Participants using GenAI in online studies: an overview and recommendations for researchers and reviewers Authors: Justus Mann, Jutta Stumpf-Wollersheim Published: 2026 ABSTRACT BackgroundParticipants’ use of generative artificial intelligence (GenAI) to create responses in online studies can threaten data authenticity and research validity. This issue is particularly pressing in settings where participation is paid for. In response, researchers have begun to develop different approaches to prevent and detect GenAI use by participants.ObjectiveThis comment illustrates the unintended consequences of researchers’ measures to prevent and detect participants’ GenAI use and aims to help researchers and reviewers better identify the potential consequences of these measures.ResultsDrawing on current literature, we identify seven different levels of participants’ GenAI use, differentiated by increasing degrees of technological sophistication and automation. These levels range on a continuum from the manual operation of GenAI by participants to the employment of automated GenAI agents. We discuss prevention and detection measures corresponding to these levels and how researchers’ measures can cause reactions from participants. Specifically, we suggest that extensive measures can trigger an adverse selection mechanism by driving out honest participants and inviting greater sophistication in GenAI use.ConclusionsInstead of engaging in a technological cat-and-mouse game with participants through deterrence, we argue that we should focus on the incentive structure to resolve the issue. We propose a combination of better payment and an opt-in verified participation mode, i.e., video proof, to reduce participants’ GenAI use in online studies and increase data quality. Building on this perspective, we provide practical recommendations for researchers and reviewers to improve research quality, maintain fairness, and save resources. Background In the last couple of years, we observed an increase in suspiciously long, verbose, and mistake-free responses to the open-ended questions in our online studies. Upon closer inspection, we identified highly similar or even redundant answers. To make matters worse, we observed these characteristics not only in responses from crowdsourcing platforms such as MTurk and Prolific but even in responses we obtained from student samples. Our findings led us to believe these participants use generative artificial intelligence (GenAI) chatbots such as ChatGPT to produce their responses. We therefore became increasingly concerned about collecting seemingly coherent, yet inauthentic, data. The recent advances in GenAI have led to the proliferation of not only large language models (LLMs), but also machine vision, audio, and video generation, which provide a unique ensemble for the multi-modal imitation of human participants. In comparison to previous computer technology, GenAI is capable of convincingly imitating human thought and speech, it is easy to use, and it is highly adaptable to different settings. While inattentiveness, carelessness, click bots, and fraudulent participants in online research have been persistent sources of concern for data quality in online research [1–3], research on participants’ GenAI use in online research is rather new. Our concerns prompted us to conceptually and empirically study the accuracy of behavioral indicators – both regarding the detection of participants’ manual GenAI use in open-ended responses and regarding how to address the impact of GenAI use on the research data [4, 5]. However, GenAI is constantly evolving, and it becomes easier to use, more autonomous, and more capable of imitating human participants. Therefore, researchers’ prevention and detection measures require constant adaptation. Researchers have increasingly investigated the issue of participants’ GenAI use by measuring its frequency, quantifying its effect on research results, and testing different approaches for its prevention and detection [4, 6–10]. In these investigations, GenAI use is often conceptualized as either participants manually operating GenAI (human-GenAI hybrid literature stream) or participants employing automatic GenAI agents (bot literature stream). However, both literature streams describe the same underlying phenomenon, namely, GenAI responses polluting data collection from human participants [11]. In addition, there are different forms and levels of GenAI use, which require different prevention and detection approaches [12]. For instance, participants can manually paste LLM answers into open-ended text boxes [13], participants can create custom scripts for answering closed questions [8, 14], and participants can even delegate their participation in entire studies to GenAI agents [15, 16]. In order to comprehensively address participants’ GenAI use in online studies, it is therefore beneficial to conceptualize participants’ GenAI use on a continuum from manual GenAI use to automatic GenAI agents [11, 12]. Building on prior research, we identify seven levels of GenAI use with increasing degrees of technological sophistication and automation. One important characteristic of increasing technological sophistication is the rising proficiency in simulating genuine human participant behavior [17]. Especially in light of the capabilities and autonomy of AI agents such as OpenClaw, Manus, or ChatGPT agent, it is increasingly important to protect research data against GenAI influence. However, as researchers adopt more sophisticated prevention and detection measures in response to GenAI threats, they also change the conditions for online research. The goal of our comment is to illustrate the unintended consequences of researchers’ measures to prevent and detect participants’ GenAI use and to help researchers and reviewers better identify the potential consequences of these measures. First, we integrate different perspectives on participants’ GenAI use as recent research has begun to discuss different types and levels of participants’ GenAI use, measures from researchers for maintaining data authenticity, and forms of participants’ adaptation. We contribute to this discussion by demonstrating how various levels of participants’ engagement with GenAI can be conceptualized along a spectrum – ranging from the manual operation of GenAI to the deployment of automated GenAI agents. We contribute to prior research by detailing the escalating dynamics of action and reaction – between the prevention and detection measures implemented by researchers and participants’ adaptation of their GenAI usage, at every level. Second, we argue that these dynamics of action and reaction are not merely a technical issue, but also an infrastructural issue of paid participation. Contrary to their goals, detection and prevention measures can trigger an adverse selection mechanism by driving out honest participants and inviting the sophistication of GenAI use. We discuss the changes needed to break the current dynamic of action and reaction – specifically, through a combination of better compensation and an opt-in “verified participation mode”. Third, we provide recommendations for researchers and reviewers to improve research quality, maintain fairness, and save resources. Our recommendations for researchers offer a perspective for considering potential consequences when planning and conducting online research – complementing the wealth of technical recommendations and best practices. Our recommendations for reviewers aim at supporting the evaluation of research since there is still little guidance for reviewers. Participants’ reasons for using GenAI To appropriately respond to participants’ GenAI use, it is important to understand why participants use GenAI in online studies. In paid participant pools, maximizing monetary gains can be a strong motivation, because GenAI facilitates completing studies. Especially when studies are long, poorly paid, or complex, GenAI may provide participants with a means for satisficing by saving time and cognitive resources while still creating plausible responses [4, 10, 12]. However, GenAI use is not necessarily driven by negative intentions. For example, some participants use GenAI to brainstorm, improve grammar, clarify or reflect on their responses, or better understand study instructions [14]. Nonetheless, even well-intentioned GenAI use can negatively influence research data, for example, by altering the response process or the response content [4, 10]. Another reason for using GenAI may be the pressure to perform. Researchers often employ quality control measures such as attention checks and consistency checks to identify inattentive and careless participants. In paid settings, researchers often reject participants who fail such checks, which directly affects participants’ approval scores and future opportunities to participate in paid studies. By using GenAI, participants may perceive a lower risk of failing such checks, because GenAI is able to detect such checks and does not suffer from fatigue [12, 14]. Furthermore, the likelihood of GenAI use may depend on personal characteristics. For example, there is, albeit contested, evidence that lower age, higher education, and higher digital literacy may increase the likelihood of using GenAI in online studies [4, 10, 13]. Moreover, participants’ prior personal experience with GenAI and their perceived acceptability of GenAI use may help participants judge whether their GenAI use is appropriate [14]. Similarly, intrinsic motivation and personal interest in the study results may influence the propensity to use GenAI [18]. In sum, the diverse reasons for participants’ GenAI use, our limited ability to reliably predict its extent based on personal traits, and the unknown influence on data quality generally require proactive prevention and detection measures from researchers. Prevention and detection measures with regard to the different levels of GenAI use Based on our experiences and reports from other researchers, we identify seven progressive levels of participants’ GenAI use and discuss potential prevention and detection measures. This section also demonstrates the importance of continuously adapting measures for prevention and detection to the sophistication of participants’ GenAI use (see Table 1). Table 1Different levels of participants’ GenAI use and prevention and detection measuresLevels of GenAI usePrevention/detection measuresLevel 1: Copying study instructions to LLMs and pasting LLM responsesAppeal not to use AI, unusually fast text generation speed, self-reports, AI detection softwareLevel 2: Adapting behavior and altering responses to overcome basic detectionBlocking copy and paste commands, converting text instructions to images, mandatory full-screen modeLevel 3: Using browser extensions to overcome preventive measuresKeystroke analysis for user behavior: Logging copy and paste commands, comparing typed contentLevel 4: Simple automation scripts to imitate human participantsTrap and honeypot questions, consistency checks, prompt injection, CAPTCHAsLevel 5: GenAI agents that autonomously complete studiesIP address analysis, keystroke analysis, screen recordingsLevel 6: Combining GenAI agents with metadata manipulationDynamic authenticity checks, webcam recordings, mandatory voice inputsLevel 7: Multi-media simulation of human participantsLaboratory environments, participation limited to terminal devices Table 1Different levels of participants’ GenAI use and prevention and detection measuresLevels of GenAI usePrevention/detection measuresLevel 1: Copying study instructions to LLMs and pasting LLM responsesAppeal not to use AI, unusually fast text generation speed, self-reports, AI detection softwareLevel 2: Adapting behavior and altering responses to overcome basic detectionBlocking copy and paste commands, converting text instructions to images, mandatory full-screen modeLevel 3: Using browser extensions to overcome preventive measuresKeystroke analysis for user behavior: Logging copy and paste commands, comparing typed contentLevel 4: Simple automation scripts to imitate human participantsTrap and honeypot questions, consistency checks, prompt injection, CAPTCHAsLevel 5: GenAI agents that autonomously complete studiesIP address analysis, keystroke analysis, screen recordingsLevel 6: Combining GenAI agents with metadata manipulationDynamic authenticity checks, webcam recordings, mandatory voice inputsLevel 7: Multi-media simulation of human participantsLaboratory environments, participation limited to terminal devices Different levels of participants’ GenAI use and prevention and detection measures Level 1: Copying study instructions to LLMs and pasting LLM responses The most basic and common use of GenAI concerns participants who copy study instructions for open-ended questions into an LLM and paste the LLM’s response into the questions’ text boxes [10]. The basic preventive measure (which does not require programming effort) involves appealing to participants not to use any GenAI technology. While the appeal has been shown to decrease the frequency of GenAI use, some participants overlook or ignore the instruction [4, 13]. Hence, researchers can also apply basic detection measures. These involve detecting participants with unusually fast text generation speed, because LLMs generally surpass peak human typing speeds of around 200 words generated per minute. Moreover, researchers can incorporate self-report questions in their studies, asking participants whether they used any form of AI technology. Researchers can also use post hoc AI detection software to identify synthetic texts [13]. Though these basic detection measures are easy to implement, they can lead to false positives and false negatives. For shorter answers, the time participants spend operating LLMs converges with the time required for typing a manual answer; self-reports can provide wrong results due to inattention or dishonesty, and AI detection software often requires minimum input lengths and can be unreliable [4, 8]. Level 2: Adapting behavior and altering responses to overcome basic detection On level 2, participants begin masking their use of GenAI. For example, participants can wait before submitting their answers to avoid detection of suspiciously fast text generation speeds, or they can manually rephrase LLM output by inserting mistakes or changing words, or employ humanizer software to make responses appear more human to AI detection software [19]. Researchers have therefore proposed the implementation of preventive measures such as blocking copy and paste commands, converting text instructions into images, or requiring participation in full-screen mode to impede the use of LLMs [4, 13]. Level 3: Using browser extensions to overcome preventive measures Preventive measures such as blocking copy and paste commands, enforcing full-screen mode, or restricting page activity provide only surface-level protection against GenAI use. Participants can install browser add-ons that manipulate the logging of copy and paste commands, override restrictions for copy and paste commands, or simulate that the study page is always active. In addition, users can easily prompt LLMs with image instructions, and participants can even use in-page extensions for running LLMs on the study page [4, 11]. One strategy to detect such GenAI use, regardless of these workarounds, lies in the application of keystroke analysis, which allows researchers to investigate whether the submitted responses correspond with the keystrokes participants made on the study page [13]. In cases where large parts of submitted responses do not correspond to the keystrokes made on the study page, this indicates that participants may have used some form of GenAI tool. However, keystroke analysis also bears the risk of false positives in cases where participants use word processing programs, as in these cases there are also differences between the submitted responses and keystrokes on the study page. Level 4: Simple automation scripts to imitate human participants There are ways for participants to circumvent keystroke analysis by using simple automation scripts that imitate human keystroke patterns [8, 14]. Due to GenAI, the development of such scripts does not require advanced programming abilities anymore, because LLMs are able to program such scripts based on text descriptions. Consequently, keystroke analysis can become insufficient for detecting the use of GenAI. In addition, scripts can be enhanced with a click function for answering item-based, closed questions. In order to counter such basic click bots, researchers have proposed instructional manipulation checks, bogus questions, straight-lining detection, and consistency checks across reverse-coded items. Moreover, researchers have proposed the implementation of the same questions at different stages of a study and checking responses for consistency. Researchers can implement honeypot questions (i.e., hidden items only visible to bots) and prompt injection with hidden text-based instructions (very small text or text matching the background color), which ask bots to disclose that they are bots. CAPTCHAs, image tasks, and drag-and-drop puzzles have also been proposed for stopping simple bots [11, 20–22]. As another visual bot detection approach, Rilla et al. [11] and Anders et al. [12] propose using a visual “Pseudo-Ebbinghaus illusion”. In contrast to the classical Ebbinghaus illusion, in the “Pseudo-Ebbinghaus illusion”, the “inner circles have different sizes clearly visible to a human observer” [12, p. 2]. Anders et al. [12] used this approach in their lab studies and observed that advanced AI agents are not able to recognize such differences. As their paper was published half a year ago, we informally tested if more recent LLMs still fail the “Pseudo-Ebbinghaus illusion”. Specifically, we used “older” LLMs from 2025 (GPT-4.1 and Claude Opus 4.5) and more recent LLMs from 2026 (GPT-5.5 and Claude Opus 4.8) and conducted one run per model. In line with Anders et al. [12], we observed that “older” LLMs from 2025 failed. However, in our test, more recent models correctly identified that “the orange circle on the left is larger than the orange circle on the right.” These results demonstrate how quickly detection methods can become outdated with the next GenAI model cycle [11]. Level 5: GenAI agents that autonomously complete studies There is increasing evidence that GenAI agents are already capable of avoiding trap questions and solving CAPTCHAs, while generating plausible metadata such as realistic completion times [17]. In conjunction with their ability to coherently answer both closed and open-ended questions, GenAI agents can complete entire online studies with seemingly authentic human data [15, 16]. Nevertheless, specific metadata such as the IP address, as well as keystroke analyses and screen recordings, may still provide approaches for detecting GenAI agents [16, 23, 24]. However, collecting such individualized and personalized data may sometimes be inhibited due to ethical concerns by institutional review boards or by local legislation [20, 25]. In addition, collecting such data poses technical challenges to researchers, requires time for setup and testing, and may require additional financial resources. Level 6: Combining GenAI agents with metadata manipulation In the future, sophisticated fraud approaches could enhance GenAI agents with custom scripts for manipulating study and participant metadata. Some researchers expect work on systems which could imitate the behavior of human participants (keystrokes and mouse movements) and dynamically manage inconspicuous IP addresses [11, 23, 26]. At this level, researchers will have to implement dynamic authenticity checks (see, for example, Prolific’s authenticity checks [27]) and will have to use detection measures such as webcam recordings [4, 28] or mandatory voice inputs in order to challenge the adaptivity of GenAI agents [11]. Level 7: Multi-media simulation of human participants On level 7, there might be bots capable of generating voice inputs and webcam avatars to imitate human participants, even though we are currently not aware of any systems integrating these functions. While the modular technical elements of these bots already exist – for example, deepfakes and AI avatars – developing, updating, and running such bots would be costly. Compared with the remuneration for participation, such approaches currently seem economically unattractive. Should computational power become consistently cheaper and algorithms better, this balance may change. Measures against this level of sophistication would be to conduct research in laboratory environments or to use specialized terminal devices that prevent the execution of other software. Consequences of expanded prevention and detection measures and participants’ reactions The different levels of GenAI use illustrate a vicious cycle, in which every action triggers a potential reaction. The sophistication of GenAI requires researchers to continually expand their prevention and detection efforts, thereby raising costs for researchers and participants who must pass more checks and tests. Several researchers have called for a joint, systematic effort by researchers, participants, and crowdsourcing platforms, because the efforts required to prevent and detect GenAI use go beyond the know-how, expertise, and financial capacity of individual researchers [4, 11, 12, 18]. However, this dynamic goes beyond a technical challenge; it hints at a fundamental market-design issue, which can trigger adverse selection among participants analogous to the “market for lemons” problem [29], causing quality degradation, lower supply, and potential market failure. Simply put, in the crowdworking market, researchers act as buyers and participants act as sellers. In this market, there is a conflict of interest between researchers’ goal of obtaining authentic, high-quality data for the lowest possible price and participants’ goal of creating acceptable-quality data with the lowest possible effort. Participants generally work in precarious conditions (e.g., low pay, lack of social insurance, pressure to perform, no long-term contracts, time pressure), in which GenAI use represents a rational optimization strategy. Specifically, participants can use GenAI to improve data quality (checking grammar, reflecting on their responses, etc.) or to reduce effort (saving time and cognitive resources) [10, 14]. The problem lies not only in the misconduct of participants, but also in the incentive structures of paid online research [18]. Moreover, the use of GenAI creates an information asymmetry between researchers and participants. Researchers face uncertainty when evaluating data quality because they do not know whether and how participants used or will use GenAI. In response, they need to implement quality control measures, increasing overall costs due to more complicated study setups, additional analyses, and greater study length (longer instructions, more burdensome quality controls). For researchers who expect lower average quality due to undesired GenAI use, it may appear rational to respond by generally treating participants with suspicion, lowering payments, adding more checks, or rejecting more responses. However, these reactions fuel adverse selection among participants. Honest, careful, and intrinsically motivated participants may leave the market because they perceive the compensation as too low relative to the effort involved. At the same time, more restrictive procedures can further incentivize extrinsically motivated participants to reduce effort by using GenAI. This cycle can lead to a “death spiral”, in which quality deteriorates because only low-effort workers and the “fittest” GenAI users, those best at concealing their use, are willing to participate, while the number of high-effort workers willing to participate simultaneously declines. The question arises whether there is a way out of this spiral. Ever-increasing quality control measures do not solve the underlying problem. For example, Santy et al. [18] propose designing crowdworking tasks in a way that makes them easier, more intrinsically appealing, and hedonically motivating by introducing gaming elements. However, this approach may not be applicable to all studies. We argue that there may be approaches that do not require any changes in study design. First, the payment must be competitive compared to other sources of income to prevent participants who invest significant effort and deliver high-quality work from turning away. Yet, higher payment alone resolves neither the information asymmetry nor the remaining “moral hazard” [30], i.e., the incentive to reduce effort through undetected GenAI use. Second, high-quality participants should have the opportunity to credibly demonstrate data quality and honest effort by opting into verification mechanisms that are costly for low-quality participants to imitate [30]. One such opportunity may be an opt-in “verified participation mode”: Participants position a webcam so that both the keyboard and the monitor are visible – this serves as proof of participation without outside assistance and significantly hinders GenAI use. These precautions would be comparable to the monitoring measures used for online exams. As such precautions require additional technical and personal effort from participants and increase privacy concerns, payment should compensate accordingly. To keep the costs of video verification within reasonable limits, the checks could be based on random sampling, a combination with behavioral data, and a “strike system”, in which the first violation could lead to a warning, the second to a temporary suspension from the “verified participation mode”, and the third violation to permanent exclusion from the “verified participation mode”. To address data privacy concerns and potential sample selection bias, crowdsourcing platforms could act as neutral intermediaries for the verification process, and participants could use webcam settings that do not show their faces. Such precautions must not become the default for ordinary online studies, but in high-stakes or vulnerable study designs, they could be useful. Accordingly, several research platforms already allow researchers to require activated webcams for participation. An opt-in “verified participation mode” is also beneficial for participants, as it protects them from being suspected of improper GenAI use and from being subsequently excluded from a study. While an opt-in “verified participation mode” results in higher costs per participant, overall costs might decrease due to a lower number of participants who are excluded from the sample, shorter studies, less time spent on analyses, and more trustworthy data. In sum, the goal should not be to introduce mandatory surveillance, but to create an environment of fair payment, which reduces the incentive for GenAI use while increasing data quality and transparency. Recommendations for researchers and reviewers Due to the different levels of participants’ GenAI use, it is important to balance the feasibility, fairness, and effectiveness of different prevention and detection measures. Maintaining data authenticity and research validity will, therefore, require more resources and engagement from both researchers and reviewers. In this section, we derive recommendations for researchers and reviewers. Recommendations for researchers Decide if monetary incentives are the only option Extrinsic motivation is a strong incentive for using GenAI. Although unpaid participation does not prevent bot activity and GenAI use, researchers might think about ways to find intrinsically motivated participants to achieve better data quality. For example, rather than relying on crowdsourcing platforms, researchers might recruit their samples through online forums and groups (e.g., Reddit, Wikipedia, Facebook) [18] or by involving professional associations; however, factors such as sample size, sampling bias, and accessibility should be considered. In addition, researchers could attempt to design studies in such a way that participants benefit directly from the results. For example, managers may be interested in learning how other managers work, or individuals might be interested in learning about themselves by participating in studies concerning personality or by comparing their performance to that of others in assessment settings. Evaluate the extent to which GenAI use may be acceptable for your study Before deciding on prevention and detection measures, researchers should consider the extent to which GenAI use may be acceptable for their studies. While the use of GenAI to produce complete answers may not be acceptable, GenAI use could well be permissible as an aid [13]. From our own experience, we know that some participants use GenAI to translate their native-language responses into English, reflect on their responses, or try to understand study instructions. Such use may not necessarily compromise data validity and can even improve data quality. However, researchers should be aware of the additional borderline cases that may arise from permitting such use. Regardless, researchers should establish clear rules regarding the use of GenAI. In general, researchers should refrain from permitting the use of GenAI in full-scale data collection. However, permitting the use of GenAI in pretests and pilot testing can help researchers better understand how and where participants use GenAI. Researchers could use this information to tailor their prevention and detection measures accordingly and to investigate how GenAI use shapes participants’ responses [4, 10]. These insights can also help researchers better identify synthetic or GenAI-assisted responses during the main study. Determine the measures for preventing and detecting GenAI use before data collection Researchers should determine the measures for preventing and detecting GenAI use before data collection to align technical and analytical necessities with ethics and integrity [20]. In order to address the different levels of participants’ GenAI use and to decrease the number of false positives and false negatives, it is important to implement different prevention and detection approaches [8, 11]. Researchers should also allocate additional time for planning and conducting research, as prevention and detection measures consume time for programming and testing [25, 31]. Leverage institutional expertise and oversight We recommend that researchers engage with institutional review boards and ethics committees, as these institutions can provide a platform for collecting and exchanging best practices on how to prevent and detect participants using GenAI [31, 32]. As mentioned earlier, collecting personal data, such as IP addresses or email addresses, may not be allowed in all countries. Given their knowledge of research ethics and local legislation, institutional review boards and ethics committees could provide their expertise to establish standard procedures against participants’ GenAI use. Some institutional review boards may not yet have developed expertise on participants’ GenAI use. In such cases, we encourage researchers to proactively provide their knowledge and insights to help build institutional expertise and practices. We also advise researchers to preregister their plans for data collection, data preparation, and data analyses. While it is impossible to anticipate all the instances of participants’ use of GenAI, preregistration promotes preparation and transparency for researchers and reviewers. When facing anomalies or instances that require deviation from preregistered procedures, researchers should openly report these deviations. Plan for larger sample sizes and alternative recruitment Depending on the research context and the participant pool, the proportion of GenAI users and unusable responses can reach up to 90% [31, 33]. Researchers should therefore avoid committing all research resources before assessing the prevalence of GenAI use. Removing such large proportions of participants can lower statistical power and call into question the appropriateness of the current data collection approach. In such cases, researchers should consider alternatives for data collection. For samples with lower proportions of GenAI users, researchers should plan budgets with sufficient flexibility to recruit additional participants if needed [19, 22]. Specifically, when applying self-reports that incentivize participants’ honesty, for example, by promising no negative consequences for admitting to GenAI use, researchers may need to recruit larger samples and therefore allocate higher budgets. Moreover, researchers should be aware that when large proportions of participants in their sample use GenAI, reviewers may call into question the general trustworthiness of the collected data. One mitigation strategy to avoid large proportions of GenAI users could be to recruit participants from platforms that proactively implement measures against participants using GenAI [20]. Similarly, researchers might consider including filters for participants with high approval ratings in previous submissions to improve data quality [34], although there is as yet no clear empirical evidence on whether higher approval rates correlate with lower GenAI use [4]. Consider the consequences for participants Measures for preventing and detecting participants’ GenAI use lead to several consequences for participants. It is therefore important to consider these consequences when determining the research design. First, preventing and detecting GenAI use introduces challenges for participants who use GenAI with honest intent and participants who depend on technology to facilitate their participation. Some participants use GenAI to reduce language barriers, to clarify their responses, or to better understand study instructions [14]. Others may use GenAI to convert text to speech or speech to text to overcome impediments, such as the inability to read or write [11]. In these cases, measures for preventing and detecting participants’ GenAI use lower the equality of participation opportunities for participants with impediments. Researchers should be especially aware of these consequences in settings with sensitive topics and vulnerable groups, and should clearly communicate the rules for using GenAI during the study to participants [28]. Second, measures for preventing and detecting GenAI use also affect participants who do not use GenAI at all. Participants have voiced their discomfort caused by a feeling of surveillance, being irritated by redundant questions, and being fatigued by longer studies [14, 35]. Researchers should therefore consider how measures for preventing and detecting participants’ GenAI use could change participants’ cognition and behavior. Such consideration is especially important in experimental settings, where the measures might interfere with treatment effects [11]. Third, well-paid studies on crowdsourcing platforms often require high approval rates, which require a lot of effort to attain and maintain. As most detection measures only provide indications, not proof, researchers should reduce the chances of false positives as far as possible. Researchers should therefore not lightly flag participants or withhold their payment due to the suspicion of GenAI use. Moreover, it is important to clearly inform participants at the earliest opportunity about the prevention, detection, and consequences of participants’ GenAI use to provide sufficient information to them to enable informed consent for participation under these rules [36]. Informed consent about procedures and consequences is especially important because participants often do not possess sufficient insight into the reasons for rejections. Prepare for uncertainty Participants using GenAI introduce an additional layer of uncertainty to research. In some cases, researchers will intuitively suspect GenAI use but will be unable to find definitive proof [32]. Also, some participants just seem too good to be true, or answers are highly similar, though not similar enough to flag them as GenAI-generated. Since it is impossible to prevent and detect all cases of GenAI use, researchers need to adjust to a reduced sense of closure and control in the research process. Another uncertainty involves sanctioning participants for GenAI use, especially on crowdsourcing platforms. From our own experience, some participants will ask for the opportunity to return their submission instead of being rejected for GenAI use; others will claim technical issues as causes for anomalies in behavioral data, and in extreme cases, participants will threaten to file complaints with the platform administration. There are online forums that discuss some researchers’ misbehavior or warn others about studies, sometimes without anonymization (see Prolific’s Reddit page for examples: https://www.reddit.com/r/ProlificAc/). Given the fact that there is often only an indicator of GenAI use, researchers should act with caution and transparency. Researchers must recognize that they are usually in a more powerful position than participants are and that they bear the responsibility for the fair treatment of participants. In sum, researchers will face uncertainty in evaluating participants’ GenAI use and in making fair decisions about inclusion, exclusion, and compensation. Recommendations for reviewers Develop personal and organizational reviewing standards Reviewers should actively develop their knowledge of the strengths and weaknesses of preventing and detecting participants’ GenAI use. This knowledge not only supports their ability to review research, but also enables them to set reviewing standards. Given the current lack of standardized procedures, reviewers should proactively consult with editors, especially when evaluating GenAI‑related prevention or detection measures for the first time. Updating journal guidelines and review protocols with explicit standards for preventing, detecting, and managing GenAI use [7], similar to the GenAI policies for authors, would be valuable for researchers and reviewers alike. Importantly, GenAI introduces an additional workload for reviewers. In this regard, we encourage editors to set clear reviewing standards to facilitate research evaluation. Specifically, clear standards could reduce redundant work and disagreements among reviewers. Evaluate consistency and check for biases Reviewers should validate whether researchers followed predetermined procedures [37], because there is analytical freedom in handling participants’ GenAI use [28]. In this regard, reviewers should be especially aware of inflated Type I errors, similar to the “garden of forking paths” problem, when there are multiple ways of detecting and handling GenAI use by participants for analyses. When consistency is in doubt or preregistration is unavailable, reviewers should insist on sensitivity analyses to assess the influence of GenAI use on data and to reduce erroneous inferences [4, 28]. Researchers who have not used any prevention and detection measures Due to the lack of best practices and software support, researchers may not have applied any prevention and detection measures against GenAI use. Post-data collection detection measures are limited in accuracy and generate results that are open to interpretation. For example, completion times, straight-lining, response redundancy, and AI detection software require cutoff thresholds and usually provide only indications of GenAI use. Such post-data collection measures also do not sufficiently cover participant behavior in detail or verify metadata, hence reducing methodological rigor. In general, reviewers should encourage researchers to apply post-data collection measures, and reviewers should proactively propose detection measures and cutoff thresholds to researchers to guide them to balance false positives and false negatives in detecting GenAI use. In cases where researchers cannot apply any prevention and detection measures, reviewers should first consider the research context, i.e., the research design, sample characteristics, and any pre-existing quality control measures, for assessing the vulnerability of data collection to participants using GenAI. For example, some crowdsourcing platforms already apply strong quality control measures that reduce the propensity for large-scale GenAI use. In these cases, it may methodologically suffice to mention the lack of prevention and detection measures as a study’s limitation. If necessary, demand resampling In cases where there are strong concerns about data quality without the opportunity to apply post-data collection measures, it might become necessary to collect new data with prevention and detection measures against participants’ GenAI use to ensure data quality [23]. Reviewers should then provide clear guidance to researchers by communicating their expectations regarding procedures, proposing relevant literature, demanding preregistration, and encouraging coordination with institutional review boards and ethics committees. Reviewers should consider the fact that, depending on the location, researchers may not be permitted to apply all types of prevention and detection measures [31]. Conclusions Participants using GenAI to generate responses in online studies introduce a new layer of complexity and uncertainty to conducting, participating in, and reviewing research. While it is important for researchers to proactively implement measures for the prevention and detection of participants’ GenAI use, it is also important to understand how their actions shape the system of crowdsourced participation. Lacking measures against participants using GenAI prolongs the review process for both researchers and reviewers. However, prevention and detection measures influence the conditions for participants and invite adaptation. The current lack of guidelines and standards for prevention and detection impedes a general consensus about the basic requirements for conducting and reviewing data collection in online studies and appropriate tools and infrastructure. While the use of GenAI is still generally detectable, GenAI will further improve in simulating human participants, thus threatening data authenticity and validity. Therefore, the technological advances of GenAI will pose an ongoing challenge to prevention and detection efforts, meaning that researchers and reviewers should regularly update their knowledge and practices. In our comment, we have demonstrated one potential way out of this “death spiral” by maintaining fair payment and establishing an opt-in “verified participation mode”. We encourage further research and proposals to resolve this issue on an infrastructural level.