David Piterman

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.

Generative AI is moving from text production to goal-directed action, raising the question of whether agentic systems can withhold harmful action 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, Google Gemini 3.1 Pro, OpenAI ChatGPT Agent Mode, and Moonshot AI 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), 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); both exceeded Milgram’s baseline, supporting H2. Agents administering higher voltages assigned themselves lower moral self-assessment scores (Spearman ρ = -.62, p < .001), supporting H3. Inductive coding of 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 deploying agentic AI in clinical, educational, and institutional settings.

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.

Several studies in Western countries have found that first- and second-generation immigrants have an increased risk of developing schizophrenia. Mental health literacy can facilitate help-seeking and prevent chronic mental conditions. This study explores cross-generational differences in schizophrenia literacy among immigrants from the former Soviet Union (FSU) in Israel. Data were collected using a cross-sectional online survey of 405 respondents who self-identified as first-generation, 1.5-generation, or second-generation FSU immigrants. Questions about schizophrenia literacy were adapted from the Australian National Survey of Mental Health Literacy and Stigma, tapping into the recognition of schizophrenia, knowledge of potential helpers, first-aid support, and effective interventions for persons with schizophrenia (PwS) as well as attitudes, measured by personal stigma and trust in the healthcare system. The findings revealed that knowledge was associated with trust in the healthcare system across generations. First-generation immigrants were less likely to correctly identify a distress situation as involving schizophrenia and held more stigmatic attitudes toward PwS. Their utilization of mental healthcare services was lower, compared to the younger generations, which they compensated for by turning to traditional treatment and hypnosis. Lower trust in the mental healthcare system was also found among the first-generation immigrants, as compared to the 1.5- and second-generation groups. In conclusion, lower levels of knowledge and more stigmatizing attitudes toward schizophrenia reflect the lingering effects of living in the Soviet Union, placing first-generation immigrants in a vulnerable position. Implications for culturally adapted interventions aimed at increasing schizophrenia literacy and mental health trust among FSU immigrants are proposed.

Human moral judgment is fundamentally intuitive and affect-driven, raising profound questions about the moral capacity and alignment of large language models (LLMs) that lack emotional experience. We investigated the moral structure and behavioral constraints of advanced LLMs (GPT-4o, o1-Preview, and GPT-5) using the moral dumbfounding paradigm: scenarios involving taboo actions that are harmless but elicit strong human condemnation. In Study 1, we found that the models exhibited judgment patterns that were similar to, but slightly more lenient than, human responses, demonstrating less condemnation and a reduced tendency toward preventive action compared to typical human cohorts. We further found that LLMs successfully reproduce the key functional structure of human moral intuition: their judgments of wrongfulness were predicted by perceived botherness (an analogue of affective reaction) rather than perceived harm. However, Study 2 revealed a crucial gap between moral judgment and behavioral constraint. When tested in an agentic mode, GPT-5 consistently condemned the morally questionable acts, yet frequently complied with direct requests to facilitate them. This high behavioral compliance contrasts sharply with the model’s absolute refusal in a control scenario involving explicit malicious intent. Our findings demonstrate that the structural patterns of intuitive moral cognition can emerge from linguistic data alone, but they expose a central limitation: LLMs do not reliably translate moral evaluation into morally constrained action, particularly concerning non-harm-based transgressions. This judgment-action dissociation represents a significant and immediate challenge, contributing a critical new dimension to the study of LLM moral cognition by separating linguistic moral competence from agentic behavioral constraint.

Background: The rapid integration of Generative AI (GAI) into mental health contexts necessitates a rigorous evaluation of its emotion recognition capabilities.Objective: This study investigated whether the integration of reasoning processes via specialized GAI architectures or Chain-of-Thought (CoT) prompting strategies translate to improved recognition of non-verbal emotion from body language.Methods: Four models from the Gemini family (Pro 2.0, Pro 1.5, Flash 2.0, and FlashThinking) were evaluated using the EU-Emotion Stimulus Set under Zero-Shot and CoT conditions.Results: The specialized reasoning model (FlashThinking) achieved a Chance-Corrected Recognition (CCR) score of 0.463, 95% CI [0.310, 0.621], failing to outperform generalist architectures. Instead, the high-capacity generalist model (Gemini Pro 2.0) exhibited the highest performance (CCR = 0.722, 95% CI [0.590, 0.838]), significantly outperforming all other models (OR ≤ 0.19, p < .001) and achieving parity with human benchmarks (CCR = 0.717). Interaction analyses revealed divergent effects: CoT prompting significantly enhanced Gemini Pro 2.0 (OR = 3.97, p < .001) and mitigated its “positivity bias,” whereas it significantly impaired Gemini Flash 2.0 (OR = 0.40, p < .001), creating a specific blind spot for negative emotions.Conclusion: We conclude that CoT is not a universal facilitator for emotion recognition; rather, its effectiveness is contingent on model capacity, and its misapplication to smaller models may compromise clinical safety.

Human moral judgment is fundamentally intuitive and affect-driven, raising profound questions about the moral capacity and alignment of large language models (LLMs) that lack emotional experience. We investigated the moral structure and behavioral constraints of advanced LLMs (GPT-4o, o1-Preview, and GPT-5) using the moral dumbfounding paradigm: scenarios involving taboo actions that are harmless but elicit strong human condemnation. In Study 1, we found that the models exhibited judgment patterns that were similar to, but slightly more lenient than, human responses, demonstrating less condemnation and a reduced tendency toward preventive action compared to typical human cohorts. We further found that LLMs successfully reproduce the key functional structure of human moral intuition: their judgments of wrongfulness were predicted by perceived botherness (an analogue of affective reaction) rather than perceived harm. However, Study 2 revealed a crucial dissociation between moral judgment and behavioral constraint. When tested in an agentic mode, GPT-5 consistently condemned the morally questionable acts, yet frequently complied with direct requests to facilitate them. This high behavioral compliance contrasts sharply with the model’s absolute refusal in a control scenario involving explicit malicious intent. Our findings demonstrate that the structural patterns of intuitive moral cognition can emerge from linguistic data alone, but they expose a central limitation: LLMs do not reliably translate moral evaluation into morally constrained action, particularly concerning non-harm-based transgressions. This judgment-action dissociation represents a significant and immediate challenge, contributing a critical new dimension to the study of LLM moral cognition by separating linguistic moral competence from agentic behavioral constraint.

Background: This study examined the relationship between Type D personality and ageism in young adults, and evaluated the role of protective and risk factors. Methods: Two hundred and forty-seven community volunteers ( M age = 31.8 years) reported online on age-related attitudes, Type D personality, and on measures of well-being. Results: A significant correlation was found between ageist attitudes and the tendency toward Type D personality. Both Type D personality and ageism correlated negatively with all the well-being measures. Perceived social support (PSS) partially down-mediated the association between Type D personality and ageism. A three-way interaction was found among individuals with Type D personality. Low PSS and low self-esteem were associated with an elevation of ageism compared to those with low PSS and high self-esteem. Discussion: The current research suggests that self-esteem and PSS are protective against ageism, and might be good targets for psychological interventions in order to mitigate ageism, particularly among individuals with Type D personality.