Integrity in (AI safety) policy advocacy
Why successful advocacy might need to look more like diplomacy
I work in AI safety policy advocacy, which involves persuading decision-makers to take actions that make the world safer from AI. As with any kind of advocacy, doing so effectively requires tailoring my arguments to different audiences, based on an understanding of what they care about and how they think. How do we do this in a way that’s high integrity, without sliding into the kind of persuasion that might seem a bit duplicitous or disingenuous?
I find it easier to start by thinking about what integrity in policy advocacy is not. Some important examples that come to my mind include:
Making blanket arguments or claims you don’t really believe or haven’t properly examined, simply because you think it will help you persuade someone of some end goal you care about.
For example, claiming “regulation definitely won’t harm innovation” because you know policymakers are concerned about that, without examining and developing a more nuanced view on exactly how regulation might impact different forms of innovation.
Totally changing your core arguments depending on the audience you’re talking to.
For example, telling an innovation-focused audience that regulation definitely won’t slow growth while telling a safety-focused audience that some slowdown is inevitable and acceptable.
Hiding or deliberately downplaying your real motivations for advocating for a particular policy.
For example, telling policymakers that you want regulation because you want to create more clarity for small businesses, when you’re actually motivated by mitigating risks.
Being totally rigid in your goal of convincing someone else of your position, and unwilling to consider the possibility of an alternative or more mutually beneficial way of achieving your goals.
For example, being overly rigid in the assertion that regulation is the only or most important way to mitigate AI risks, and shutting down alternative ideas as going against your advocacy goals.
High integrity policy advocacy, by contrast, I think involves:
Being open and honest about where you might be wrong, and where possible being rigorous in examining the evidence for your claims. Asking yourself, “do I really believe this claim, or am I a bit tempted to say what’s convenient?”; “am I happy to defend this claim if someone really presses me on it?”
Keeping the same consistent core message across audiences, even where context-specific details and supporting arguments might change. Asking yourself: “If these different people I’m talking to compared notes, would either feel they’d been misled about my position?”
Being transparent about your motivations and objectives in advocating for specific policies; distinguishing between your main reasons for supporting a policy, and what might be other good reasons for doing so. For example, I might genuinely believe that regulating frontier AI companies would be beneficial to small businesses in the UK, but if my main motivation for pursuing regulation has more to do with risk mitigation, I should lead clearly with that. This doesn’t mean I can’t also talk about other benefits, but being transparent about your goals makes it easier for others to assess your claims, including by taking into account any biases or blind spots you may have.
Being open to the possibility that the policy you are advocating for might not be the best or most effective way to achieve your goals, and working with other stakeholders to find mutually beneficial actions where possible. Rather than asking “how can I make the case for my preferred policy in terms this other person cares about?” I might instead try asking “is there a policy idea that serves my goals and is also good for what this person cares about?”
This suggests an approach to advocacy that might look quite different from the “stereotypical” picture of advocacy where one person tries to persuade another of a specific policy proposal. My experience has been that often the best policy change happens through a more collaborative, iterative process with decision-makers, based on mutual understanding of goals, assumptions, and uncertainties. It won’t always be possible to find options that perfectly serve everyone’s goals, but even acknowledging areas of misalignment openly can help with avoiding an overly zero-sum approach to policy advocacy. For example, it might be the case that key decision-makers care most about specific types of impact AI regulation could have on the economy, and that by understanding this, we could develop regulatory proposals which avoid that downside in particular. In this sense, high-integrity policy advocacy might actually have a lot in common with diplomacy and have a lot to learn from what has been found to work well there.