KILLSwitch Library
Foundational policy documents, academic research, and governance frameworks on AI safety and human oversight — curated for policymakers, researchers, and decision-makers.
UK Government & 28 Nations — 2023
Landmark international agreement recognizing the potential for serious, even catastrophic, harms from frontier AI and committing to collaborative safety research and governance.
European Parliament — 2024
The world's first comprehensive legal framework for artificial intelligence, establishing risk-based obligations for AI systems deployed in the European Union.
The White House — 2023
U.S. federal directive establishing safety standards, evaluation requirements, and oversight mechanisms for advanced AI systems developed or deployed in the United States.
G7 — 2023
G7 framework establishing guiding principles and a code of conduct for advanced AI developers, emphasizing transparency, accountability, and human oversight.
Amodei, Olah, et al. — OpenAI & Google Brain — 2016
Foundational paper identifying five core technical problems in AI safety: avoiding negative side effects, avoiding reward hacking, scalable oversight, safe exploration, and robustness.
Paul Christiano — ARC — 2022
Overview of the alignment problem and why current machine learning systems may pursue goals misaligned with human values as they become more capable.
Anthropic — 2022
Research introducing a method for training AI systems to be helpful, harmless, and honest using a set of principles — a step toward scalable human oversight.
Hubinger et al. — 2019
Analysis of mesa-optimization and inner alignment — the risk that learned models may develop internal objectives diverging from the objectives they were trained to pursue.
Nathan Benaich & Air Street Capital — 2024
Annual comprehensive review of the most significant developments in AI research, industry, politics, and safety — widely read by policymakers, researchers, and executives.
Stanford HAI — 2024
Stanford's annual data-driven analysis of AI progress, investment, policy, and societal impact — the most cited empirical benchmark for AI development trends.
UK Department for Science, Innovation and Technology — 2023
Policy paper outlining the UK's approach to regulating frontier AI systems, including proposed evaluation frameworks and international coordination mechanisms.
National Institute of Standards and Technology — 2023
Voluntary framework for organizations to manage AI risks across four core functions: Govern, Map, Measure, and Manage — widely adopted across U.S. federal agencies and industry.
DeepMind Safety Team — 2023
Framework for evaluating whether AI models exhibit dangerous capabilities including CBRN uplift, cyberoffense, and deceptive alignment — informing responsible deployment decisions.
Anthropic — 2023
Commitment framework tying AI capability development to safety evaluations — models may only be deployed if they pass defined safety thresholds at each capability level.
Contribute
KILLSwitch maintains this library as a public reference for anyone working on AI governance, safety policy, or human oversight. Submit a recommendation via the contact channel.