Responsible AI An umbrella term for AI ethics, AI safety, AI governance, trustworthiness and related topics. We understand it as the responsible development of AI: all considerations during AI research and development that concern societal impact, morality, safety, security and the environment. Beyond academia, the term also covers corporate guidelines, national policies and regulations.
Multidisciplinary Multiple fields working on AI, each on its own.
Interdisciplinary Multiple fields collaborating on AI, for example computer science together with philosophy, political science and law. Responsible AI needs this kind of collaboration, which does not always happen.
Areas of Responsible AI Our (subjective) division of Responsible AI into seven subfields, mapped along two axes: technological vs. socio-political solutions, and immediate impact vs. long-term consequences. The areas are Ethics & Fairness, Safety & Alignment, Governance & Law, Environment, Security & Privacy, Trustworthiness, and Mind, Justification & Values.
Ethicists vs. Long-termists A recurring dispute inside Responsible AI. "Ethicists" focus on harms AI causes today, such as discrimination. "Long-termists", usually from the AI safety community, focus on long-term technological and existential risks. Each side sometimes dismisses the other, and they compete for attention and funding. We see both agendas as important and not in contradiction.
Practical ethics of AI: bias and discrimination, human rights, and how the resources for AI are sourced. It targets immediate problems, though these are often deeply rooted. The debate is mostly socio-political, with some technical solutions.
Fairness Whether AI systems are fair and unbiased, meaning they do not discriminate. Algorithmic fairness develops tools to audit and debias machine learning systems.
Human rights (in AI) Whether AI is used in accordance with human rights and for their benefit, not against them.
Resource ethics (data and resource ethics) Whether the resources AI needs (data, semiconductors, human labour) are produced ethically, without exploiting people. Examples: data collection in countries with weak protections, and cheap crowdsourced labelling work such as Amazon Mechanical Turk.
Domain-related AI ethics How to use AI ethically in a specific field such as journalism or art, without breaching intellectual property or norms around plagiarism and trust.
Safe design of capable AI systems and their monitoring. Mostly technical, it tries to pre-empt future (possibly existential) problems for humanity. Once ridiculed, it became much more respected with the rise of GPT models.
Alignment Developing AI models aligned with human values or value systems. Which values those should be is itself a long philosophical debate.
Monitoring Tracking the capabilities of models, which are mostly built without safety guarantees. Approaches include benchmarks and tests, or interpretability, which tries to understand what is going on inside the model.
Superintelligence AI vastly more capable than humans, and the scenario in which it threatens humanity. The key reference is Nick Bostrom's book Superintelligence.