Faking It With AI: What Psychologists Need to Know About Assessment Validity

The study, "ChatGPT Helps Students Feign ADHD: An Analogue Study on AI-Assisted Coaching," published in Psychological Injury and Law, sheds light on how readily accessible AI tools like ChatGPT can be used to coach individuals in feigning psychological conditions. Let's break down the study and discuss what it means for our practices. If you’re on the go, you can listen to the podcast linked above.

The Study: Methods and Design

This analogue study aimed to investigate whether AI-generated coaching could help university students successfully feign ADHD during a clinical neuropsychological assessment. The researchers employed a rigorous methodology:

• AI Coaching Material Generation: First, they gathered questions from 22 students about how to feign ADHD during an assessment. These questions, along with input from the research team, were used to conduct an extensive query with ChatGPT-4. The goal was to develop a concise AI-generated information sheet to coach individuals on feigning ADHD, specifically considering the types of assessments used. The prompt given to ChatGPT specifically asked for information on the assessment's primary goal, the instruments used, how to align responses and behavior with genuine ADHD, how validity is assessed, and strategies to avoid detection.

• Analogue Study: Next, 110 university students were randomly assigned to one of three groups: a control group, a symptom-coached simulation group (provided with DSM-5 criteria for ADHD), and an AI-coached simulation group (provided with the ChatGPT-generated information sheet).

• Assessment Battery: All participants underwent a clinical neuropsychological assessment that included:

• Conners’ Adult ADHD Rating Scale (CAARS)

• Weiss Functional Impairment Rating Scale (WFIRS)

• The Perceptual and Attention Functions–Selective Attention (WAFS)

• Measure of working memory (Reliable Digit Span - RDS) derived from the WAIS-III.

• The battery also included embedded indicators of symptom validity (SVTs) within the CAARS and performance validity (PVTs) within the WAFS and RDS.

• Feigning Instructions: Participants in the simulation groups were instructed to complete the assessment as if they were pretending to suffer from ADHD. They received a vignette emphasizing potential benefits of feigning. Pre- and post-experimental checks ensured participants understood and adhered to the instructions.

Key Findings and Results

The study yielded significant results highlighting the potential impact of AI-assisted coaching:

• ChatGPT Provided Useful Coaching: The AI generated an information sheet that was concise, easy to understand, and contained accurate information about the assessment instruments, including example items and scoring methods. Importantly, it also provided insights into how these instruments detect feigning, such as by analyzing response consistency and extreme scores, and cautioned against exaggeration, emphasizing subtlety.

• AI-Coached Group Was More Nuanced: Compared to the symptom-coached group, the AI-coached simulation group consistently moderated their symptom overreporting and cognitive underperformance. While both simulation groups reported more symptoms and performed worse on cognitive tests than the control group (as expected), the AI-coached group's presentation was less extreme.

• Lower Detection Sensitivity: The detection rates (sensitivity) of the SVTs and PVTs were generally lower for the AI-coached group compared to the symptom-coached group, indicating that the AI-coached individuals were more successful at feigning ADHD in a way that evaded detection. This difference was statistically significant for commission errors on the WAFS.

Implications for Psychologists

This study has profound implications for all psychologists involved in assessment:

• Increased Sophistication in Feigning: The study highlights that readily available AI tools can provide individuals with effective strategies to feign ADHD in a more sophisticated manner, potentially making it harder to distinguish genuine cases from those where a student might be seeking accommodations or other benefits without truly having the disorder.

• Increased Threat to Assessment Validity: The findings demonstrate that publicly accessible AI tools can provide effective strategies for feigning ADHD, posing a significant threat to the validity of our assessments.

• Beyond ADHD: Feigning Other Disorders: My personal take, which aligns with the broader implications of this research, is that this will likely be a problem for many other disorders that rely heavily on checklists/rating scales and self-report, including depression, anxiety, and even autism. If AI can coach individuals to present nuanced symptoms of ADHD, it stands to reason that similar coaching could be developed for other conditions with well-defined diagnostic criteria and commonly used self-report measures.

• The Flip Side: AI and "Faking Good": Conversely, my take is that it's also plausible that individuals could use AI to "fake good" and hide mental health problems, particularly in high-stakes cases such as fitness-for-duty evaluations for law enforcement or in custody disputes. AI could provide guidance on how to minimize reported symptoms and present an image of psychological stability.

• The Crucial Role of Validity Testing: This research underscores the critical importance of incorporating symptom validity tests (SVTs) and performance validity tests (PVTs) into our assessment batteries. These measures become even more essential in an era where AI can assist in more sophisticated feigning attempts. We must be vigilant in interpreting these measures.

• The Necessity of Multimethod Assessment: Relying solely on self-report questionnaires is increasingly risky. Comprehensive, multimethod assessments that integrate clinical interviews, behavioral observations, collateral information, and objective testing are crucial for a more accurate understanding of an individual's psychological state.

• Test Security Concerns: The authors rightly recommend that researchers and clinicians exercise caution when sharing assessment materials, example items, and scoring methodologies. The more information available online, the easier it is for AI to incorporate these details into coaching materials.

• Continuous Learning and Adaptation: As AI technology rapidly evolves, psychologists must remain informed about these advancements and their potential impact on assessment practices. We need to continuously refine our skills in detecting response bias and malingering in this changing landscape.

• Understanding Motivation and Context: A thorough understanding of the individual's motivations for seeking assessment and the potential secondary gains (or losses) associated with a particular diagnosis or presentation remains paramount in interpreting assessment results.

Moving Forward

The findings of this study serve as a timely reminder of the challenges posed by readily available information and increasingly sophisticated AI tools. As psychologists, we must adapt our assessment practices to mitigate these risks, prioritize validity testing, employ multimethod approaches, and remain vigilant in our interpretation of assessment data. The age of AI-assisted feigning (and potentially "faking good") is likely upon us, and our professional responsibility demands that we are prepared.