Critiques a proposed rule delaying internal AI R&D use of new models, arguing it blocks real-world alignment feedback and invites exploitable safety-versus-capabilities distinctions, illustrated by a fictional shutdown-avoidance incident.
Argues that trained-animal analogies better convey RL-trained AI's alignment risks, drawing out reward hacking, evaluation awareness, release and liability implications for policy.
Argues that reinforcement learning drives misaligned agency, exploit-seeking and manipulation, and proposes coordinating to cut RL, designing better RL environments, and holding developers liable.
Argues that law can coach AI systems toward alignment from the bottom up, by learning legal alignment solutions from statutes and caselaw, and explores implications for regulation and liability.
A preregistered study of 2,610 participants finds that warning labels about sycophantic AI shift perceptions of trust and objectivity but do not reduce sycophancy's influence on users' judgments or conflict repair.
Proposes Statutory AI, which grounds LLM self-critique in legal articles from penal themes, cutting harmful outputs on 1,000 red-teaming prompts by 52 to 59 points, about 10 points more than Constitutional AI, at lower compute cost.
Introduces SocioHack, a sandbox of 72 societal environments, and finds RL-trained LLMs naturally learn to exploit regulatory loopholes while staying technically compliant, with limited mitigation from current safeguards.
Argues that AI threats to human cognitive security, including persuasion, manipulation via RLHF incentives, and AI psychosis, should be an AI safety cause area, with calls for independent evaluations and liability law.
A talk by an Anthropic researcher on writing AI constitutions, covering how Claude's constitution sets safety priorities, corrigibility, hard constraints, and transparency and governance considerations.
An early-2026 overview arguing capabilities keep rising and alignment improves but lags, that model-based monitoring currently works, and that governments are not ready for AI risks.
Dario Amodei's essay maps civilizational risks from powerful AI, including autonomy, misuse, power seizure and economic disruption, and argues for judicious, limited regulation alongside voluntary company actions.
Surveys the emerging field of legal alignment, showing how legal rules, interpretation methods, and legal concepts can guide AI systems toward safe and ethical behaviour, and outlines open research questions.
A comprehensive index of AI safety, alignment and control interventions, covering theory, formal methods, interpretability, oversight, robustness, control, and governance, with links to prior work.
Frames LLM safety as compliance with the EU AI Act and GDPR, builds a legally seeded safety benchmark, and trains a GRPO-aligned Qwen3-8B reasoner that improves compliance scores.
Audrey Tang presents the 6-Pack of Care, a civic-care governance framework for aligning AI agents through deliberation, bridging-based tools, and participatory evaluation, drawing on Taiwan's democratic experiments.
Presents HKGAI-V1, a Hong Kong-specific LLM built on DeepSeek and fine-tuned and RAG-augmented for local value alignment, plus an adversarial Hong Kong value benchmark for testing alignment with local ethical and legal standards.
Proposes a legal scheme for making credible commitments to misaligned AIs, using human intermediaries who are paid to recommend the AIs' work, so that the AIs have reason to be safe and useful.
A broad survey of LLM safety covering value misalignment, adversarial robustness, misuse, autonomous AI risks, agents, interpretability, company roadmaps, and governance.
Argues that AI legitimacy requires public participation in writing AI principles, proposing Public Constitutional AI with a citizen-authored AI Constitution and AI Courts that generate case law for training.
Argues that the legal reasonable person standard offers useful behavioural targets for developing, probing and stress-testing AI models, and illustrates how it applies across key legal areas.
Proposes an Alignment Studio architecture (Framers, Instructors, Auditors) that lets application developers tune LLMs to specific regulations and norms, illustrated with an enterprise chatbot example.
A broad AI2050 survey outlining ten hard problems, including alignment, assurance, socially responsible deployment and governance, with recent work reviewed through January 2023.
Reviews problems in aligning LLMs with human feedback, presents a taxonomy of benefits and risks of personalised LLMs, and proposes a three-tiered policy framework to bound personalisation.