Ziv Ben-Zion

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.

After the October 7, 2023, attack and the prolonged war that followed, research on resilience, recovery, and adaptation in Israel expanded rapidly. We conducted a prospectively registered scoping review of 147 peer-reviewed quantitative publications drawn from 125 independent samples to examine how resilience and adaptation were conceptualized, which resources were associated with favorable outcomes, and what evidence supported protective mechanisms. No publication operationalized resilience relative to the adversity participants experienced. Instead, resilience was predominantly measured as a self-reported capacity, and adaptation was most often inferred from low symptom levels rather than preserved functioning or broader positive adjustment. Personal resilience and individual psychological resources showed comparatively strong associations with favorable outcomes. However, the reliance on self-report resilience measures limits the interpretation of these associations as effects of distinct resources. Belonging showed a more consistently favorable pattern in a smaller literature, whereas findings for social support, family resources, and collective resources were less consistent. Formal tests of adversity-buffering effects were uncommon, predominantly cross-sectional, and indicated context-specific rather than general protection; prospective evidence regarding mechanisms was similarly sparse. No randomized trial tested a resilience-promoting intervention against a non-active comparator. The literature, therefore, identifies numerous correlates of favorable outcomes but provides substantially less evidence about when these resources constitute resilience or alter adaptation in the face of adversity. Future research should prioritize exposure-sensitive definitions, multidimensional outcomes, longitudinal designs, and temporally ordered tests of mechanisms.

Background: The October 7, 2023, attack in Israel was followed by a prolonged war marked by repeated threat, displacement, reserve mobilization, and hostage-related uncertainty. This context generated a rapidly expanding yet heterogeneous empirical literature on psychological distress. This scoping review mapped the scope, temporal patterning, distribution, and evidence gaps of peer-reviewed publications among individuals residing in Israel after October 7, foregrounding prolonged war as the central context.Methods: This prospectively registered review followed PRISMA-ScR guidance. Searches were conducted in PubMed/MEDLINE, APA PsycINFO, Scopus, and Web of Science Core Collection for publications from October 7, 2023, to April 2026. Eligible publications reported quantitative findings on mental health symptoms, functioning, risk or protective factors, or service use. Findings were synthesized narratively by publication characteristics, temporal patterns, exposure contexts, population groups, and evidence gaps.Findings: Searches yielded 11,285 records; screening identified 240 eligible publications, of which 205 reported distress-related outcomes and were retained here. Cross-sectional publications predominated (120/205; 58.5%), followed by longitudinal or repeated-assessment publications (80/205; 39.0%); intervention or experimental publications were rare (5/205; 2.4%). Findings documented acute increases in PTSD/PTSS, depression, anxiety, acute stress, sleep disturbance, and general distress, followed by partial but uneven symptom reduction during the first year. Evidence beyond the first year was limited but suggested persistent or possible cumulative burden under prolonged-war conditions. Burden was concentrated in groups marked by direct trauma, displacement, bereavement or hostage-related uncertainty, military or caregiving burden, minority or age-related vulnerability, and moral injury or betrayal.Interpretation: Post-October 7 mental health burden is best understood not as a bounded post-event aftermath, but as distress evolving within a prolonged crisis. Future research should move beyond early symptom documentation toward longitudinal, functional, service-oriented, and intervention-informed evidence.

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.