Reducing the impact of AI-generated misinformation on memory and perception of a police incident.

Generative artificial intelligence (AI) systems are increasingly used to generate police incident reports.Yet such systems can produce errors, which may detrimentally impact legal processes.The present study examined the impact of an AI-generated report on memory and perceptions of a police incident and whether safeguards can reduce those impacts.Participants (N = 432) read an accurate summary of a domesticviolence incident; 24 hr later, they read an AI-generated report that was either accurate or contained misinformation, before completing memory tests and rating their perceptions of the incident.Two safeguards were tested: (a) active error monitoring and (b) watching body-worn camera footage.Misinformation significantly impaired memory and influenced perceptions of the incident-especially in participants who strongly identified with police-and the safeguards were largely effective at reducing these impacts.Findings demonstrate the risks of generative AI usage in forensic settings, requiring the implementation of appropriate safeguards for combatting its detrimental effects. General Audience SummaryGenerative artificial intelligence (AI) systems can quickly create human-like content, offering solutions to many real-world problems.In the legal system, such systems are used to produce police reports from audio sources, including footage from officers' body-worn cameras.While AI report-writing tools are marketed to improve report-writing efficiency, early evidence suggests that they may not save time and could introduce false information into officers' reports.In this study, we examined how a misleading, AI-generated report influences memory and perceptions of a police incident and whether safeguards can reduce the impact of misinformation.Participants read a summary of police responding to an alleged domestic-violence incident; 24 hr later, they read an AI-generated report that was either accurate or contained misinformation-suggesting that the suspect behaved more aggressively toward police than they did.We tested the effectiveness of two safeguards: (a) actively identifying and correcting errors in the misleading report and (b) watching the body-worn camera footage before reading the report.Participants completed cued-recall and recognition memory tests and rated their perceptions of the incident, as well as how strongly they identified with police.As expected, the misleading report significantly reduced memory accuracy and changed how participants perceived the incident, especially among those who identified strongly with the police.Safeguards were largely effective at reducing the impact of misinformation on memory accuracy and perceptions.These findings suggest that misleading content can influence how people remember events even when they are warned that it is AI-generated and that available safeguards are somewhat effective at reducing those effects.The current research highlights the risks of using generative AI systems in legal contexts and the importance of using appropriate safeguards to prevent misinformation from influencing memory and perceptions.