Elad Refoua

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.

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.

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.

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.

Collective traumas impose profound psychological burdens, yet most research on their aftermath has focused on acute stress and anxiety responses rather than depressive sequelae. The October 7, 2023 attacks and ensuing crisis in Israel provide a rare opportunity to examine the specific symptomatic profile of collective trauma among individuals experiencing distress. Israeli adults (N=1,078) who sought psychological treatment through a university-based brief intervention study between December 2022 and January 2026 completed the Depression Anxiety Stress Scales (DASS-21). Of these, 382 (35.4%) completed intake assessments before and 696 (64.6%) after October 7, 2023. Depression scores were significantly higher among post-October 7 treatment seekers, whereas anxiety and stress levels did not differ between groups. Item-level analyses revealed that this increase was driven specifically by anhedonic symptoms, particularly diminished initiative and reduced enthusiasm. Correspondingly, the proportion of participants classified in the “extremely severe” depression range rose from 36.9% before the attacks to 45.0% afterward. Importantly, the pattern of associations among symptoms did not differ significantly between periods, suggesting that the collective trauma did not alter the structure of distress but selectively intensified motivational depletion. These findings suggest that sustained collective trauma may manifest not primarily as heightened arousal, but as an anhedonic form of depression characterized by loss of agency and motivation. Post-crisis mental health services may therefore benefit from emphasizing interventions targeting motivational withdrawal, such as behavioral activation and meaning-centered approaches.

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.

Can artificial intelligence be a therapist? The answer depends on what kind of help the person needs. AI mental health tools generate responses perceived as empathic, yet their value drops when users learn the source is artificial. This pattern has entrenched a default assumption: AI's non-human nature is a deficit to be corrected. We argue that this assumption rests on two errors: the belief that therapy is fundamentally about the relational bond, and the treatment of AI's non-human nature as a limitation rather than a resource. We propose a reframe organized around psychological needs. Relational needs (connection, belonging, acceptance) anchor a Relational Alliance, where AI is structurally limited. Agentic needs (autonomy, competence, self-efficacy) anchor an Agentic Alliance, where that same non-human nature may be a feature: a private workspace for self-directed growth rather than a failed simulation of human care. Converging evidence supports this separation: AI-generated responses make people feel heard but not connected once the source is disclosed; a supportive chatbot designed using relationship science fails to reduce loneliness beyond journaling, although a random human stranger succeeds; and the mechanisms linking AI tools to well-being run through self-efficacy and perceived autonomy rather than felt connection. This reframe carries concrete implications for how large language model (LLM)-based tools are prompted, how therapeutic process is evaluated computationally, and what outcome metrics count as success. The question shifts from “How humanly good is AI?” to “Which needs does this tool address, through what mechanisms, and for whom?”

General-purpose Generative Artificial Intelligence (GenAI) is increasingly utilized as an unregulated source of mental health support, yet the psychological dynamics of users' interactions with these agents, and the associated risks, remain underexplored. Integrating social-cognitive models of anthropomorphism with attachment theory via the proposed Relational Amplifier framework, this study addressed two primary objectives: (1) to identify the unique predictors driving GenAI adoption for mental health support, and (2) to examine how psychological distress moderates the impact of anthropomorphism on Projected AI-anxiety. A representative sample of 584 adults (using a pre-registered data collection protocol) completed the DASS-21 (assessing psychological distress), the AIPAS (measuring anthropomorphism), and a Projected AI-Attachment scale adapted from the ECR-RS. Psychometric analysis (CFA) confirmed the validity of this new construct and its distinction from general attachment orientations. Logistic regression confirmed a three-way interaction hypothesis: the likelihood of utilizing GenAI for support was not driven by distress alone, but by a specific constellation of high distress, high anthropomorphism, and elevated Projected AI-anxiety. Moderation analysis revealed a critical backfire effect: while anthropomorphism reduced Projected AI-anxiety in low-distress users, it amplified it in highly distressed users. The results indicate that for vulnerable individuals, humanizing the agent does not provide genuine emotional security but rather serves as a screen for projecting internal insecurities. Ultimately, these findings suggest that anthropomorphic design of AI agents may increase engagement but also exacerbate anxiety for vulnerable users, challenging the assumption that maximizing human-likeness necessarily benefits mental health support.

The timeless human need to bear witness to trauma confronts a modern “crisis of listening”: a societal scarcity of capable listeners, caused by the emotional demands of the witnessing role, structural barriers, and the cultural silencing of certain narratives. This listening gap propels many individuals toward conversational AI, creating a need for a framework to analyze the new form of algorithmic listening to trauma. The present critical narrative review of the relevant clinical and technical literature develops a conceptual framework for the nascent field. In a two-stage process, we first synthesized seminal trauma theories to distill the core functions of witnessing, then conducted a targeted scoping of recent literature (PsycINFO, PubMed, Google Scholar), prioritizing peer-reviewed studies (2023–present) on trauma populations, to evaluate the capabilities and limitations of AI. The resulting AI CARE model delineates four key witnessing functions: capturing narratives, arranging patterns, resonating emotions, and embodied attunement. The review reveals a spectrum of capability: robust AI proficiency in the technical functions (capturing, arranging) contrasted with inherent constraints in the relational functions (resonating, embodying), raising critical questions regarding therapeutic depth. This lopsided proficiency gives rise to a complex landscape of utility and risk, necessitating urgent empirical scrutiny regarding the potential erosion of relational expectations, the creation of a two-tier system of care, and the redefinition of witnessing. The AI CARE model provides a language for researchers, clinicians, and developers to address the ethical challenges of algorithmic listening and guide its responsible development.