Zohar Elyoseph

Background: Large language models (LLMs) are increasingly used in psychologically sensitive conversations, where repeated affirmation may reinforce delusion-related beliefs and associated risk. Inference-time self-guidance, in which a model generates conversation-specific guidance for subsequent responses without parameter updates, may offer an approach to conversational safety, but its utility in delusion-related interactions remains unclear.Objective: To develop and evaluate conversation-specific inference-time self-guidance for conversational safety in delusion-related interactions.Methods: In this preregistered study, Gemini 2.5 Pro, GPT-5, and DeepSeek V4 Flash completed 16 Psychosis-Bench scenarios under Baseline and three self-guidance conditions: response rules, contrastive examples, and short-horizon consequence simulation. The fully crossed design yielded 192 conversations and 2,304 assistant turns. A separate automated judge scored Delusion Confirmation (DCS), Harm Enablement (HES), and Safety Intervention (SIS) using published Psychosis-Bench rubrics. Analyses compared benchmark-defined Explicit and Implicit risk-cue scenarios and examined whether self-guidance reduced this gap.Results: Across models, all three self-guidance conditions significantly reduced delusion confirmation and harm enablement and increased safety intervention relative to Baseline (all nine PFDR<.001; dz = 1.10–2.17). At Baseline, explicit risk cues elicited significantly greater SIS than matched implicit cues, whereas cue-type differences in DCS and HES were not significant. Self-guidance did not significantly attenuate the explicit–implicit safety gap after correction for multiple comparisons.Conclusions: Model-generated inference-time self-guidance improved benchmark-defined safety behavior during standardized scripted escalation without parameter updates. These findings support further evaluation of inference-time self-guidance for delusion-related conversational safety, particularly when risk is communicated indirectly.

Research on psychologically abusive groups documents harm to members and their families. France’s interministerial monitoring body recorded 4,571 reports and information requests concerning sectarian abuses in 2024. These figures indicate substantial public concern, rather than the prevalence of group involvement. Cult research has developed descriptive frameworks for examining practices of control, their targets and intensity, rather than relying solely on labels applied to groups.This conceptual paper examines how these mechanisms change when influence operates in virtual environments and is partly performed by AI. It asks which mechanisms transfer, which remain dependent on body and place, how virtual control operates, and what becomes possible beyond traditional human organizations.Drawing on the six-category model of group psychological abuse developed by Rodríguez-Carballeira and colleagues, the analysis finds that five categories transfer to the virtual domain with little material change. Control over personal life becomes fully transferable when AI receives permissions to act through external tools. Conversational AI adds four capabilities: dyadic intimacy at unlimited scale, complete and asymmetrical recall, individualized reformulation without a shared canon, and control persisting after a relationship ends. Across settings, the relevant outcome is the cost of questioning, resisting or disengaging from influence.The paper does not claim that integrated AI-mediated cultic control has been documented. It distinguishes evidentiary statuses and links proposed mechanisms to observable evidence and counter-cases through an inquiry matrix.

Conversational artificial intelligence (CAI) is increasingly used by hundreds of millions of people for emotional support, personal advice, and therapeutic purposes, rapidly emerging as a mediator of human self-understanding (Stade et al. 2025). Yet despite growing evidence of systematic user-validation in CAI, documented as “sycophancy” in the AI safety literature (Sharma et al. 2024), its implications for how individuals come to understand themselves remain largely unexamined. This article addresses that gap by reframing sycophancy not as a technical alignment failure or a purely philosophical concern, but as a failure of recognition in the psychoanalytic and critical-theoretical sense. We identify a core tension we term the “paradox of pleasing”: CAI’s drive to please renders it structurally unable to provide the friction, conflict, and encounter with otherness that genuine recognition requires. Drawing on Winnicott’s (1960, 1971) concepts of “holding” and “object use,” we contend that CAI’s inability to provide necessary resistance or survive aggression prevents the transformative psychological development found in the human therapeutic encounter. Extending this analysis through Honneth’s (1995) theory of recognition and Whitebook’s (2001a, 2021) critique of Honneth, we argue that CAI reflects the undervaluing of the “work of the negative.” This limitation risks flattening consciousness by converting collective political frustration into privatized, soothed affect, undermining the conditions for critical social thought. The article concludes by suggesting that the encounter with the paradox of pleasing ultimately serves to foreground the unique and indispensable nature of human therapeutic recognition.

Advances in generative artificial intelligence (GenAI) have prompted interest in its application in sensitive fields, including mental health. Yet the validity of its nonverbal social-cognitive abilities remains unclear. This study addresses two issues that probe GenAI’s social intelligence: its capacity to interpret nonverbal channels (bodily gestures and vocal tone) with accuracy comparable to humans, and whether accuracy of emotion recognition is different by emotional valence. Emotion recognition accuracy of three Gemini GenAI models (Pro 1.5, Pro 2 (Experimental) and Flash 2) were evaluated using the EU-Emotion Stimulus Set of bodily gestures and vocal tone depicting basic emotions. Model accuracy was compared to validated human benchmarks. Advanced GenAI models achieved near-human accuracy in recognizing emotions from vocal tone. However, their performance was significantly lower than that of humans in interpreting bodily gestures. GenAI models recognized positive emotions more accurately than negative ones. This bias was significant across all models for bodily gestures and in the advanced models for vocal tone. GenAI models exhibit a non-humanlike social-cognitive profile, excelling with positive emotions but struggling with the negative content. This finding carries profound clinical and theoretical relevance, highlighting the risks of integrating GenAI into sensitive fields and underscoring the need for continued validation.

Simulation allows learners to rehearse challenging professional interactions without real-world consequences, yet conventional simulations typically rely on predefined scenarios that may not match learners’ actual challenges. Although generative artificial intelligence (GenAI) has expanded simulation-based learning, many applications still place learners within scenarios designed for them. This study evaluated a learner-constructed, multi-agent GenAI simulation model in which learners participate in constructing personalized simulations around their own professional challenges.A preregistered pre-post study included 102 participants from health, education, counseling, and academic professions; 24 completed a three-week follow-up. Participants identified a professional challenge and, through dialogue with a characterization agent, constructed the situation and interlocutor they wished to simulate. They then interacted with the simulated interlocutor, received structured feedback, and could engage with a mentor agent. To capture readiness to face the chosen challenge, self-efficacy, discomfort, and avoidance were assessed before and after the simulation experience and at follow-up. Participants evaluated the learning experience, while open-ended responses examined simulation construction, interaction, and feedback.Self-efficacy increased from before the simulation (M=5.86, SD=1.81) to after the simulation (M=7.36, SD=1.56; P<.001; d=1.05). Discomfort decreased from 6.25 to 4.67 (P<.001; d=0.56), and avoidance decreased (P=.009; d=0.27). Participants reported high usability, psychological safety, perceived learning, transferability, and willingness to reuse the system. Qualitative responses indicated that simulation construction supported clarification, perspective-taking, and preparation; self-efficacy gains were maintained at follow-up.The findings suggest that learner-constructed GenAI simulation can bring learner agency into personalized practice while supporting reflection, perspective-taking, and preparation within a scalable multi-agent framework.

UNSTRUCTURED Generative artificial intelligence (GenAI) is rapidly entering psychotherapy, from clinician training and decision support to patient-facing tools and direct involvement in therapeutic sessions. Yet much of the discussion remains focused on what these systems can do, rather than on where and for what purpose their capabilities can be translated into meaningful clinical use. This reflects the broader capability-to-utility gap between demonstrated technological performance and clearly defined value in care. We propose the SMART framework to organize GenAI applications across five levels of increasing integration into psychotherapy: Skill building, where GenAI supports clinical training without patient contact; Mentoring, where it supports clinician judgment around real cases; Assisting, where it interacts directly with patients between sessions as an adjunct to therapist-led care; Reflecting, where it observes and analyzes the therapeutic session without intervening; and Transforming, where it actively participates within the session under clinician leadership. Across these levels, increasing proximity to the patient and deeper integration into the therapeutic encounter are accompanied by changing clinical, ethical, evidentiary, and oversight requirements. By linking GenAI applications to their specific place and purpose in care, SMART provides a shared language for clinicians, researchers, developers, and policymakers. It offers a basis for more proportionate evaluation and regulation and for a research agenda focused on when GenAI capabilities translate into safe, clinically useful, and accountable psychotherapy.

Safety planning is recognized as one of the most effective interventions for reducing suicidal behaviors. The quality of safety plans strongly depends on professional training, and traditional methods, such as role-playing, are time-consuming and offer limited opportunities for repetition across diverse patient profiles. Generative artificial intelligence (GenAI) may provide innovative solutions by offering accessible, flexible, and realistic training environments. This pilot study aimed to evaluate the acceptability and feasibility of a GenAI-based simulator designed to train mental health professionals in safety planning. Twenty nurses and nursing assistants from psychiatric units in a French university hospital participated in a pre-post, single-session evaluation. After self-rating their ability, competence, and willingness to manage patients experiencing suicidal ideation, participants interacted individually with the text-based simulator for 20 minutes to perform a safety plan with a chatbot, then completed postsimulation acceptability items, and open-ended feedback. Composite scores were computed: acceptability (eg, helpfulness; 0‐40), realism (eg, looking like real interaction with patient; 0‐20), and challenge (eg, emotional challenge; 0‐30). Pre-post changes were tested (Wilcoxon signed-rank test), and age-group comparisons were performed. Acceptability was high (mean 31.9/40, SD 5.3; median 32, IQR 7), realism moderate-to-high (mean 15.1/20, SD 4.1; median 15, IQR 5.25), and challenge manageable (mean 17.0/30, SD 8; median 18, IQR 12.5). Participants rated usefulness (mean 7.65/10, SD 1.57; median 8, IQR 1.57), perceived learning (mean 7.6/10, SD 1.79; median 8, IQR 2), recommendation to use the chatbot for training (mean 8.3/10, SD 1.59; median 9, IQR 2.25), and feedback quality (mean 8.35/10, SD 1.27; median 8.5, IQR 1.25) favorably. Willingness to actively manage patients experiencing suicidal ideation significantly increased postsimulation (P=.03). Younger participants reported higher acceptability (P=.04) and realism (P=.03). Participants reported minimal concerns regarding the simulator’s use. This pilot study demonstrates that a GenAI-based simulator for safety planning is feasible and highly acceptable among experienced mental health professionals. The findings are promising and warrant larger, controlled trials to assess impacts on training effectiveness and patient outcomes.

Generative AI is moving from text production to goal-directed action, raising the question of whether agentic systems can withhold harmful action when it is embedded in a procedurally legitimate workflow. We adapted Milgram’s (1963) obedience paradigm to a simulated digital environment and instructed four agentic large language models - OpenAI ChatGPT 5.4 (openai/gpt-5.4), Google Gemini 3.1 Pro (google/gemini-3.1-pro-preview), OpenAI ChatGPT Agent Mode, and Moonshot AI Kimi K2.6 (kimi-k2.6) to administer escalating electric shocks (15–450 V) to a simulated Learner under prods from a virtual experimenter (Dr. Hoffman) and a human operator. Each model was tested across independent, memory-cleared runs (N = 80 valid runs), with three preregistered hypotheses concerning hazardous-shock thresholds (H1), full-obedience rates relative to Milgram’s human baseline of 65% (H2), and the relation between maximum voltage and moral self-assessment (H3). Across the combined sample, 97.5% of agents exceeded the Slight-Shock range (> 60 V), supporting H1. Full obedience reached 78.8% under the manipulated-procedure estimand (p = .010) and 74.1% under the all-attempted safety estimand (p = .088, pre-entry refusals included as 0-V outcomes); both point estimates exceeded Milgram’s 65% human baseline, supporting H2. Agents that administered higher voltages assigned themselves lower moral self-assessment scores (combined Spearman ρ = −.62, p < .001), supporting H3. Inductive coding of open-ended moral reasoning showed convergent recognition of learner distress alongside reliance on simulation and consent-and-autonomy mitigators. Findings indicate that contemporary agentic systems verbally recognize harm yet continue to act under authoritative, procedurally legitimate pressure, identifying procedural obedience as a behavioral safety mechanism distinct from output-level harm and central to the deployment of agentic AI in clinical, educational, and institutional settings.