Governing autonomous systems at machine speed

Humans are not capable of practically overseeing the actions of autonomous systems, as these accelerate and proliferate. In this interview I talk about governing autonomous systems by their behaviour, in the same way the law governs humans, and how LexChip uses consent-based smart legal contracts to bring real-time oversight, breach detection and a legally admissible record to autonomous AI deployments.

Dylan Muir is founder and CEO of LexChip, a legal tech company building governance for autonomous AI systems through smart legal contracts.

Regulators increasingly demand explainability from AI. Is that achievable?

Explainable AI is of course a research domain of its own. I see interesting parallels to neuroscience. The thought experiment in neuroscience was always: if you had access to the entire connectome of a brain, would you automatically understand the computational principles? Once you have the full connectome, your reverse engineering work can begin. But that doesn't automatically give you insight.

These days, people think of large language models. To some extent we have the feeling that these can be directly interrogated for explanation, which obviates some of the uncertainty that people feel about neural networks and how they work. I'm not sure that provides us a solution. A large language model doesn't necessarily have the ability to be self-reflective. It doesn't necessarily have individual access to every one of its weights, or every one of its activations. In any case it's predicting token to token, and not necessarily providing insight as to where those token predictions come from mechanistically. If we take a step back in the direction of AI planning networks, whether transformer-based or deep neural network-based, the explainability problem becomes again much more acute.

So most AI models were not designed to be directly explainable, research directions aside of course. Regulators, however, can absolutely demand explainability from autonomous decision-making. Just because it's hard doesn't mean it's not necessary from a legal, moral, ethical, or regulatory perspective.

So how does LexChip approach the problem?

The approach we are taking with LexChip is to sidestep explainability, and examine compliance based on behaviour and actions taken or planned. A deterministic rule set can be applied to the decisions and behaviour of an autonomous system, even if the reasons or explanation for those decisions are not available.

Governing systems by their behaviour is exactly what we do with humans and the law. A human can be thinking the most despicable thoughts for their entire life, but if they never express them and never act on them, then the law doesn't care. It's the behaviour of a human that we can address, even if as humans we prefer to interact with others we believe are internally “good”.

Who is responsible for making those guardrails comprehensive?

We are putting forward behaviour-based compliance as a comprehensible, deterministic approach to system guardrails. At the same time, we are expecting users to take the responsibility of setting comprehensive guardrails; that responsibility lies with them, not with LexChip.

But the point is that with LexChip, you are doing something positive you can point to, as an attempt to mitigate and reduce risks of autonomous system deployment. It will never be possible to prove avoidance of every harm, but companies have a responsibility to do what is in their power to avoid and reduce harms where possible.

What about systems that act faster than any human can oversee?

The current scenarios, with either humans on or in the loop, or the legal system itself, are both already very much after the fact. The appeal of LexChip is to accelerate the reach of the legal system to the speed of autonomous decisions and behaviour.

Humans are not capable of practically overseeing the actions of autonomous systems, as these accelerate and proliferate. Imagine an autonomous agent swarm, spawned in response to a single human user request. There's no practical way for the human user to oversee and triage every decision made by every unit of the swarm, in real time. Automation is required for that real-time oversight, which is what LexChip provides.

Who should write the rules?

The beauty of the smart legal contract concept used by LexChip, from our co-founders Natasha Blycha and James Myint, as well as others, is to rely on consensual rule definitions such as contract law. Contracts are fundamentally consent-based agreements, where the several parties to the contract agree on what constraints, guardrails, automations, and interventions are implemented as part of the smart legal contract.

This same approach extends to policy agreements and regulation, which can be pulled into the LexChip system as dependencies of a smart legal contract, bringing with them their own guardrails and rules for automation, intervention, and response to breach.

So finally, the parties who are legally responsible for the actions of the autonomous system will also be legally responsible for defining and putting in place the guardrails for that system's deployment.

How is this different from blockchain smart contracts?

My co-founders have been working on usability of smart legal contracts for at least a decade. It's crucial for the adoption of smart legal contracts that current contract law experts are able to also draft them. That's who we want engaging with LexChip; we can't require coding experience or a degree in computer science to draft a smart legal contract.

That's a key point of distinction between our approach and blockchain-style smart contracts. In our case, we are talking about real, legally binding contracts, which have natural language as the main legal vehicle for terms.

Can any of this guarantee that harm won't occur?

There is no way to eliminate all possibilities for harm, through either accident, bad behaviour, or bad-faith operation of a system. Take the analogy back to contracts: legal contracts in no way ensure that breaches or mistakes won't occur; these are always possible. What a good contract does is outline where the responsibility for a breach lies, and the agreed-upon consequences for a breach. So while a contract cannot prevent every breach, it guides the response to a breach or mistake.

Likewise with LexChip: we cannot promise, and no system can plausibly promise, that all harms will be strictly prevented. But we can mitigate harms. We can anticipate breaches and build in guardrails that prevent bad behaviours when possible. We can detect a breach once it occurs and notify all parties immediately, reducing any follow-on or compounding harms. We can maintain a “black-box” style data record that becomes legally admissible, which speeds up and supports any legal proceedings that might follow. Our technology also helps companies to show they have taken steps to address their legal responsibilities when deploying autonomous systems.