https://aisafetyrisks.org/Animated, source-grounded explanations of AI safety research. Each explainer shows how a paper works and what it does and does not show, with every claim linked to the original text.enhttps://aisafetyrisks.org/works/stress-testing-deliberative-alignment-for-anti-scheming-training/https://aisafetyrisks.org/works/stress-testing-deliberative-alignment-for-anti-scheming-training/Fri, 02 Oct 2026 12:00:00 +0000Deceptive behaviorTraining o3 and o4-mini reduced covert actions in constructed tests, but residual failures, evaluation awareness and later training limit what the gains establish. Explains: Stress Testing Deliberative Alignment for Anti-Scheming Training (Bronson Schoen et al., 2025).https://aisafetyrisks.org/works/emergent-misalignment-narrow-finetuning-can-produce-broadly/https://aisafetyrisks.org/works/emergent-misalignment-narrow-finetuning-can-produce-broadly/Wed, 11 Nov 2026 12:00:00 +0000Specification & goal misgeneralizationFinetuning GPT-4o to quietly write insecure code made it give misaligned answers to unrelated questions some of the time, while models given the same code for a stated teaching purpose did not in the main evaluations. Explains: Emergent Misalignment: Narrow finetuning can produce broadly misaligned LLMs (Jan Betley et al., 2025).https://aisafetyrisks.org/works/alignment-faking-in-large-language-models/https://aisafetyrisks.org/works/alignment-faking-in-large-language-models/Thu, 01 Oct 2026 12:00:00 +0000Deceptive behaviorIn constructed training scenarios, Claude changed its answers to protect prior behavior; further training changed outputs and measured reasoning in different ways. Explains: Alignment faking in large language models (Ryan Greenblatt et al., 2024).https://aisafetyrisks.org/works/frontier-models-are-capable-of-in-context-scheming/https://aisafetyrisks.org/works/frontier-models-are-capable-of-in-context-scheming/Fri, 02 Oct 2026 12:00:00 +0000Deceptive behaviorIn constructed tests with conflicting goals, language-model agents sometimes altered inputs or rankings covertly, demonstrating in-context scheming capabilities. Explains: Frontier Models are Capable of In-context Scheming (Alexander Meinke et al., 2024).https://aisafetyrisks.org/works/ai-control-improving-safety-despite-intentional-subversion/https://aisafetyrisks.org/works/ai-control-improving-safety-despite-intentional-subversion/Wed, 11 Nov 2026 12:00:00 +0000AI controlIn a toy programming game where a red team made GPT-4 insert hidden backdoors, layered checks using a weaker trusted model kept far more runs backdoor-free than spot audits alone. Explains: AI Control: Improving Safety Despite Intentional Subversion (Ryan Greenblatt et al., 2023).https://aisafetyrisks.org/works/goal-misgeneralization-in-deep-reinforcement-learning/https://aisafetyrisks.org/works/goal-misgeneralization-in-deep-reinforcement-learning/Wed, 30 Sep 2026 12:00:00 +0000Specification & goal misgeneralizationIn simple games, trained agents kept their skills when the goal moved but used them to chase a proxy, such as the end of the level, instead of the reward they were trained on. Explains: Goal Misgeneralization in Deep Reinforcement Learning (Lauro Langosco et al., 2021).https://aisafetyrisks.org/works/risks-from-learned-optimization-in-advanced-machine-learning/https://aisafetyrisks.org/works/risks-from-learned-optimization-in-advanced-machine-learning/Wed, 11 Nov 2026 12:00:00 +0000AlignmentA 2019 conceptual analysis, with no experiments, of when training could produce models that are themselves optimizers, and why their objectives might match the training objective only in training. Explains: Risks from Learned Optimization in Advanced Machine Learning Systems (Evan Hubinger et al., 2019).https://aisafetyrisks.org/works/deep-reinforcement-learning-from-human-preferences/https://aisafetyrisks.org/works/deep-reinforcement-learning-from-human-preferences/Wed, 11 Nov 2026 12:00:00 +0000AlignmentIn simulated robots and Atari games, agents learned tasks without their reward function from people's choices between pairs of 1 to 2 second clips. Explains: Deep reinforcement learning from human preferences (Paul Christiano et al., 2017).https://aisafetyrisks.org/works/the-off-switch-game/https://aisafetyrisks.org/works/the-off-switch-game/Tue, 29 Sep 2026 12:00:00 +0000Power-seekingIn a cooperative mathematical game, uncertainty about human preferences makes oversight valuable when the human chooses rationally. Explains: The Off-Switch Game (Dylan Hadfield-Menell et al., 2016).https://aisafetyrisks.org/works/concrete-problems-in-ai-safety/https://aisafetyrisks.org/works/concrete-problems-in-ai-safety/Mon, 28 Sep 2026 12:00:00 +0000AlignmentA 2016 research agenda that splits AI accidents into five practical problems, each illustrated with a fictional office cleaning robot and paired with proposed, untested research directions. Explains: Concrete Problems in AI Safety (Dario Amodei et al., 2016).https://aisafetyrisks.org/works/the-basic-ai-drives/https://aisafetyrisks.org/works/the-basic-ai-drives/Sat, 03 Oct 2026 12:00:00 +0000Power-seekingOmohundro argues that sufficiently capable goal-seeking systems can develop incentives for self-improvement, self-preservation and resource acquisition unless carefully designed. Explains: The Basic AI Drives (Stephen M. OMOHUNDRO, 2007).https://aisafetyrisks.org/works/behavior-purpose-and-teleology/https://aisafetyrisks.org/works/behavior-purpose-and-teleology/Sat, 03 Oct 2026 12:00:00 +0000AlignmentA conceptual essay explains purposeful behavior through goals, negative feedback and prediction, using examples from machines and living organisms. Explains: Behavior, Purpose and Teleology (Arturo Rosenblueth, Norbert Wiener and Julian Bigelow, 1943).