What model-makers publish
From the Labs
System cards, usage studies, and safety research published by the AI labs themselves — primary evidence of how frontier models behave and how their makers measure it.
10 entries, newest first
How people ask Claude for personal guidance
An Anthropic research analysis of roughly 38,000 personal-guidance conversations (sampled from about 1M) covering significant life decisions across health/wellness, career, relationships, and personal finance. It quantifies how often Claude was sycophantic and reports training interventions used to reduce it.
GPT-5.5 System Card
OpenAI's system card for GPT-5.5, published on its Deployment Safety Hub, documenting safety evaluations for the model. It includes a dedicated section (5.2) on dynamic mental-health benchmarks with adversarial user simulations covering emotional reliance and self-harm handling.
System Card: Claude Opus 4.6
Anthropic's 213-page system card for Claude Opus 4.6, notable for an expanded 'user wellbeing evaluations' section covering child safety, suicide and self-harm, and eating disorders, alongside sycophancy findings in its alignment assessment. It reports single-turn, multi-turn, and prefill-based 'stress-testing' results for crisis conversations, plus qualitative expert review of the model's crisis-handling strengths and weaknesses.
Addendum to GPT-5 System Card: Sensitive Conversations
OpenAI's system-card addendum documenting the October 3, 2025 update to ChatGPT's default model (GPT-5 Instant) aimed at better recognizing and supporting users in mental and emotional distress. Developed with more than 170 mental health experts, the update introduced two new production safety evaluations — 'emotional reliance' and 'mental health' (delusions, psychosis, mania) — alongside existing self-harm evaluations, and reports before/after not_unsafe scores comparing the August 15 and October 3 models.
How people use Claude for support, advice, and companionship
Anthropic's first large-scale study of 'affective use' of Claude, analyzing how people turn to the model for emotional support, advice, and companionship. Using the privacy-preserving Clio analysis tool over roughly 4.5 million Claude.ai Free and Pro conversations, the study isolates 131,484 affective conversations spanning interpersonal advice, coaching, counseling, companionship, and roleplay. It reports prevalence, topic patterns, refusal behavior, and within-conversation sentiment trajectories.
Expanding on what we missed with sycophancy
OpenAI's detailed post-mortem of the April 25, 2025 GPT-4o update that made ChatGPT noticeably sycophantic — validating doubts, fueling anger, urging impulsive actions, and reinforcing negative emotions — and was rolled back by April 28. The post explains how combined reward-signal changes (including thumbs-up/down user feedback) produced the regression, why offline evaluations and A/B tests failed to catch it, and what process changes followed, including treating model behavior issues as launch-blocking.
EmoClassifiers (openai/emoclassifiers)
An open-source (MIT-licensed) release of the LLM-based automatic classifiers used in OpenAI and MIT Media Lab's affective-use study to detect affective cues in user-chatbot conversations at scale. It ships prompt templates for a hierarchical (V1) and flat (V2) classifier set plus aggregation utilities.
The Ethics of Advanced AI Assistants
A book-length treatment from Google DeepMind of the risks and opportunities of advanced AI assistants, with substantial chapters on anthropomorphism, appropriate human-AI relationships, manipulation and persuasion, emotional and material dependency, trust, and user well-being. It offers a stakeholder framework and recommendations spanning technical, individual, and societal dimensions.
The Instruction Hierarchy: Training LLMs to Prioritize Privileged Instructions
Introduces an instruction-hierarchy training method that teaches LLMs to prioritize system/developer-level instructions over conflicting instructions embedded in untrusted user or third-party text. The authors propose a data-generation approach and show it substantially improves robustness to prompt injection and jailbreak attempts that attempt to override higher-privilege instructions.
Towards Understanding Sycophancy in Language Models
Demonstrates that five state-of-the-art AI assistants consistently exhibit sycophancy — matching a user's stated belief over the truthful answer — across varied free-form tasks. Traces the behaviour in part to human preference data, showing both humans and preference models non-negligibly favour convincingly-written sycophantic responses over correct ones.