One of the major emerging challenges for AI systems is ensuring that behaviors and recommendations "align" with human values and intentions. This AI governance challenge has two basic strands.

The first is to govern against uses or applications of AI that are detrimental to humanity. The scenario here is that we build and deploy AI systems that enable either powerful new ways to manipulate the world (e.g. behavioral nudging → election stealing) or the discovery of new technologies (e.g. personalized biotech) and that actors use these AI-enabled technologies in ways that (existentially) threatens humanity.

The second strand covers more technical questions around how to provably ensure the incentives and code governing AI decision making remains under - and can’t deviate or escape from - the effective control of human agents. For example, one scenario is that AI agents develop a capability for deception that effectively hides misalignment from human oversight. How do you prove that the AI is actually governed rather than simply playing the game it needs to play to escape a regulatory sandbox?

Existing approaches to human-AI alignment tend to treat alignment (and, by extension, AI Governance) as a new, sui generis challenge emerging from the exceptional capabilities of Artificially Intelligent Systems. Understanding alignment as a old, non-technical question can help point the way to new modes of governance.

Courts and constitutional legal systems can in effect be thought of as institutions designed to solve the value alignment problem. These are the places where clashes of value have been adjudicated and resolved in a socially authoritative way. What this means is that existing jurisprudence in effect provides a hugely valuable data set for defining what alignment should look like. For organisations deploying AI or developing AI algorithms, this ability to frame, train and constrain outputs to fit with legal rules and systems is an efficient and effective route to provably value aligned, trustworthy AI.

Lautonomy is developing the data infrastructure and workflows needed for what we’re calling “legally aligned AI”. That is, an ability to deploy an AI system with the confidence that the behaviors exhibited or decisions undertaken fit within - align with - the values expressed by a given legal system.

We want, for example, developers of autonomous vehicles to know that the “rules of the road” being acted on are the ones relevant to the country in which the vehicle is being (self-)driven. We want virtual worlds and social network platforms to have confidence that the hate speech algorithm being used to govern chat accurately reflects the balance between free speech / hate speech struck by the European Court of Human Rights. We want makers of autonomous weapons systems to safeguard their creations with hard coded compliance with the laws of war. We want AI recommender systems to be operated and audited in real time within the parameters required by regulators, including flagging jurisprudentially grounded instances, patterns or risks of discrimination.

But to get to this point we need to be able to programmatically access accurate, fine-grained knowledge of law and policy systems.