of cheating via different methods was recorded.These were coded according to the highest level of the framework proposed by Noorbehbahani (Noorbehbahani et al., 2022), i.e. group vs individual.More granular analysis was not possible due to the number and nature of the included studies.
Response rate.De ned as " the percentage of people who completed the survey after being asked to do so" (Halbesleben & Whitman, 2013).
Method of sampling.As one of the following; convenience sampling, where all members of the population were able to complete the survey, but data were analysed from those who voluntarily completed it.'Unclassi able' where it was not possible to determine the sampling method based on the data provided (no other sampling methods were used in the included studies).
Ethics.Was it reported whether ethical/IRB approval had been obtained.
Anonymity.Were participants assured that they were answering anonymously?Students who are found to have cheated in exams can be given severe penalties, and so a statement of anonymity (not just con dentiality) is important for obtaining meaningful data.
Data are reported as mean ± SEM unless otherwise stated.Datasets were tested for normal distribution using a Kolmogorov-Smirnov test prior to analysis and parametric tests were used if the data were found to be normally distributed.The details of the speci c tests used are in the relevant results section.
Results  (Owens, 2015)).Thus, these samples were treated as distinct in the analysis since they represent different participants.Multiple studies asked the same groups of participants about different types of cheating, or the conditions under which cheating happens.The analysis of these is explained in the relevant results subsection.A summary of the studies is in Table 1.The detail of each individual question asked is in supplementary online data S1.Response rate.Fifteen of the samples did not report su cient information to allow a response rate to be calculated.The ten remaining samples returned an average response rate of 55.6% ±10.7, with a range from 12.2-100%.
Anonymity.Eighteen of the 23 samples (72%) stated that participant responses were collected anonymously Ethics.Seven of the 25 samples (28%) stated that ethical approval was obtained for the study.
How common is self-reported online exam cheating in Higher Education?44.7% of participants (2088/4672) reported engaging in some form of cheating in online exams.This analysis included those studies where total cheating was not recorded, and so the most commonly reported form of cheating was substituted in.To check the validity of this inclusion, a separate analysis was conducted of only those studies where total cheating was recorded.In this case, 42.5% of students (1574/3707) reported engaging in some form of cheating.An unpaired t-test was used to compare the percentage cheating from each group (total vs highest frequency), and returned no signi cant difference (t(23) = 0.5926, P = 0.56).
Did the frequency of online exam cheating increase during COVID?The samples were classi ed as having been collected pre-COVID, or during COVID (no samples were identi ed as having been collected 'post-COVID').One study (Jenkins et al., 2022) asked the same students about their behaviour before, and during, COVID.For the purposes of this speci c analysis, these were included as separate samples, thus there were 26 samples, 17 pre-COVID and 9 during COVID.Pre-COVID, 29.9% (629/2107) of participants reported cheating in online exams.Post-COVID this gure was 54.7% (1519/2779).
To estimate the variance in these data, and to test whether the difference was statistically signi cant, the percentages of students who reported cheating for each study were grouped into pre-and post-COVID and the average calculated for each group.The average pre-COVID was 28.03% ± 4.89, (N = 17), whereas during COVID the average is 65.06 ± 9.585 (N = 9).An unpaired t-test was used to compare the groups, and returned a statistically signi cant difference (t(24) = 3.897, P = 0.0007).The effect size (Hedges g) was 1.61, indicating that the COVID effect was substantial (Fig. 1).
To test the reliability of this result, we conducted a split sample test as in other systematic reviews of the prevalence of academic misconduct (Newton, 2018), wherein the data for each group were ordered by size and then every other sample was extracted into a separate group.So, the sample with the lowest frequency of cheating was allocated into Group A, the next smallest into Group B, the next into Group A, and so on.This was conducted separately for the pre-COVID and 'during COVID'.Each half-group was then subject to an unpaired t-test to determine whether cheating increased during COVID in that group.Each group returned a signi cant difference (t(10) = 2.889 P = 0.0161 for odd-numbered samples, t(12) = 2.48, P = 0.029 for evennumbered samples.This analysis gives con dence that the observed increase in self-reported online exam cheating during the pandemic is statistically robust, although there may be other variables which contribute to this (see discussion).
Comparison of Group vs Individual online exam cheating in Higher Education.In order to consider how best to address cheating in online exams, it is important to understand the speci c behaviours of students.Many studies asked multiple questions about different types of cheating, and these were coded according to the typology developed by Noorbehbehani which has a high-level code of 'individual' and 'group' (Noorbehbahani et al., 2022).
'Individual' cheating meant that, whatever the type of cheating, it could be achieved without the direct help of another person.This could be looking at notes or textbooks, or searching for materials online.'Group' cheating meant that another person was directly involved, for example by sharing answers, or having them sit the exam on behalf of the participant (contract cheating).Seven studies asked their participants whether they had engaged in different forms of cheating where both formats (Group and Individual) were represented.For each study we ranked all the different forms of cheating by the frequency with which participants reported engaging in it.For all seven of the studies which asked about both Group and Individual cheating, the most frequently reported cheating behaviour was an Individual cheating behaviour.For each study we calculated the difference between the two by subtracting the frequency of the most commonly reported Group cheating behaviour from the frequency of the most commonly reported Individual cheating behaviour.The average difference was 23.32 ± 8.0 percentage points.These two analyses indicate that individual forms of cheating are more common than cheating which involves other people.
Effect of Proctoring/Lockdown Browsers.The majority of studies did not make clear whether their online exams were proctored or unproctored, or whether they involved the use of other software such as lockdown browsers.Thus it was di cult to conduct de nitive analyses to address the question of whether these systems reduce online exam cheating.Two studies did speci cally address this issue in both cases there was a substantially lower rate of self-reported cheating where proctoring systems were used.Jenkins et al, in a study conducted during COVID, asked participants whether their instructors used 'anti cheating software (e.g., Lockdown Browser)' and, if so, whether they had tried to circumvent it.16.5% admitted to doing this, compared to the overall rate of cheating of 58.4%.Owens asked about an extensive range of different forms of misconduct, in two groups of students whose online exams were either proctored or unproctored.The total rates of cheating in each group did not appear to be reported.The most common form of cheating was the same in both groups ('web search during an exam') and was reported by 39.8% of students in the unproctored group but by only 8.5% in the proctored group (Owens, 2015).
Reasons given for online exam cheating.Ten of the studies asked students why they cheated in online exams.These reasons were initially coded by both authors according to the typology provided in (Noorbehbahani et al., 2022).Following discussion between the authors, the typology was revised slightly to that shown in Table 1, to better re ect the reasons given in the reviewed studies.2. Direct comparison between the reasons is not fully valid since different studies asked for different options, and some studies offered multiple options whereas some only identi ed one.However in the four studies that offered multiple options to students, three of them ranked 'opportunities to cheat' as the most common reason (and the fourth study did not have this as an option).Thus students appear to be most likely to cheat in online exams when there is an opportunity to do so.
Table 3 Different reasons for cheating.The speci c statement from the study was coded by the authors here to facilitate comparison between studies.Many studies asked these questions hypothetically (e.g.what would motivate you to cheat) and so the % of students selecting an option may be higher than the percentage who report actually cheating.*qualitative 'path nder analysis' was undertaken to identify motivations for cheating.These were not quanti ed.

Discussion
We reviewed data from 19 studies, including 25 samples totaling 4672 participants.We found that a substantial proportion of students, 44.7%, were willing to admit to cheating in online summative exams.This total number masks a nding that cheating in online exams appeared to increase considerably during the COVID-19 pandemic, from 29.9-54.7%.These are concerning ndings.However, there are a number of methodological considerations which in uence the interpretation of these data.These considerations all lead to uncertainty regarding the accuracy of the ndings, althoughr a common theme is that, unfortunately, the issues highlighted seem likely to result in an under-reporting of the rate of cheating in online exams.
There are numerous potential sources of error in survey-based research, and these may be ampli ed where the research is asking participants to report on sensitive or undesirable behaviours.One of these sources of error comes from non-respondents, i.e. how con dent can we be that those who did not respond to the survey would have given a similar pattern of responses to those that did (Goyder et al., 2002;Halbesleben & Whitman, 2013;Sax et al., 2003).Two ways to minimize non-respondent error are to increase the sample size as a percentage of the population, and then simply to maximise the percentage of the invited sample who responds to the survey.However only nine of the samples reported su cient information to even allow the calculation of a response rate, and only two reported the total population size.Thus for the majority of samples reported here, we cannot even begin to estimate the extent of the non-response error.For those that did report su cient information, the response rate varied considerably, from 12.2% to 100, with an average of 55.6%.Thus a substantial number of the possible participants did not respond.
Most of the surveys reviewed here were conducted using convenience sampling, i.e. participation was voluntary and there was no attempt to ensure that the sample was representative, or that the non-respondents were followed up in a targeted way to increase the representativeness of the sample.People who voluntarily respond to survey research are, compared to the general population, older, wealthier, more likely to be female and educated (Curtin et al., 2000).In contrast, individuals who engage in academic misconduct are more likely to be male, younger, from a lower socioeconomic background and less academically able (reviewed in (Newton, 2018).Thus the features of the survey research here would suggest that the rates of online exam cheating are under-reported.
A second source of error is measurement error -for example, how likely is it that those participants who do respond are telling the truth?Cheating in online exams is clearly a sensitive subject for potential survey participants.Students who are caught cheating in exams can face severe penalties.
Measurement error can be substantial when asking participants about sensitive topics, particularly when they have no incentive to respond truthfully.Curtis et al conducted an elegant study to investigate rates of different types of contract cheating and found that rates were substantially higher when participants were incentivized to tell the truth, compared to traditional self-report (Curtis et al., 2022).No studies reviewed here reported any incentivization, and many did not report IRB approval or that participants were guaranteed anonymity in their responses.Absence of evidence is not evidence of absence, but it again seems reasonable to conclude that the majority of the measurement error reported here will also lead to an underreporting of the extent of online exam cheating.
However, there are very many variables associated with likelihood of committing academic misconduct (also reviewed in (Newton, 2018).For example, in addition to the aforementioned variables, cheating is also associated with individual differences such as personality traits (Giluk & Postlethwaite, 2015; Williams & Williams, 2012), motivation (Park et al., 2013), age and gender (Newstead et al., 1996) and studying in a second language (Bretag et al., 2019) as well as situational variables such as discipline studied (Newstead et al., 1996).None of the studies reviewed here can account for these individual variables, and this perhaps explains, partly, the wide variance in the studies reported here, where the percentage of students willing to admit to cheating in online exams ranges from essentially none, to all students, in different studies.However, almost all of the variables associated with differences in likelihood of committing academic misconduct were themselves determined using convenience sampling.In order to begin to understand the true nature, scale and scope of academic misconduct, there is a clear need for studies using large, representative samples, with appropriate methodology to account for non-respondents.
There are some speci c issues which must be considered when determining the accuracy of the data showing an increase in cheating during COVID.In general, the pre-COVID group appears to be a more homogenous set of samples, for example, 11 of the 16 samples are from students studying business, and 15 of the 16 pre-COVID samples are from the USA.The post-COVID samples are from a much more diverse range of disciplines and countries.
However the increase in self-reported cheating was replicated in the one study which directly asked students about their behaviour before, and during, the pandemic; Jenkins and co-workers found that 28.4% of respondents were cheating pre-COVID, nearly doubling to 58.4% during the pandemic (Jenkins et al., 2022), very closely mirroring the aggregate results.
It is di cult to quantify the potential impact of these issues on the accuracy of the data analysed here, since objective measures of cheating in online exams are di cult to obtain in higher education settings.One way to achieve this is to set up traps for students taking closed-book exams.One study tested this in the context of a 2.5 hour online exam administered for participants to obtain credit from a MOOC.The exams was set up in such a way that participants would "likely not bene t from having access to third-party reference materials during the exam".Students were instructed not to access any additional materials or to communicate with others during the exam.The authors built a 'honeypot' website which had all of the exam questions on, with a button 'click to show answer'.If exam participants went online and clicked that button then the site collected information which allowed the researchers to identify the unique i.d. of the test-taker.This approach was combined with a more traditional analysis of the originality of the free-text portions of the exam.Using these methods, the researchers estimated that ~ 30% of students were cheating (Corrigan-Gibbs et al., 2015b).This study was conducted in 2014-15, and the data align well with the pre-COVID estimates of cheating found here.
The challenges of interpreting data from small convenience samples will also affect the analysis of the other measures made here; that students are more likely to commit misconduct on their own, because they can.The overall pattern of ndings though does align somewhat, suggesting that concerns may be with the accuracy of the numbers rather than a fundamental qualitative problem (i.e. it seems reasonable to conclude that students are more likely to cheat individually, but it is challenging to put a precise number to that nding).For example, the apparent increase in cheating during COVID is associated with a rapid and near-total transition to online assessment.Pre-covid, the use of online exams would have been a choice made by education providers, presumably with some efforts to secure the security and integrity of the assessment.During COVID lockdown, the scale and speed of the transition to online exams made it much more challenging to put security measures in place, and this would therefore almost certainly have increased the opportunities to cheat.
Thus an aggregate portrayal of the ndings here is that students are committing misconduct in signi cant numbers, and that this has increased considerably during COVID.Students appear to be more likely to cheat on their own, rather than in groups, and most commonly motivated by the simple fact that they can cheat.What can we do about it?
One obvious suggestion is to increase the use of remote proctoring, wherein student behaviour during online exams is monitored, for example, through a webcam, and/or their online activity is monitored or restricted.We were unable to draw meaningful conclusions about the effectiveness of remote proctoring or other software such as lockdown browsers to reduce cheating in online exams, since very few studies stated de nitively that the exams were, or were not, proctored.The two studies that examined this question did appear to show a substantial reduction in the frequency of cheating when proctoring was used.Con dence in these results is bolstered by the fact that these studies both directly compared unproctored vs proctored/lockdown browser.Other studies have used proxy measures for cheating, such as time engaged with the exam, and changes in exams scores, and these studies have also found evidence for a reduction in misconduct when proctoring is used (e.g.(Dendir & Maxwell, 2020)).The effectiveness (or not) of remote proctoring to reduce academic misconduct seems like an important area for future research.However there is considerable controversy about the use of remote proctoring, including legal challenges to its use and considerable objections from students as reviewed in the introduction, and so it remains an open question whether this is a viable option for widespread general use.
Honour codes are a commonly cited approach to promoting academic integrity, and so (in theory) reducing academic misconduct.However, empirical tests of honour codes show that they do not appear to be effective at reducing cheating in online exams (Corrigan-Gibbs et al., 2015a, b).In these studies the authors likened them to 'terms and conditions' for online sites, which are largely disregarded by users in online environments.However in those same studies the authors found that replacing an honour code with a more sternly worded 'warning', which speci es the consequences of being caught, was effective at reducing cheating.Thus a warning may be a simple, low-cost intervention to reduce cheating in online exams.
Another potential approach to deterring cheating could be to deliberately set traps, of 'honeypots'; websites which appear to show students the answers to an exam but which are actually a way of catching students who commit academic misconduct.This approach has been used to quantify cheating in online exams, as described above (Corrigan-Gibbs et al., 2015b).There are some obvious ethical issues associated with such an approach, and from a practical perspective it seems reasonable to assume that students would get wise to this, and that the identity and nature of any honeypot could be easily shared by students cheating collaboratively.
Another option to reduce cheating in online exams is to use open-book exams.This is often suggested as a way of simultaneously increasing the cognitive level of the exam (i.e. it assesses higher order learning) (e.g.(Varble, 2014), and was suggested as a way of reducing the perceived, or potential increase in academic misconduct during COVID (e.g.(Nguyen et al., 2020;Whisenhunt et al., 2022).This approach has an obvious appeal in that it eliminates the possibility of some common forms of misconduct, such as the use of notes or unauthorized web access (Noorbehbahani et al., 2022;Whisenhunt et al., 2022), and can even make this a positive feature, i.e. encouraging the use of additional resources in a way that re ects the fact that, for many future careers, students will have access to unlimited information at their ngertips, and the challenge is to ensure that students have learned what information they need and how to use it.This approach certainly ts with our data, wherein the most frequently reported types of misconduct involved students acting alone, and cheating 'because they could'.Some form of proctoring or other measure may still be needed in order to reduce the threat of collaborative misconduct.Perhaps most importantly though, it is unclear whether open-book exams truly reduce the opportunity for, and the incidence of, academic misconduct, and if so, how might we advise educators to design their exams, and exam question, in a way that delivers this as well as the promise of 'higher order' learning.These questions are the subject of ongoing research.
In summary then, there appears to be signi cant levels of misconduct in online examinations in Higher Education.Students appear to be more likely to cheat on their own, motivated by assessment design and delivery which makes it easy for them to do so.Future research in academic integrity would bene t from large, representative samples using clear and unambiguous survey questions and guarantees of anonymity.This will allow us to get a much better picture of the size and nature of the problem, and so design strategies to mitigate the threat that cheating poses to assessment validity.

Supplementary Files
This is a list of supplementary les associated with this preprint.Click to download.

Figures
Figures

Table 1
Studies and samples included in the review.Data = year the data was collected.RR = response rate.P/L = was remote proctoring or a lockdown browser used (ns = not stated).See text for further details.Sampling method.23/25 samples were collected using convenience sampling.The remaining two did not provide su cient information to determine the method of sampling.

Table 2
Reasons given for cheating in online exams, coded by the authors based on the examples given in the studies reviewed here.Descriptive statistics (the percentages of students reporting the different reasons as motivations for cheating) are shown in Table