Mirko Casu

This paper examines the impact of cognitive biases on decision-making in forensics and digital forensics, exploring biases such as confirmation bias, anchoring bias, and hindsight bias. It assesses existing methods to mitigate biases and improve decision-making, introducing the novel “Impostor Bias”, which arises as a systematic tendency to question the authenticity of multimedia content, such as audio, images, and videos, often assuming they are generated by AI tools. This bias goes beyond evaluators’ knowledge levels, as it can lead to erroneous judgments and false accusations, undermining the reliability and credibility of forensic evidence. Impostor Bias stems from an a priori assumption rather than an objective content assessment, and its impact is expected to grow with the increasing realism of AI-generated multimedia products. The paper discusses the potential causes and consequences of Impostor Bias, suggesting strategies for prevention and counteraction. By addressing these topics, this paper aims to provide valuable insights, enhance the objectivity and validity of forensic investigations, and offer recommendations for future research and practical applications to ensure the integrity and reliability of forensic practices.

Mental health disorders are a leading cause of disability worldwide, and there is a global shortage of mental health professionals. AI chatbots have emerged as a potential solution, offering accessible and scalable mental health interventions. This study aimed to conduct a scoping review to evaluate the effectiveness and feasibility of AI chatbots in treating mental health conditions. A literature search was conducted across multiple databases, including MEDLINE, Scopus, and PsycNet, as well as using AI-powered tools like Microsoft Copilot and Consensus. Relevant studies on AI chatbot interventions for mental health were selected based on predefined inclusion and exclusion criteria. Data extraction and quality assessment were performed independently by multiple reviewers. The search yielded 15 eligible studies covering various application areas, such as mental health support during COVID-19, interventions for specific conditions (e.g., depression, anxiety, substance use disorders), preventive care, health promotion, and usability assessments. AI chatbots demonstrated potential benefits in improving mental and emotional well-being, addressing specific mental health conditions, and facilitating behavior change. However, challenges related to usability, engagement, and integration with existing healthcare systems were identified. AI chatbots hold promise for mental health interventions, but widespread adoption hinges on improving usability, engagement, and integration with healthcare systems. Enhancing personalization and context-specific adaptation is key. Future research should focus on large-scale trials, optimal human–AI integration, and addressing ethical and social implications.

While the challenge of distinguishing AI-generated from real images is widely acknowledged, the specific cognitive biases that systematically shape human judgment in this domain remain poorly understood. It is particularly unclear how a general awareness of AI capabilities fosters novel biases, like a pervasive skepticism (“impostor bias”), and how this interacts with established phenomena like “automation bias”. This study addresses this gap by providing the first quantitative analysis of how these two biases operate across five distinct experimental variants designed to test the context-dependency of human perception. Through a mixed-methods study with 746 participants, we demonstrate that human authentication accuracy hovered around chance levels (ranging from 47.0% to 55.5%). However, our analysis provides robust evidence for the systematic operation of cognitive biases. We validate the presence of “impostor bias” through a consistent pattern of higher doubt for AI-generated images and confirm “automation bias” through significant opinion changes following algorithmic suggestions. Our findings reveal that these biases are not uniform across populations: gender was a consistent predictor of automation bias, with males in all five variants showing a significantly stronger and more consistent tendency (Cohen’s d = 0.254–0.683) to be influenced by algorithmic suggestions. In contrast, age and academic background had minimal and highly localized effects. Furthermore, we identified a significant interaction between experimental stimuli and performance over time, isolating a pronounced fatigue effect to a single questionnaire variant where accuracy progressively declined (by approximately 1.7% per trial). By integrating human feedback with Grad-CAM visualizations, we confirm a divergence between human holistic evaluation and the localized focus of machine learning models. These findings carry direct implications for policy, as discussed within the context of the European AI Act, and inform the design of human–AI systems and media literacy programs aimed at mitigating these critical cognitive vulnerabilities.

The landscape of synthetic media has been irrevocably altered by text-to-video (T2V) models, whose outputs are rapidly approaching indistinguishability from reality. Critically, this technology is no longer confined to large-scale labs; the proliferation of efficient, open-source generators is democratizing the ability to create high-fidelity synthetic content on consumer-grade hardware. This makes existing face-centric and manipulation-based benchmarks obsolete. To address this urgent threat, we introduce SynthForensics, to the best of our knowledge the first human-centric benchmark for detecting purely synthetic video deepfakes. The benchmark comprises 6,815 unique videos from five architecturally distinct, state-of-the-art open-source T2V models. Its construction was underpinned by a meticulous two-stage, human-in-the-loop validation to ensure high semantic and visual quality. Each video is provided in four versions (raw, lossless, light, and heavy compression) to enable real-world robustness testing. Experiments demonstrate that state-of-the-art detectors are both fragile and exhibit limited generalization when evaluated on this new domain: we observe a mean performance drop of 29.19% AUC, with some methods performing worse than random chance, and top models losing over 30 points under heavy compression. The paper further investigates the efficacy of training on SynthForensics as a means to mitigate these observed performance gaps, achieving robust generalization to unseen generators (93.81% AUC), though at the cost of reduced backward compatibility with traditional manipulation-based deepfakes. The complete dataset and all generation metadata, including the specific prompts and inference parameters for every video, will be made publicly available at [link anonymized for review].