I was interviewed for the story Can a president consult an AI oracle? on AJP. I've been busy, but given how impressed I am with South Korea's efforts to hold their presidents' power to account, I could not say "no." But I could say "may I publish this all on my blog?"
Dear Eric (if I may),
I always prefer real-time interviews so we both agree what is worth spending time on, but given your charming letter I'll try to do this quickly
- How can we tell whether a president using an AI assistant remains meaningfully in control rather than simply endorsing a machine-generated conclusion?
This is a problem however powerful the individual is. The more powerful and important their decisions, the more important it is that they are immersed in a network that holds that power to account. There's no simple way to surveil for good use. We need to be vigilant in ensuring our political systems are safe and effective.
- Is there a principled difference between using a chatbot to research a legal question and publicly displaying its answer to lend authority to a political argument?
I believe it is always important to say where knowledge derives from, so people can check your sources and ensure they agree with your interpretation. This is the problem with genAI chatbots, even if they report source there's no reason to believe that is the part of their data from which they generated their predictive output. But whether you talk about a chatbot or a person, you can always go check the sources they mention and see whether these align. But we do need to dedicate some resources to doing these checks.
For example, just a few days ago I skyped the one slide from my talk with an AI generated image, and showed that it also had the prompt I'd used and the engine on the slide https://bsky.app/profile/did:plc:tkbweudpy6tvzjqdiza4z3p5/post/3my5a7w4xlc2a In this case, the image clearly did no harm and it doesn't matter much whether it's exactly replicable, but it's a matter of modelling good practice. I think leaders and experts should always try to do that.
- Could the apparent fluency and certainty of AI encourage political leaders to overestimate its reliability? What would meaningful human oversight require in this context?
Again, this is a problem for everyone, but I think the message is getting out that this is just the way the systems will always work, provided that the majority of the data input used for training is also fluent and confident.
- Your work stresses human-centred governance in demanding security settings. What lessons transfer to constitutional interpretation and other sensitive executive decisions?
My work actually emphasises this in all contexts, you must have just read perhaps my NATO paper? [Human-Centred Control of AI Is Both Possible and Necessary – that links to the PDF, and yes he had and was referencing it.] But here I make such recommendations for any user: https://joanna-bryson.blogspot.com/2025/02/generative-ai-use-and-human-agency.html There are two separate concerns: how you use genAI well, and how you make "sensitive executive decisions" well. These both have rules independent of each other, and both sets of rules apply if someone is using genAI for sensitive decisions.
- Should rules for AI use by elected leaders be based on the type of model, the sensitivity of the issue, the public consequences, or some combination of these?
My answer to 5 is not much different than 4, but I will take advantage of this opportunity to point out that sensitive national security contexts or any very important context you must also think about cybersecurity and espionage. But in my academic contexts, I to be honest don't use genAI very much, but if it isn't immediately evident and easy to get what I want, I try several independent models, preferably from different countries and knowledgebases. Just as when I'm trying to solve hard problems, I ask several people from a variety of relevant academic backgrounds.
How will you use this? May I have permission to include your email & my answers on my own blog after you've gone to press?
[Evidently, he said yes!]
Comments