Psychologists navigate AI’s growing role in assessment

The most widespread use of AI in psychological assessment is report writing: A psychologist conducts the interviews, administers the tests, and takes the notes, while an AI tool summarizes the material, drafts report sections, and in some cases, offers interpretations as well.

From disability claims to prison sentences, psychological assessments often have lasting consequences. In clinics, schools, workplaces, courtrooms, and beyond, psychological evaluations help people access appropriate care and can inform decisions that alter the course of their lives. Now, as artificial intelligence starts to shape how these tests are created, interpreted, and even taken, psychologists are faced with new opportunities—as well as new risks.

AI is already being used to streamline report writing and other time-consuming tasks, freeing up psychologists to focus more on evaluation, interpretation, and decision-making. Researchers are also exploring how AI can be used to generate new test items and to embed assessment into real-world settings, including to help detect mental health risks.

But the same tools are also introducing new vulnerabilities. General-use chatbots can provide test takers with increasingly sophisticated coaching on how to present symptoms or manipulate responses, raising concerns about test security. And across assessment contexts, experts warn that overreliance on AI—especially for interpreting test results—can lead to misdiagnosis, incorrect classifications, and breaches of patient privacy.

“Many people in our field are being really thoughtful on questions of confidentiality, privacy, and bias. But in some cases, the technology is outpacing our ethical analysis,” said Linda McGhee, JD, PsyD, a clinical psychologist and attorney who is a visiting professor at George Washington University and president of the assessment psychology section of APA Div. 12 (Society of Clinical Psychology). “Our field has a special and unique responsibility to carefully analyze and responsibly integrate AI.”

AI tools, including both popular generative AI chatbots such as ChatGPT and Claude and more specialized tools built for supporting psychological assessment, can be highly useful. But because the stakes are so high, psychologists say the field must proceed with caution.

To provide guidance on navigating AI in assessment contexts, APA’s Committee on Psychological Tests and Assessment (CPTA) has released a report on the responsible use of AI in assessment (see Further reading). The comprehensive document addresses ethical and practical considerations across domains, including clinical, industrial and organizational, and school psychology.

“When it comes to psychological assessment, we are often gatekeepers to resources and accommodations. That means we can negatively impact lives very seriously if we make an erroneous decision,” said A. Jordan Wright, PhD, a clinical associate professor of applied psychology at New York University (NYU) who has authored several widely used books on psychological assessment.

The most widespread use of AI in psychological assessment is to aid in report writing, for which multiple tools now exist. A psychologist conducts interviews, administers tests, and takes notes; an AI tool summarizes the notes and test data, writes various sections, and—in some cases—provides interpretations.

“It streamlines the process so we can spend less time writing reports and more time actually conducting assessments,” Wright said.

Some psychologists also use general-use AI chatbots such as ChatGPT as clinical brainstorming partners to help with case conceptualization, assessment, and even diagnosis. That’s due in part to their ease of accessibility and use, said Tony Rousmaniere, PsyD, executive director of the nonprofit Sentio Counseling Center based in Los Angeles, California.

Some baseline ethical guidance for using AI in assessment reports: Review outputs, disclose use, and do not enter protected health information (PHI) into general purpose AI apps, said Paul Ingram, PhD, an associate professor of psychological sciences at Texas Tech University who recently coauthored a paper on the ethical use of AI in assessment (Professional Psychology: Research and Practice, Vol. 57, No. 3, 2026). Keep in mind that tools such as ChatGPT are not HIPAA compliant, and any data shared may be retained or, depending on the platform and settings, used to train future models.

Fortunately, ethical AI use does not have to come at the expense of usefulness, said Alexandre Vaz, PhD, a clinical psychologist and director of training at Sentio Counseling Center who created a free training program on the ethical use of AI for therapists. With appropriately de-identified information, AI can still provide valuable background such as historical or cultural context relevant to a case, Vaz said.

But even when PHI is removed, using general purpose chatbots for case analysis is problematic, said Adam Natoli, PhD, an assistant professor of psychology at Montclair State University in New Jersey and a member of the CPTA, who is studying applications of AI in psychological assessment. Research shows that these tools can perpetuate stigmatizing ideas about mental health and bias toward certain groups, such as African Americans, which could exacerbate existing disparities in diagnosis and assessment (Moore, J., et al., in FAccT ’25: Proceedings of the 2025 ACM Conference on Fairness, Accountability, and Transparency, Association for Computing Machinery, 2025opens in new window; Hofmann, V., et al., Nature, Vol. 633, 2024opens in new window). Natoli is now comparing popular chatbots by giving them the same clinical scenarios to see whether some over pathologize certain groups of people or underdetect issues in others.

One of the risks of using AI is that some clinicians may become complacent and rely on these tools in place of their own judgment, even though the tools can make errors, Natoli said. A flawed interpretation can have serious consequences—a missed or incorrect diagnosis, unnecessary or denied accommodations, or other decisions that can affect a person’s education, employment, custody, or legal standing.

“We need to ask ourselves the same questions about AI tools that we would about any psychological test,” Natoli said. “Has this been validated for the individual in the way I intend to use it? Do I know how to use it? And can I explain it in a way that’s understandable?”

Many AI tools used to support assessment lack guardrails that require psychologists to apply their own interpretations, Wright said. That means psychologists must build their own workflows to critically evaluate AI-generated outputs.

Using these tools is particularly risky in forensic settings, experts argue, because AI tools can produce inconsistent outputs and rely on reasoning that is hard to explain or justify. Legal standards are still evolving, and psychologists who use AI in these cases could be held responsible for errors made by AI (Grabb, D. J., & Angelotta, C., Journal of the American Academy of Psychiatry and the Law, Vol. 51, No. 4, 2023opens in new window).

Beyond its power to streamline writing and analysis, AI may offer a way to transform how psychological assessments are designed and delivered. An emerging approach known as generative psychometrics uses AI to create assessments that adapt to test takers in real time based on their responses.

“One of the biggest criticisms of many psychological tests is that they are static and may not be adaptable to a world that’s continually changing,” Wright said.

AI-generated psychological assessments would still need to meet core psychometric standards, including reliability, validity, cultural fairness, and usefulness. More work is needed to reach that point, Wright said, but this tailored approach could make measurements more precise and reduce test length by ruling out irrelevant items as the assessment progresses.

Psychologist Liberty Munson, PhD, director of psychometrics at Microsoft and a member of the CPTA, is starting to apply this approach. She oversees tests that lead to Microsoft certifications, which are standardized credentials that verify a worker’s technical skills, such as cloud engineering capabilities on Microsoft Azure. Her team built an AI item-generation tool that creates new test questions, which are then reviewed by a subject matter expert. They’re also using AI to build in “flow of work” assessments that allow users to earn recognition in the course of their daily work. For example, workers using Microsoft Copilot to write code or analyze data could have their skills verified automatically as they work.

“There’s a growing interest in embedding assessment into Microsoft technologies so that people can demonstrate their skills without sitting for an exam,” Munson said.

AI is also at the heart of other efforts to embed psychological assessment within real-world settings. Research has shown that machine learning can help predict risk for psychosis based on speech patterns (Olah, J., et al., Frontiers in Psychiatry, Vol. 14, 2024opens in new window). Building on that work, Natoli and his colleagues from the NaphCare Charitable Foundation are now exploring whether the technology can help flag at-risk individuals during the booking stage in correctional facilities.

Another use of AI is to refine existing tests. Natoli used natural language processing to compare translations of psychological assessments across languages, finding substantial differences in the emotional tone of translated test items that could affect how they are interpreted (Assessment, Vol. 33, No. 2, 2025opens in new window).

“We’re now looking beyond validity to explore whether the sentiments are consistent between languages,” he said. “For example, are items that are worded optimistically in English also worded optimistically in Spanish?”

AI chatbots are also being used on the other side of the equation—helping test takers manipulate responses to either exaggerate or downplay their symptoms or abilities. Information about how to answer psychological assessments has floated around the internet for years, but AI makes that information much more accessible, said Iulia Crișan, PhD, a clinical psychologist and researcher studying AI-supported feigning at West University of Timișoara in Romania.

“Someone can now create a virtual agent that trains them on how to correctly feign on a certain instrument,” Natoli said. “It would be like meeting with a lawyer-psychologist-acting-coach hybrid.”

Early research on feigning with AI suggests that even standard AI chatbots can provide sophisticated guidance, such as instructing test takers to only make “mistakes” on certain items and to avoid extreme scores that trigger detection. But in studies of feigning depression, attention-deficit/hyperactivity disorder (ADHD), and head injuries, validity measures were still able to catch most AI-supported feigners (Battista, F., et al., Current Psychology, Vol. 45, No. 379, 2026opens in new window; Bensink, S., The Role of AI-Consultation in Malingering Detection During ADHD Assessments: An Insight in the Student Population [unpublished master’s thesis], University of Groningen, 2025opens in new window; Voin, S., et al., Applied Neuropsychology: Adult, online first publication, 2025opens in new window).

That security may be short-lived, however, as users learn more effective prompting strategies, Ingram said. In one study, researchers gave students more guidance on using ChatGPT to feign ADHD symptoms—and detection rates dropped (Fuermaier, A. B. M., & Niesten, I. J. M., Psychological Injury and Law, Vol. 18, 2025opens in new window). Other research on feigning psychosis and neuropsychological symptoms suggests that ChatGPT and Gemini pose a real threat to test security (Lavigne, S., et al., The Clinical Neuropsychologist, Vol. 40, No. 3, 2026opens in new window; Gershan, S. A., et al., Journal of the American Academy of Psychiatry and the Law (Vol. 54, No. 2, 2026opens in new window).

“AI poses a substantial risk, because many of our feigning measures rely on very simple detection strategies,” said Richard Rogers, PhD, a Regents Professor of Psychology at the University of North Texas who conducts research on how and why people feign mental disorders.

For example, some stand-alone performance validity tests are designed specifically to detect feigning. If AI tools can identify them, it may become easier for people to game the test. Rogers is exploring ways to make psychological assessments more secure, such as developing more complex, pattern-based detection strategies that are too cognitively demanding to manipulate.

“The idea is to anticipate that people will try to prepare for these tests and to build detection strategies that create cognitive overload for those who are trying to feign,” Rogers said.

Experts agree training in AI can no longer be considered a specialty area within psychology, given how ubiquitous these tools have become.

“AI is already shaping practice, and we can’t claim to be providing a complete education if we don’t prepare students to use these tools responsibly,” said McGhee, who spoke at APA 2026 about AI and assessment instruction.

As director of NYU’s clinical/counseling psychology doctoral program, Wright allows students to use AI but builds in both ethical and clinical guardrails to ensure they do so safely. To protect patient privacy and confidentiality, trainees must know exactly where the data they share is being stored and how the data will be used by the company behind a given tool. On the clinical side, Wright asks students to evaluate and compare AI outputs to assess their accuracy and usefulness.

Part of today’s challenge is that general purpose chatbots are accessible and easy to use, while many specialized assessment tools are still being refined and validated. But that could change as more research-backed tools reach the market, said Sarah Kopelovich, PhD, a professor in the Department of Psychiatry and Behavioral Sciences at the University of Washington School of Medicine. For example, Kopelovich helped build an AI tool for clinicians with a Small Business Technology Transfer grant, an NIH initiative designed to quickly bring evidence-based tools to market.

“We need to get research-backed tools out there as soon as possible,” she said. “Otherwise, people will continue to rely on ChatGPT and other general purpose chatbots.”

Above all, psychologists must stay engaged in monitoring and intervening in how AI is used in assessment, Natoli said, including by developing formal guidance on the responsible use of AI assessment tools. The new report from CPTA provides a helpful framework for how to vet AI tools, interpret reports, and navigate ethical issues such as bias and confidentiality. Psychologists can also use their voices to address the increasing number of apps that provide access to psychological test content without the guidance of a trained practitioner. Without appropriate oversight, such tools could undermine test security, encourage self-diagnosis, and make it harder to ensure that tests are interpreted in the right clinical or legal context.

“We need to be assertive in areas of policy and industry, saying what is and is not OK when it comes to the use of AI in psychological assessment,” Natoli said.