This is a super important and deep concern about the limits of agentic representatives. Inattention and abstention can be very important parts of how people delegate in a democracy. Agents that artificially remove that might make many political problems worse, not better.
This gets back to one of my main questions about how we design good agentic representatives---they have to do more than naively mirror our unconsidered preferences. I was mostly thinking that means pushing back on poorly considered ideas we have that don't actually match our underlying preferences, but Brad's comments show that there's another very important aspect, which is developing a notion of how important an issue is to the user, and knowing when to push hard on the issue and when to abstain.
More broadly, Brad also raises the important point that democracy is more than a technological engineering problem---it's inevitable that we will shift the technology we use to do democracy and that agents will be a part of it in the long run, but we shouldn't expect that to magically solve fundamentally human problems.
So much more to think about here!
Some Saturday morning musings, as I await the World Cup: this is an interesting approach and one that's really worth thinking about. It gets a lot right in saying that you don't want concentrated power in control of these agents. That an arguably important problem in politics (if not the most important one) is that of principal-agent misalignment. That you wouldn't want your agent to simply be a digital twin that mirrored both your best and worst impulses, but rather should be something that interrogates the latter. All good.
That said, it viscerally strikes me as problematic, and while I don't have a nuanced theory as to why I find it so troublesome, perhaps it becomes down to discomfort with trying to make the messy part of politics simply a design failure rather than the residue of legitimate, incommensurable bargaining that can't be reduced to an algorithm. It all seems a very Silicon Valley thing: recasting a tragic or constitutive feature of politics as a tooling problem.
Stated slightly differently, the argument seems to assert that the major problem in American politics (or perhaps politics in a democratic regime more broadly) is a failure of representation technology, i.e., that we elect people who inadequately reflect the actual public will and who can pass policies that represent that of a concentrated minority rather than the general public, or perhaps even pursue some maximization of their own preferences at public expense.
But is this the major problem? And, if so, can it be solved without introducing a more serious principal-agent misalignment?
It seems to me that selective inattention is part of a democratic equilibrium, and I would worry if my agent were to represent every preference I have in every particular debate.
Let me give you an example from my time in Congress. I was asked about overturning the Clinton administration's prohibition on snowmobiling in Yellowstone. I will confess that, representing rural eastern Oklahoma, I had no strong views about snowmobiling in Yellowstone and actually never even contemplated the topic. Over the course of a couple of beers, I am sure I can offer some views about the merits of snowmobiling, all ones not particularly well tutored. It would probably lead me to think that snowmobiling should be limited or even restricted in Yellowstone. This take is basically relying on my casual, almost aesthetic preferences, rather than any deep philosophical or political economy view.
So would I want my agent to fight for this at the 4 a.m. debate over snowmobiling in Yellowstone? Absent my agent, I would probably pay no attention to the question at all and leave it to be fought out among local stakeholders: businesses, environmentalists, ranchers, people who live nearby, affected Indian tribes, animal welfare specialists, conservationists, and more. Each of these people would have more considered views and a much greater stake in the outcome. They are far more likely to get it right than my rather gratuitous opinion from a small town in eastern Oklahoma about what the fate of snowmobiling in Yellowstone should really be.
So the assertion that a major problem in politics is misalignment of principal and agents seems to suggest that a major deficit of democracy is something along the lines of "I have preferences, but I lack the time and information, and so those preferences are not adequately expressed." But it seems to me, as in my Yellowstone example, a truer description might be "I have no serious view, no stake, no developed judgment, and no real claim to intervene." So having my constantly vigilant agent transform my faint, half-baked, low-salience attitudes into an active political force seems perverse. It makes actionable what is, in fact, almost complete indifference. Maybe I shouldn't be involved in questions of Yellowstone management (even if, as a federal park, I have standing to do so)? If my agent infers a bundle of preferences about how Yellowstone should be handled, then it actually disrupts the democratic process, which today, through the mechanism of salience, weighs interested parties' views more heavily.
So the real problem is my agent might represent me too well. My low-stakes preferences become of equal weight to the people who are most directly affected by it. My views on Yellowstone should, properly in a democracy, remain weak and deferential. Sometimes not caring is the proper political approach.
Of course, you might argue that the agents can develop some kind of model of salience that would solve the problem that I just discussed. This seems possible, and indeed you gesture at it when you talk about the agent not simply being a digital twin. The agent presumably could act on some kind of rule that says only support my preference when the issue reaches a particular salience threshold and disregard other times.
This is where, if this is the solution, the implementation gets really murky because salience is a morally loaded concept. Sometimes issues are not salient to me simply because I am being negligent. Sometimes it is humility. Sometimes distant interventions are paternalism, and sometimes outside interventions are necessary to save local communities from themselves. It's very hard to know.
So the Hallian agent must be able to discern considered conviction from gratuitous opinion, moral necessity from meddling, and decide when to represent my opinion, when to reinterpret my opinion, when to mute my opinion. That's why creating such "democratic agents" is about solving more than a principal-agent problem driven by the triage of attention. It is a deep question of political and moral philosophy. This comes back to why it seems to me, perhaps unfairly, to be such a Silicon Valley approach. It takes the most fundamental questions of human existence and tries to find a technical solution to them.
Creating agents that can engage in democratic deliberation seems hard to the point of impossible. And it might even be dangerous: every political question becomes a universalized conflict among all possible preferences.
