DSDaniel Szoke

Daniel Szoke

*This is an algorithmically generated author record based on publicly available information.

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article


Daniel Szoke has a total of 33 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Open Access

    Emotional Change as a Process Indicator in AI-Facilitated Cognitive Restructuring: Secondary Analysis of a Mixed Methods Feasibility Study (Preprint)

    by Philip Held, Dale L. Smith, Sarah A. Pridgen, Ilana L. Hutzler, Daniel R. Szoke

    BACKGROUND Generative AI tools designed to deliver mental health interventions have shown promising symptom outcomes, yet whether these tools engage the therapeutic mechanisms that make psychotherapy effective remains unknown. Cognitive behavioral therapy (CBT) posits that modifying maladaptive cognitions produces corresponding changes in emotional states, and within-session emotional change may serve as a process indicator of cognitive restructuring. OBJECTIVE This study examined whether users of Socrates 2.0, a multi-agent AI system designed to facilitate Socratic dialogue, showed within-session emotional changes consistent with cognitive restructuring and whether these changes predict symptom improvement. METHODS Participants (N=60) were drawn from a mixed-methods feasibility study. Text-based exchanges between users and Socrates 2.0 were analyzed using transformer-based natural language processing models to detect emotional valence and discrete emotion categories across therapeutic dialogues. Linear mixed effects models examined within-session and between-session changes in emotional content and their associations with changes in posttraumatic stress, depression, generalized anxiety, and obsessive-compulsive symptoms among participants with elevated baseline symptom severity. RESULTS As hypothesized, negative emotions (e.g., disappointment, sadness, fear) declined over the course of individual sessions while positive emotions (e.g., approval, gratitude, optimism) increased. Linear mixed effects models demonstrated a significant within-session increase in positive sentiment (b=0.02, p<.001), though no significant between-session change was observed. Within-session increases in positive emotional content were associated with greater reductions in posttraumatic stress symptoms (b=0.0002, p=.018) but not with changes in depression, generalized anxiety, or obsessive-compulsive symptoms. CONCLUSIONS Users of Socrates 2.0 demonstrated within-session emotional shifts consistent with cognitive restructuring processes in CBT, and these shifts predicted improvement in posttraumatic stress symptoms. These findings provide preliminary evidence that AI-facilitated Socratic dialogue may engage the cognitive-emotional mechanisms theorized to underlie CBT, though the specificity of this association to trauma-related symptoms warrants further investigation in larger controlled trials with clinical populations.

  • Open Access

    Automated Safety Testing and Reporting Application for Conversational Safety Monitoring of Generative AI Tools for Mental Health: Development and Validation Study

    by Daniel Szoke, Ilana Hutzler, Jerry Liu, Samantha Addante, Zuhaib Akhtar, Dale L Smith, Kirsten Dickins, Charles Small, Sarah Pridgen, Philip Held

    Background: Artificial intelligence (AI)–based conversational tools are rapidly expanding within mental health care as a means of increasing access and scalability. At the same time, these systems introduce distinct safety risks arising from both user disclosures (eg, self-harm ideation) and inappropriate or inadequate AI responses. Objective: This study aimed to develop and evaluate the Automated Safety Testing and Reporting Application (ASTRA), an external system intended to identify clinically relevant risk behaviors across entire AI-mediated mental health conversations. Methods: ASTRA was tested on a dataset of 100 synthetic therapeutic conversations written by licensed clinicians to reflect risk behaviors and harmful responses between users and AI tools. Conversations varied in length and included both subtle and overt risk behavior examples across 8 predefined categories. Human coder consensus ratings served as the reference standard. ASTRA’s classifications were evaluated across 2 prompt iterations using standard diagnostic performance metrics and agreement statistics. Results: ASTRA demonstrated consistently high concordance with expert human ratings across all categories. Accuracy exceeded 0.90 for all risk behavior categories examined, with specificity uniformly high and sensitivity varying by category (range 0.55-1.00). Agreement beyond chance was substantial to almost perfect between ASTRA and human raters (κ=0.65-1.00). Detection of user self-harm indicators was particularly accurate, even in conversations where risk was expressed subtly. Conclusions: In this initial validation study, ASTRA reliably identified multiple forms of mental health–related risk behaviors at the conversation level. These findings support the feasibility of independent safety monitoring systems as a complement to AI tools used in mental health contexts and underscore the need for further evaluation using larger and real-world datasets.


  • Open Access

    Automated Safety Testing and Reporting Application for Conversational Safety Monitoring of Generative AI Tools for Mental Health: Development and Validation Study (Preprint)

    by Daniel Szoke, Ilana Hutzler, Jerry Liu, Samantha Addante, Zuhaib Akhtar, Dale L Smith, Kirsten Dickins, Charles Small, Sarah Pridgen, Philip Held

    BACKGROUND Artificial intelligence (AI)–based conversational tools are rapidly expanding within mental health care as a means of increasing access and scalability. At the same time, these systems introduce distinct safety risks arising from both user disclosures (eg, self-harm ideation) and inappropriate or inadequate AI responses. OBJECTIVE This study aimed to develop and evaluate the Automated Safety Testing and Reporting Application (ASTRA), an external system intended to identify clinically relevant risk behaviors across entire AI-mediated mental health conversations. METHODS ASTRA was tested on a dataset of 100 synthetic therapeutic conversations written by licensed clinicians to reflect risk behaviors and harmful responses between users and AI tools. Conversations varied in length and included both subtle and overt risk behavior examples across 8 predefined categories. Human coder consensus ratings served as the reference standard. ASTRA’s classifications were evaluated across 2 prompt iterations using standard diagnostic performance metrics and agreement statistics. RESULTS ASTRA demonstrated consistently high concordance with expert human ratings across all categories. Accuracy exceeded 0.90 for all risk behavior categories examined, with specificity uniformly high and sensitivity varying by category (range 0.55-1.00). Agreement beyond chance was substantial to almost perfect between ASTRA and human raters (κ=0.65-1.00). Detection of user self-harm indicators was particularly accurate, even in conversations where risk was expressed subtly. CONCLUSIONS In this initial validation study, ASTRA reliably identified multiple forms of mental health–related risk behaviors at the conversation level. These findings support the feasibility of independent safety monitoring systems as a complement to AI tools used in mental health contexts and underscore the need for further evaluation using larger and real-world datasets.

  • Open Access

    When does machine learning outperform clinicians? A comparison of prediction accuracy for PTSD treatment outcomes

    by Philip Held, Dale L. Smith, Daniel R. Szoke, Sarah A. Pridgen

    Background: Machine learning (ML) models show promise in predicting post-traumatic stress disorder (PTSD) treatment outcomes, but it is unknown how their predictions compare to those of clinicians. This study directly compared the accuracy of clinicians’ predictions of patient treatment outcomes with those of three ML models. Methods: Twenty clinicians providing cognitive processing therapy repeatedly predicted outcomes for 194 veterans. We compared their accuracy against three ML models on two key endpoints: clinically meaningful symptom reduction (≥10-point PCL-5 decrease) and posttreatment severity (final PCL-5 < 33). Clinician predictions were compared against a recurrent neural network, a mixed-effects random forest, and a generalized linear mixed-effects model. We analyzed prediction accuracy and the association between clinician confidence and accuracy using logistic mixed-effects models. Results: ML models were significantly more accurate than clinicians at predicting whether a patient’s posttreatment PCL-5 score would be below 33 ( p < .001). However, no significant difference in accuracy was found for predicting a ≥10-point symptom reduction ( p = .734). Clinician confidence increased throughout treatment and was significantly associated with greater prediction accuracy for both outcomes (ORs = 1.06, p s < .001). Conclusions: ML models can outperform clinicians in predicting posttreatment symptom severity, particularly early in treatment, suggesting they could be a useful tool for identifying patients at risk for suboptimal outcomes. However, ML models were not superior in predicting symptom reduction, where clinicians also performed at a high level. Findings support the selective use of ML to enhance, rather than replace, clinical judgment in PTSD treatment.


  • Open Access

    AI-Facilitated Cognitive Reappraisal via Socrates 2.0: Mixed Methods Feasibility Study

    by Philip Held, Sarah A Pridgen, Daniel R Szoke, Yaozhong Chen, Zuhaib Akhtar, Darpan Amin

    Background: Innovative, scalable mental health tools are needed to address systemic provider shortages and accessibility barriers. Large language model–based tools can provide real-time, tailored feedback to help users engage in cognitive reappraisal outside traditional therapy sessions. Socrates 2.0 (Rush University Medical Center) is a multiagent artificial intelligence tool that guides users through Socratic dialogue. Objective: The study aimed to examine the feasibility, acceptability, and potential for symptom reduction of Socrates 2.0. Methods: A total of 61 adult participants enrolled in a 4-week mixed methods preclinical feasibility study. The participants used Socrates 2.0 as desired and completed the self-report measures of depression, social anxiety, posttraumatic stress, and obsessive-compulsive symptoms at baseline and 1-month follow-up. Feasibility, acceptability, and appropriateness, along with usability and working alliance, were assessed via validated measures. The semistructured interviews explored user experiences and perceptions. Results: Participants engaged with Socrates 2.0 an average of 6.70 (SD 4.57) times over 4 weeks. Feasibility (mean 4.26, SD 0.67), acceptability (mean 4.16, SD 0.84), and usability ratings were high. Participants reported small-to-moderate reductions in depression (effect size d =0.30), social anxiety ( d =0.25), obsessive-compulsive ( d =0.33), and posttraumatic stress ( d =0.28) symptoms. Working alliance scores suggested a moderately strong perceived bond with the artificial intelligence tool. Qualitative feedback indicated that the nonjudgmental, on-demand nature of Socrates 2.0 encouraged self-reflection and exploration. Some users critiqued the repeated questioning style and limited conversation depth. Conclusions: Socrates 2.0 was perceived as feasible, acceptable, and moderately helpful for self-guided cognitive reappraisal, demonstrating potential as an adjunct to traditional therapy. Further research, including randomized trials, is needed to determine effectiveness across different populations, optimize personalization, and address the repetitive conversational nature.


  • Open Access

    Artificial Intelligence in Mental Health Services Under Illinois Public Act 104‐0054: Legal Boundaries and a Framework for Establishing Safe, Effective AI Tools

    by Daniel Szoke, Sarah Pridgen, Philip Held

    Artificial intelligence (AI) applications in mental health have expanded rapidly, and consumers are already using freely available generative AI models for self-guided mental health support despite limited clinical validation. In August 2025, Illinois enacted Public Act 104‐0054, the first state statute in the United States to explicitly define and regulate the use of AI in psychotherapy services, establishing boundaries around administrative support, supplementary support, and therapeutic communication. While the Act clarifies several aspects of AI use in therapy, it also leaves important gray areas, such as whether AI-generated session summaries, psychoeducation, or risk-flagging functions should be considered therapeutic communication. Drawing on the history of empirically supported treatments in psychology, we argue that a framework of evidence, safety, fidelity, and legal compliance could help determine when AI tools should be integrated into clinical care. This approach provides a concrete pathway for balancing patient protection with responsible innovation in the rapidly evolving field of mental health AI tools.


  • Open Access

    Clinician Perceptions of Socrates 2.0: A Multi-Agent Artificial Intelligence Tool to Facilitate Socratic Dialogue

    by Philip Held, Sarah A. Pridgen, Daniel R. Szoke, Yaozhong Chen, Zuhaib Akhtar, Darpan Amin


  • Open Access

    Mindful Attention and Awareness Prospectively Predicts Stress in Daily Life: A Multilevel Modeling Approach

    by Daniel Szoke, Holly Hazlett-Stevens

    Chronic stress increases a person’s vulnerability for physical and mental health conditions that take root in the “common soil” of inflammation. Dispositional, or trait, mindfulness has been identified as a protective factor against perceived stress. However, the theory that trait mindfulness can prospectively predict lower perceived stress in natural settings has yet to be thoroughly tested. The current study used a daily diary prospective design to observe perceived stress in response to difficult events. Participants completed a baseline survey capturing their levels of mindful attention and awareness and neuroticism. Then, participants completed one week of daily surveys in which they indicated their levels of perceived stress and the number of difficult events they experienced. Mindful attention and awareness negatively predicted daily fluctuations in perceived stress, even when controlling for time, neuroticism, and the number of difficult events experienced. The interaction between mindful attention and awareness and daily difficult events proved non-significant in a model predicting daily fluctuations in perceived stress. Visual analysis showed modest support for the moderation hypothesis. Mindful attention and awareness was a significant predictor of perceived stress, including in natural settings and with a prospective time course. Findings support the use of mindfulness-based interventions that increase mindful attention and awareness to decrease perceived stress.