I recently spoke with Emily Gillett from the Institute for Law and AI who told me about the concept of a personal theory of change. I’ve done a lot of thinking about the ToC for Safe AI in Law; but this was the first time I’d heard about someone refer to it in the context of an individual.
I have a strong belief in the utility of having a 3-5-year plan, and it’s advice I find myself giving others often - without a destination every direction can feel arbitrary/aimless. The idea of a personal ToC resonated, so here’s a first attempt at one.
Starting at the end. I want to have contributed to a world in which our institutions have adapted to advanced AI and met a challenge that I fear they do not currently see. The ultimate goal (or destination) is that stronger institutions help society manage wider AI risks.
What this means in practice is difficult to articulate, and something I’ve struggled to explain in a satisfying way to those involved in AI safety, governance, and research. My experience in law has shown me (often viscerally) how institutional failure affects people. I am concerned that the rush to implement AI systems can compound those failures, and undermine or overwhelm the institutions on which we depend. Institutional capacity is both something AI can damage, and something society needs to prevent harm. Courts, governments and other institutions help monitor risk, enforce obligations, and coordinate response to incidents. Weakening those (i) makes other risks more likely, and (ii) if they eventuate, harder to manage.
There are two limbs to the threat that advanced AI poses to existing institutions - the loss of human oversight, and the inability to match the speed of AI generated/directed outputs.
Working backwards from this end: I have a background in law (across legal practice, government policy, civil liberties advocacy, and most recently as a barrister). I’ve worked on changing the law from outside government, within government, and arguing about it or applying it in court. I’ve developed technical AI research skills, completing behavioural LLM evaluations during a research sprint. I loved the experience of turning a question into a methodology, gathering and analysing the data, and producing a research submission. I’ve also co-authored a paper on AI incident response that I later used as the basis of a submission to an Australian parliamentary joint select committee hearing. I can combine practical legal and institutional knowledge with a growing technical understanding.
I can contribute to the ultimate goal in several ways, (1) my SPAR research with Elija Perrier can help decision-makers understand the reliability and limits of mechanistic interpretability, it can form the basis of adapting our existing institutions or the creation of entirely new institutions to deal with the novel problems that arise with advanced AI; (2) Safe AI in Law (SAIL) can help practitioners and executives understand AI risk, improve scrutiny and safeguards, shift corporate culture, and impact the diffusion/adoption of AI systems; (3) policy/governance work can provide another path to implementing necessary institutional reform. I enjoy advocacy. One aspect in particular is translating ideas/concepts for different audiences. It’s what I also love about teaching. I get excited when I see someone finally understand a concept I’ve been teaching/explaining.
Factors relevant to any decision: quality and impact of the work I do, external demand, personal sustainability, and likely contribution.
Assumptions I’m making:
- institutional capacity and resilience is tractable and neglected
- my experience and skills add marginal value
- mechanistic interpretability can address institutionally relevant questions or help set the bounds/limits of institutional reforms
- I can find ways in which to sustain any contribution or involvement
Table:
| test | what would justify more investment | |
|---|---|---|
| SPAR research | Complete SPAR project; connect a technical AI safety outcome to a legal question. | external reviewers find the work useful; it produces follow-up questions; I want to continue the day-to-day work. |
| Safe AI in Law | Publish a series of LinkedIn posts explaining select MIT AI Risk scenarios in the context of executives/legal practitioners; monitor the practical response. Develop a CPD curriculum for lawyers. | practitioners change a relevant practice or decision; the work feels worthwhile relative to its demands. |
| Policy or governance research | Explore collaborations where my institutional knowledge and research skills address a clear need (e.g. incident response, independent auditing). | demand for my contribution, good mentorship, a path to influence/impact, the work fits my interests. |
I intend to reflect on this in a week or so, and then perhaps a month later (plus after any major update/change to my personal circumstances or after any major AI-related event).