Faking Personality Assessment

Employers increasingly recognize that success in many roles depends not only on a candidate’s technical skills but also on how they typically behave – whether they are dependable, adaptable, cooperative, or resilient, for example. To make informed hiring decisions, organizations try to match candidates to job profiles based not only on what they can do, but also on who they are. 

This approach assumes a predictive relationship between test scores and future workplace behavior: If we know how someone tends to behave across time and in different situations, we can make educated guesses about how they will perform in a given job context (e.g., Tett et al., 1991). But this predictive power hinges on a critical condition – that the test scores accurately reflect the candidate’s true personality traits. 

Since the early days of personality testing, however, candidates have found ways to intentionally distort their responses. This behavior, known as faking, has been a long-standing concern in personnel psychology. Faking is broadly defined as a deliberate, goal-directed behavior where individuals consciously shape their responses to create a favorable impression – often based on what they believe the employer wants to see (Ziegler et al., 2011). In this sense, faking is not a personality trait in itself, but a motivated behavior. 

The consequence of faking is straightforward but serious: If a person’s answers are not an honest reflection of their traits, then any predictions made based on those scores – such as their likelihood to be reliable, collaborative, or stress-resilient – become less valid. Research confirms that faked scores undermine the validity of personality assessments, making it harder for organizations to identify the right people for the job (MacCann et al., 2011; Tett & Simonet, 2021; Ziegler et al., 2011). 

Importantly, not everyone fakes, and not everyone fakes to the same extent. Whether someone engages in this behavior depends on the interaction of personal and situational factors (Heggestad, 2011). For instance, high-stakes situations such as competitive job applications create a strong incentive to fake, especially if the candidate understands the role’s requirements and believes they can game the system. Yet not everyone will act on that motivation. A person who is anxious about being dishonest – or fears getting caught – might still respond truthfully, even in a high-pressure context (MacCann et al., 2011). 

Distinguishing between faked and honest responses is not always straightforward. An unusually high or low score might raise suspicion, but it could just as well reflect an authentic response from an unusual person. There are methods that allow us to evaluate whether a person has faked: Validity items can provide information about the test taker’s attention level during the assessment (e.g., Meade & Craig, 2012). Impression management scales can offer insight into how much a person tends to alter their answers to reflect more socially desirable traits and behaviors (Crowne & Marlowe, 1960). Bogus items can be used in the overclaiming technique to gauge how much people inflate the presentation of their knowledge in certain areas (Paulhus et al., 2003). As a test publisher, Hogrefe has long relied on these and other psychometric safeguards to support the responsible use of personality assessments. All of these methods work to some degree, which will also be discussed further in a later part of this white paper. However, they never allow for complete certainty if a person has faked their re - sponses, only whether responses should be treated with caution. This inherent ambiguity makes it difficult to confidently identify faking, even when there is reason for concern (MacCann et al., 2011). 

The emergence of generative artificial intelligence (AI) tools further complicates this picture. For the first time, individuals have access to powerful systems that can generate highly plausible, tailored responses to personality questionnaire items (Phillips & Robie, 2024a, 2024b; Robie et al., 2025), making it even harder to tell whether a test result reflects the person or the prompt they gave to an AI. Importantly, traditional methods that have proven useful against human faking may not work as expected when applied to responses that were generated or assisted by AI (Robie et al., 2025). This underscores the need for specific guidelines and new methods to detect and prevent faking in the age of generative AI.

Empirical data indicate that approximately 30–50% of applicants have, at some point, misrepresented themselves during a hiring or selection process (Griffith & Converse, 2011). More specifically, Griffith and Converse (2011) found that 32% of respondents admitted to exaggerating positive traits on questionnaires, and 15% acknowledged providing outright false responses. With the increasing accessibility of generative AI, this behavior appears to be evolving. Recent surveys suggest that between 29% and 38% of respondents report being at least somewhat likely to use AI to assist them in completing recruitment assessments (Westfall, 2024; Thresher, 2024). 

Until recently, individuals who sought to fake their responses faced a number of cognitive and strategic challenges. Effective faking required an understanding of the target role, insight into what the employer was seeking, and the ability to map those expectations onto specific questionnaire items. The success of faking was, in part, limited by an individual’s ability to strategically manage impressions – something that not all candidates were equally capable of doing (e.g., McFarland & Ryan, 2000). In some cases, this behavior may even have been interpreted as a form of test-taking skill or situational intelligence. 

The introduction of generative AI technologies has changed this dynamic. Tools such as large language models (LLMs) – for example, ChatGPT – are now capable of producing coherent and contextually appropriate responses to simple user prompts (Budhwar et al., 2023). By lowering the effort required to fake effectively, these tools may allow a much broader range of applicants to generate responses that align closely with role expectations. Initial evidence suggests that generative AI can, in some cases, outperform human fakers (Phillips & Robie, 2024a, 2024b), and that traditional detection methods may not be well equipped to identify such responses (Robie et al., 2025). In the next section, we review recent studies that directly examine the capacity of LLMs to fake personality assessments.

These initial studies by Phillips and Robie (2024a, 2024b) provide a valuable starting point, but several open questions remain. The current research sheds little light on how to detect AI-generated responses once they are submitted. It is difficult to draw conclusions from patterns in aggregated data to make decisions about individual profiles. More importantly, the models themselves are evolving rapidly. Future versions may be better able to simulate traits like integrity, decode more complex question formats, and tailor answers with even greater nuance. 

This highlights the need for ongoing research and vigilance. At Hogrefe, we are actively working to investigate and monitor AI-assisted faking behaviors, and to explore mechanisms for identifying such patterns in assessment data. This includes both psychometric and technological solutions. 

In the meantime, what can organizations and practitioners do now to minimize the risk of AI-assisted faking? To answer this question, we review traditional strategies for detection and prevention as well as the avenues opened by the introduction of generative AI, in the next section.

The results of the study by Robie et al. (2025) offer two key takeaways: First, LLMs are not consistently superior at faking, but they are already capable of producing responses that rival or exceed those of human fakers in some contexts. Second, traditional detection methods show important limitations, underscoring the need for refinement and revalidation as AI tools continue to develop.

Paradoxically, while generative AI presents a new avenue for faking, it may also offer new opportunities for faking-detection. To use an LLM to assist in faking, a candidate must typically provide the model with a job profile or summary, as well as the actual test items or closely paraphrased prompts. This interaction can leave detectable behavioral traces. 

For example, system metadata such as response times, tab-switching events, and browser activity patterns can be used to identify anomalous test-taking behavior. Candidates who copy test items into an AI tool may display atypical pausing, refocusing, or tab navigation behaviors that suggest external tool usage. These so-called page-focus events can be recorded in online assessment tools, such as the Hogrefe Test System (HTS), which provides a report on test-taking behavior after a test session has been completed. This report also includes response times, which could be used to evaluate the likelihood of AI-assisted faking. However, the mentioned indicators are not definitive on their own, and it can never be guaranteed that a person has actually faked their responses. However, they could inform the development of risk indices to help flag potentially AI-assisted responses. 

The usefulness of metadata to detect AI-assisted faking is complicated even more by the emergence of AI agents (e.g., Open AI, 2025). This new feature of some AIs lets users give an LLM a task which it then carries out on its own – for example, to book a table at a restaurant via online booking services. This kind of AI behavior still has to be investigated and might produce patterns of metadata that are different from candidates interacting with an AI during an assessment. 

However, it is important to emphasize that no faking-detection system can guarantee perfect accuracy. False positives remain a concern, and the current evidence base is too limited to justify widespread adoption of such tools. More research is needed to explore how and under what conditions AI-assisted faking can be identified. At Hogrefe, we are actively contributing to this research and are committed to developing evidence-based solutions. Until more robust and validated detection methods are available, we recommend that practitioners exercise caution when interpreting personality test results that appear inconsistent, implausibly extreme, or unusually well-aligned with job requirements. In such cases, supplementary assessment methods – such as retesting, structured interviews, or reference checks – may help clarify remaining uncertainties. However, a sole focus on detection is unlikely to provide a complete solution. It is equally important to consider preventive strategies that reduce the likelihood of faking before it occurs. Prevention holds a distinct advantage: It eliminates the need for difficult, and often legally sensitive, decisions regarding how to treat suspected cases of faking (MacCann et al., 2011). In contrast to detection-based approaches, preventive measures mitigate risk without requiring post hoc judgment about a candidate’s honesty. 

In the following section, we outline a range of prevention strategies that can be integrated into assessment design and administration to help safeguard the integrity of personality assessments – particularly in the age of AI.

One strategy to mitigate both traditional and AI-based faking is to use assessment formats that are inherently more difficult to manipulate. For example, asking candidates to provide verifiable information may increase accountability and discourage exaggeration. A question such as “How many hours of overtime did you work last month?” is more likely to elicit a truthful answer than a general statement like “I take on extra work,” because it is more concrete and potentially subject to verification (MacCann et al., 2011). 

Situational judgment tests (SJTs) offer another approach. These assessments present test takers with realistic workplace scenarios and ask them to evaluate or select from possible responses. Evidence suggests that SJTs are less vulnerable to faking than traditional Likert-type rating scales, potentially because evaluating specific behavioral responses in context demands more cognitive effort than endorsing generalized statements (Nguyen et al., 2005), and the “correct” answer is harder to determine. One example is Hogrefe’s Leadership Judgment Indicator-2, which measures leadership behavior through hypothetical situations in which candidates must choose how they would act, from four possible options. By grounding responses in context, such instruments make it more difficult to simply select the “ideal” answer. However, even with SJTs, faking remains possible for humans (Peeters & Lievens, 2005), and it is not yet known how effectively AI models can simulate desirable responses in SJT formats. 

Forced-choice formats have also been shown to be less fakable, particularly in recent studies discussed above (see the section AI Can Outperform Humans – With Some Limitations; Phillips & Robie, 2024a, 2024b). In this format, test takers are required to choose between equally socially desirable statements, limiting their ability to endorse all positive traits at once. While effective in curbing faking, forced-choice formats result in ipsative or partially ipsative scores, which complicates comparisons across candidates – an important limitation for use in selection contexts.

Additionally, the use of warnings raises important ethical and legal considerations. If test takers are warned that deception will be detected and punished, but no valid detection method is in place, a warning may be considered misleading. This could create legal exposure for organizations and reduce trust in the assessment process (MacCann et al., 2011). In addition, warnings may have unintended consequences for honest test takers, particularly those who are anxious. Some individuals may respond more conservatively or alter their answers unnecessarily out of fear of being falsely flagged as deceptive. 

For these reasons, warnings must be used thoughtfully. Their wording should reflect the actual capacities of the testing system, and they should be designed to reduce faking without introducing undue stress or confusion for genuine candidates. In the case of generative AI, test takers may benefit from clear communication explaining that the use of AI tools to influence test results could misrepresent their suitability for the role – and may ultimately work against their best interests.

As generative AI becomes more accessible, technological safeguards are gaining importance as a means of preventing its misuse during assessment. Although empirical data on their effectiveness is still limited, several practical interventions have emerged as promising options. 

Proctored testing environments, either in person or online, offer one of the most reliable means of deterring AI use. By restricting access to electronic devices and monitoring candidate behavior, these settings can significantly limit opportunities for external assistance. When in-person testing is not feasible, remote assessments can be administered through digital platforms that incorporate browser restrictions – such as blocking copy–paste functionality, disabling screenshots, or preventing candidates from leaving the testing tab. Using assessments with time limits reduces the opportunity for individuals to consult AI tools during the test session. 

Webcam-based proctoring may serve as an additional safeguard, although this approach raises privacy concerns and may not be appropriate in all contexts. In such cases, it is essential that candidates are informed in advance about the use of webcam monitoring, that they provide explicit consent, and that they are given clear information on how their data will be handled, stored, and protected. Regardless of the method, it is important to maintain transparency: Test takers should be informed about any monitoring tools or restrictions that are in place. This not only ensures compliance with data protection regulations but also helps support candidate trust and reduce dissatisfaction with the testing process. 

It is important to acknowledge that no single method can fully prevent faking, especially as AI models continue to improve. However, a combination of assessment selection, warnings, and technological controls can substantially reduce the likelihood of dishonest responding and help preserve the validity of test results.

Finally, some researchers have begun to propose a shift in perspective. Rather than treating AI use solely as a threat to test validity, it may be appropriate in certain contexts to consider AI use as a relevant skill in its own right. Lievens and Dunlop (2025) argue that if the ability to use AI is essential to the job being assessed, then AI-assisted performance on assessments might reflect job-relevant behavior rather than response distortion. In such cases, AI use could be incorporated into the design of the assessment itself, transforming what would otherwise be considered error variance into a construct of interest. 

While this perspective is not yet widely adopted, it underscores the importance of context in assessment design. Not all uses of AI are inherently deceptive or undesirable, and future practices may benefit from a more nuanced understanding of when and how such tools are used.

In selection contexts, the primary role of personality assessment is often to identify candidates who best fit the requirements of the position. In this sense, personality tests are commonly used to “select-in” the strongest applicants. However, research suggests that faking poses less of a threat when personality tests are used differently – namely, to “select-out” candidates who clearly do not fit the position (MacCann et al., 2011). 

The reasoning is straightforward: Faking primarily affects rank order at the upper end of the distribution, where individuals may displace one another by faking higher scores. By contrast, those who engage in faking are not necessarily drawn from the very bottom of the distribution. As a result, using personality tests to exclude the lowest-scoring group of applicants reduces the disruptive impact of faking on the overall selection outcome. 

Supporting this view, Griffith et al. (2007) looked at cases where the selection ratio is high – specifically, when more than 50% of candidates are selected to continue in the selection process. In such cases, even if some candidates inflate their scores, the relative rank among top scorers is less important, because the majority of them are advancing anyway. What matters more is that those with genuinely low scores are correctly identified and excluded. Put differently, personality assessments may be most effective when used as a screen-out tool to remove the lowest 10–20% of candidates, rather than to identify only the very top performers. 

Taking these considerations into account, the next section of this paper will present a set of recommendations designed to help practitioners respond effectively to the risks of AI-assisted faking in personality assessment.

Using personality assessment to exclude the candidates with inappropriately low levels of desired attributes will give “fakers” a smaller chance to displace desirable candidates in the selection procedure.

Transparent communication about the purpose of the assessment, the importance of honest responding, and any technical safeguards in place can increase test taker compliance, trust, and engagement. Consider also including a note that AI-assisted responses may not reflect a candidate's actual fit for the role and could work against their interests. At Hogrefe Consulting we provide practical advice on how to communicate the implementation of assessments and the measures against faking in place.

Warnings about the detection and consequences of faking reduce deceptive behavior but must be used with care. Avoid overstating the capabilities of detection systems and be mindful of the potential for increasing anxiety among honest respondents. Where possible, frame warnings in an educational way, emphasizing the value of honest responses for achieving good person–job fit, which is in the best interest of both the candidate and the employer.

When existing detection methods suggest that faking has occurred, we advise practitioners to retest or interpret scores cautiously rather than to exclude the identified fakers or try to correct their scores. Supplementary assessment methods, such as structured interviews or reference checks, may help clarify whether results reflect genuine self-presentation and should be applied before making final decisions.

Generative AI technology is evolving rapidly. Assessment providers and HR professionals should continue to monitor new research, updates to AI capabilities, and developments in detection and prevention methods. Hogrefe remains committed to contributing to this research and providing clients with timely guidance.

Finally, it is important to remember that not all uses of AI are inappropriate. In some cases, particularly where AI use is part of the job, applicant’s ability to leverage these tools may represent a job-relevant skill rather than deception. Assessment strategies should remain flexible and context-sensitive.

© 2025 Hogrefe Publishing Group