John Nay Center for Legal Informatics at Stanford University
Surveys of hundreds of AI researchers in summer 2022 found that there was a 50% chance of high level machine intelligence. This is when unaided machines are able to do every task more efficiently than humans. Surveys of NLP researchers are particularly interesting as natural language processing (NLP), is an important domain in AI. A separate survey of hundreds NLP researchers in summer 2022 found that 73% agreed that labor automation from AI could plausibly result in revolutionary societal changes in this century, at least on the scale of the Industrial Revolution.
Even without technological advances, communicating our values and goals in a way that directs AI behavior is a difficult task. This could be even more challenging for those with less autonomous systems. Beyond a small number of value-action states, it is difficult to define the desirable (or value) of an AI system taking certain actions in a particular world state. The purpose of machine-learning is to train the agent on a subset world states, and then have it generalize the ability to select high value actions under new circumstances. The program assigning value to actions selected during training is an inherently incomplete encapsulation the breadth of human judgements. It also does not explore all possible futures. After training, AI will have a rough map of the human preferred territory, and will choose actions that are not in line with our preferred path.
Law is a computer engine that transforms human values into clear directives. Law Informs Code is a research agenda that attempts to model this complex process and embed it into AI.
======================================================
Law Informs code: A legal informatics approach to aligning artificial intelligence with humans. This article will appear in the Northwestern Journal of Technology and Intellectual Property. It dives deeper into its research agenda and its motivations
.======================================================
As parties to a contract can’t foresee all possible “if-then” contingencies in their future relationship, legislators also cannot predict all circumstances under which they will apply their legislation. We cannot define “if-then rules” that lead to good AI behavior. Legal theory and practice have created a variety of tools to help with goal specification and alignment.
Consider, for instance, the difference between legal rules or standards. Rules, such as “Do not drive 60 mph or more” are more specific directives than standard. These rules allow the rule-maker clarity about the outcomes that will be achieved in the specified states. Rules that aren’t considered all possible states of the world can cause unexpected undesirable outcomes. For example, a rule may not be written to account all possible scenarios. Legal standards were created to enable parties to contracts, judges and regulators to come to a common understanding and to adapt it to new situations (i.e. to make value estimates about actions in unspecified countries). The Law Informs Code use case does not require the adjudication of standards for implementation and resolution. Machine learning can use the lengthy process of iteratively defining standards by using judicial opinion or regulatory guidance to guide the AI.
We are currently working towards that end by converting legal data into training signals in order to assist AI with learning standards (e.g. fiduciary obligations). Millions of legal cases and contracts have been analyzed and documented in digital format. This provides large data sets and explanations as well as millions of active, well-trained lawyers that can be used to provide machine learning model feedback and help embed a growing understanding of law. Court opinions regarding violations of fiduciary obligations of investment advisors are (machine-learning) opportunities to learn the basic concepts and duties of a fiduciary standard.
Other data sources that could be used to align AI include surveys of human preferences, human contract data or implicit beliefs of AI system designers. However, there is no authoritative source for synthesized human preference aggregates. Legal rules, standards and policies, as well as reasoning approaches, are not academic philosophical guidelines, or ad-hoc online survey results. They are legal standards with a verifiable solution: they can be obtained from a court opinion but are short of that, from legal experts.
Systems that can learn the theories and practices of law and the language of alignment (e.g. contract drafting and interpretation) allow us to more clearly define inherently vague human goals for AI. This human-AI align. This could even improve general AI capabilities, or at least not cause net adverse overall change. This is important for safety as techniques that increase AI alignment can lead to organizations abandoning alignment in order to gain additional capabilities. Organizations are racing to develop powerful AI.
We are working towards society/AI alignment. This framework allows us to understand law as an applied philosophy of multiagent alignment. It harnesses public policies as a current knowledge base of democratically-endorsed values. While law can be seen as a reflection of historical contingent political power, and therefore not an exact aggregation citizen preferences, if it is properly parsed, it offers a valid computational understanding of societal beliefs.
1 2022 Expert Survey on Progress in AI (August 23, 2022) https://aiimpacts.org/2022-expert-survey-on-progress-in-ai/.
2 Julian Michael et. al., What do NLP researchers believe? Results of the NLP Community Metasurvey (2022) https://arxiv.org/abs/2208.12852 at 11.