Regularly updated resource list to get up to speed on the main ideas in the field of long-term AI safety. The resources are selected for relevance and/or brevity, and the list is not meant to be comprehensive.
- Technical overviews
- Technical work
- Collections of technical works
For a popular audience:
Sutskever and Amodei (2017). Wall Street Journal: Protecting Against AI’s Existential Threat
Cade Metz (2017). New York Times: Teaching A.I. Systems to Behave Themselves.
OpenPhil (2015). Potential risks from advanced artificial intelligence. An overview of AI risks and timelines, possible interventions, and current actors in this space.
Robert Miles. Computerphile videos on AI safety. Introductory videos on AI safety concepts and motivation.
Max Tegmark (2017). Life 3.0: Being Human in the Age of AI.
For a more technical audience:
- The long-term future of AI (longer version) (2015). A video of Russell’s classic talk, discussing why it makes sense for AI researchers to think about AI safety, and going over various misconceptions about the issues.
- Concerns of an AI pioneer (2015). An interview with Russell on the importance of provably aligning AI with human values, and the challenges of value alignment research.
- On Myths and Moonshine (2014). Russell’s response to the “Myth of AI” question on Edge.org, which draws an analogy between AI research and nuclear research, and points out some dangers of optimizing a misspecified utility function.
Robert Miles. AI Safety YouTube channel. Introductory videos on technical challenges in AI safety.
Scott Alexander (2015). No time like the present for AI safety work. An overview of long-term AI safety challenges, e.g. preventing wireheading and formalizing ethics.
Grace, Salvatier, Dafoe, Zhang, Evans (2017). When Will AI Exceed Human Performance? Evidence from AI Experts. A detailed survey of >300 authors at NIPS and ICML conferences, on when AI is likely to outperform humans in various domains.
Victoria Krakovna (2015). AI risk without an intelligence explosion. An overview of long-term AI risks besides the (overemphasized) intelligence explosion / hard takeoff scenario, arguing why intelligence explosion skeptics should still think about AI safety.
Stuart Armstrong (2014). Smarter Than Us: The Rise Of Machine Intelligence. A short ebook discussing potential promises and challenges presented by advanced AI, and the interdisciplinary problems that need to be solved on the way there.
Allan Dafoe (2018). AI Governance: A Research Agenda.
Everitt, Lea, Hutter (2018). AGI safety literature review.
Brundage, Avin, Clark, Toner, Eckersley, et al (2018). The Malicious Use of AI: Forecasting, Prevention and Mitigation.
Soares and Fallenstein (2017). Aligning Superintelligence with Human Interests: A Technical Research Agenda
Amodei, Olah, Steinhardt, Christiano, Schulman, Mané (2016). Concrete Problems in AI safety. Research agenda focusing on accident risks that apply to current ML systems as well as more advanced future AI systems.
Taylor, Yudkowsky, LaVictoire, Critch (2016). Alignment for Advanced Machine Learning Systems
Jacob Steinhardt (2015). Long-Term and Short-Term Challenges to Ensuring the Safety of AI Systems. A taxonomy of AI safety issues that require ordinary vs extraordinary engineering to address.
Nate Soares (2015). Safety engineering, target selection, and alignment theory. Identifies and motivates three major areas of AI safety research.
Nick Bostrom (2014). Superintelligence: Paths, Dangers, Strategies. A seminal book outlining long-term AI risk considerations.
Steve Omohundro (2007). The basic AI drives. A classic paper arguing that sufficiently advanced AI systems are likely to develop drives such as self-preservation and resource acquisition independently of their assigned objectives.
Malik, Palaniappan, Fisac, Hadfield-Menell, Russell, Dragan (ICML 2018). An Efficient, Generalized Bellman Update For Cooperative Inverse Reinforcement Learning. Uses the property of CIRL that the human is a full-information agent to reduce the complexity of the problem by an exponential factor.
Ratner, Hadfield-Menell, Dragan (R:SS 2018). Simplifying Reward Design through Divide-and-Conquer. An approach for inferring a common reward across multiple environments with separately specified rewards.
Fu, Luo, Levine (ICLR 2018). Learning Robust Rewards with Adversarial Inverse Reinforcement Learning. Using an adversarial approach to IRL to recover reward functions that are invariant to changing environment dynamics and distributional shift, by factoring out reward shaping from the learned reward function. This resolves IRL’s reward ambiguity in favor of safer reward functions.
Hadfield-Menell, Milli, Abbeel, Russell, Dragan (NIPS 2017). Inverse Reward Design. Formalizes the problem of inferring the true objective intended by the human based on the designed reward, and proposes an approach that helps to avoid side effects and reward hacking.
Fisac, Gates, Hamrick, Liu, Hadfield-Menell, Palaniappan, Malik, Sastry, Griffiths, Dragan (ISRR 2017). Pragmatic-Pedagogic Value Alignment. A cognitive science approach to the cooperative inverse reinforcement learning problem.
Milli, Hadfield-Menell, Dragan, Russell (IJCAI 2017). Should robots be obedient? Obedience to humans may sound like a great thing, but blind obedience can get in the way of learning human preferences.
Amin, Jiang, Singh (NIPS 2017). Repeated Inverse Reinforcement Learning. Separates the reward function into a task-specific component and an intrinsic component. In a sequence of task, the agent learns the intrinsic component while trying to avoid surprising the human.
Armstrong and Leike (2016). Towards Interactive Inverse Reinforcement Learning. The agent gathers information about the reward function through interaction with the environment, while at the same time maximizing this reward function, balancing the incentive to learn with the incentive to bias.
Hadfield-Menell, Dragan, Abbeel, Russell (NIPS 2016). Cooperative inverse reinforcement learning. Defines value learning as a cooperative game where the human tries to teach the agent about their reward function, rather than giving optimal demonstrations like in standard IRL.
Evans, Stuhlmüller, Goodman (AAAI 2016). Learning the Preferences of Ignorant, Inconsistent Agents. Relaxing some assumptions on human rationality when inferring preferences.
Irving, Christiano, Amodei (2018). AI safety via debate. Agents are trained debate topics and point out flaws in one another’s arguments, and a human decides who wins. This could allow agents to learn to perform tasks at a superhuman level while remaining aligned with human preferences. (blog post)
Saunders, Sastry, Stuhlmueller, Evans (AAMAS 2018). Trial without Error: Towards Safe Reinforcement Learning via Human Intervention. (blog post)
Christiano, Leike, Brown, Martic, Legg, Amodei (NIPS 2017). Deep reinforcement learning from human preferences. Communicating complex goals to AI systems using human feedback (comparing pairs of agent trajectory segments).
Abel, Salvatier, Stuhlmüller, Evans (2017). Agent-Agnostic Human-in-the-Loop Reinforcement Learning.
Specification gaming / wireheading:
Manheim and Garrabrant (2018). Categorizing Variants of Goodhart’s Law. Explores different failure modes of overoptimizing metrics at the expense of the real objective (which often results in specification gaming).
Lehman, Clune, Misevic, et al (2018). The Surprising Creativity of Digital Evolution: A Collection of Anecdotes from the Evolutionary Computation and Artificial Life Research Communities.
Everitt, Krakovna, Orseau, Hutter, Legg (IJCAI 2017). Reinforcement learning with a corrupted reward channel. A formalization of the reward misspecification problem in terms of true and corrupt reward, a proof that RL agents cannot overcome reward corruption, and a framework for giving the agent extra information to overcome reward corruption. (blog post)
Amodei and Clark (2016). Faulty Reward Functions in the Wild. An example of reward function gaming in a boat racing game, where the agent gets a higher score by going in circles and hitting the same targets than by actually playing the game.
Jessica Taylor (AAAI Workshop 2016). Quantilizers: A Safer Alternative to Maximizers for Limited Optimization. Addresses objective specification problems by choosing from the top quantile of high-utility actions instead of the optimal action.
Everitt and Hutter (2016). Avoiding Wireheading with Value Reinforcement Learning. An alternative to RL that reduces the incentive to wirehead.
Laurent Orseau (2015). Mortal universal agents and wireheading. An investigation into how different types of artificial agents respond to opportunities to wirehead (unintended shortcuts to maximize their objective function).
Interruptibility / corrigibility:
Aslund, El Mhamdi, Guerraoui, Maurer (2018). Virtuously Safe Reinforcement Learning. Designing interruptible agents in the presence of adversaries.
Ryan Carey (AIES 2018). Incorrigibility in the CIRL Framework. Investigates the corrigibility of value learning agents under model misspecification.
El Mhamdi, Guerraoui, Hendrikx, Maurer (NIPS 2017). Dynamic Safe Interruptibility for Decentralized Multi-Agent Reinforcement Learning.
Hadfield-Menell, Dragan, Abbeel, Russell (AIES 2017). The Off-Switch Game. This paper studies the interruptibility problem as a game between human and robot, and investigates which incentives the robot could have to allow itself to be switched off.
Orseau and Armstrong (UAI 2016). Safely interruptible agents. Provides a formal definition of safe interruptibility and shows that off-policy RL agents are more interruptible than on-policy agents. (blog post)
Soares, Fallenstein, Yudkowsky, Armstrong (2015). Corrigibility. Designing AI systems without incentives to resist corrective modifications by their creators.
Side effects / impact measures:
Zhang, Durfee, Singh (IJCAI 2018). Minimax-Regret Querying on Side Effects for Safe Optimality in Factored Markov Decision Processes. Whitelisting the features of the state that the agent is permitted to change, and allowing the agent to query the human about changing a small number of features outside the whitelist.
Krakovna, Orseau, Martic, Legg (2018). Measuring and avoiding side effects using relative reachability. A general side effects measure that penalizes reducing reachability of states that would be available by default, and avoids introducing certain bad incentives. (blog post)
Eysenbach, Gu, Ibarz, Levine (ICLR 2017). Leave no Trace: Learning to Reset for Safe and Autonomous Reinforcement Learning. Reducing the number of irreversible actions taken by an RL agent by learning a reset policy that tries to return the environment to the initial state.
Armstrong and Levinstein (2017). Low Impact Artificial Intelligences. An intractable but enlightening definition of low impact for AI systems.
Yudkowsky and Soares (2017). Functional Decision Theory: A New Theory of Instrumental Rationality. New decision theory that avoids the pitfalls of causal and evidential decision theories. (blog post)
Garrabrant, Benson-Tilsen, Critch, Soares, Taylor (2016). Logical Induction. A computable algorithm for the logical induction problem.
Andrew Critch (2016). Parametric Bounded Löb’s Theorem and Robust Cooperation of Bounded Agents. FairBot cooperates in the Prisoner’s Dilemma if and only if it can prove that the opponent cooperates. Surprising result: FairBots cooperate with each other. (blog post: Open-source game theory is weird)
Yudkowsky and Herreshoff (2013). Tiling Agents for Self-Modifying AI, and the Löbian Obstacle.
Babcock, Kramar, Yampolskiy (2017). Guidelines for Artificial Intelligence Containment.
Eliezer Yudkowsky (2002). The AI Box Experiment.
Leike, Martic, Krakovna, Ortega, Everitt, Lefrancq, Orseau, Legg (2017). AI Safety Gridworlds. A collection of simple environments to illustrate and test for different AI safety problems in reinforcement learning agents. (blog post, code)
Stuart Armstrong. A toy model of the control problem. Agent obstructs the supervisor’s camera to avoid getting turned off.
Michaël Trazzi. A Gym Gridworld Environment for the Treacherous Turn. (code)
Note: I did not include literature on less neglected areas of the field like safe exploration, distributional shift, adversarial examples, or interpretability (see e.g. Concrete Problems or the CHAI bibliography for extensive references on these topics).
Collections of technical works
FLI grantee publications (scroll down)
Berkeley AGI Safety course (CS294)
Paul Christiano. AI Alignment. A blog on designing safe, efficient AI systems (approval-directed agents, aligned reinforcement learning agents, etc).
Rohin Shah. Alignment newsletter. A weekly newsletter summarizing recent work relevant to AI safety.
If there are any resources missing from this list that you think are a must-read, please let me know! If you want to go into AI safety research, check out 80000 Hours guidelines and the AI Safety Syllabus.
(This list of resources was originally published as a blog post.)