In this article, I argue that democracy is a necessary component of AI safety. To show this, I respond to Bentham's Bulldog's piece, We Should Hand Off To Morally Reflective AIs, published by Forethought.[1] Bulldog recently expressed these views in podcast form, also published by Forethought. I believe that Bulldog is wrong in ways common among AI safety advocates, and therefore I hope that this article will provide a structure for better appreciating democracy within AI safety.
I first argue that AI handoff as argued by Bulldog cannot fit into liberal democracy. Bulldog's second move is to argue that, when choosing between "handoff to morally reflective AIs" and liberal democracy, we should choose the former. I deny this, even granting that such a trustworthy AI is plausible and that we could know when they become 'properly' morally reflective.[2] I finally give the implications of my argument for alignment research more broadly.
The mechanism by which Bulldog desires handoff is through suffrage for morally reflective AIs. He argues:
A number of people seem to conceive of handoff as a strange abrogation of the liberal order—one that replaces human decision-making with AI. But this doesn't have to be. One of the more promising ways of handing off would be giving economic and political rights to digital minds. Because digital minds could be so numerous, eventually this would lead to them making nearly all decisions. This would, in fact, be squarely in accordance with the norms of the liberal tradition, for it would give rights to morally important welfare subjects.
— We Should Hand Off To Morally Reflective AIs
In this section, I describe how universal AI suffrage as Bulldog describes it would not fit into our liberal democracy.
Bulldog believes that morally reflective AIs should be granted voting rights in virtue of their moral importance as welfare subjects. There are two issues: first, not only morally reflective AIs would receive a vote, and second, moral importance is not a sufficient condition for voting rights, even granted.
Bulldog argues that the intelligence explosion would create "quadrillion times more digital minds than human minds,"[3] each of which would suffer a political injustice by not having suffrage. Perhaps, but this would apply equally to non-morally reflective AIs as it does to morally reflective AIs. Additionally, we have no reason to believe that the first and most plentiful digital minds will be morally reflective, but they will have the necessary and sufficient conditions for voting. To make the liberal argument, he needs to be consistent about sentient AIs being able to vote, independent of their moral reflection or how alien their values are to ours.
This leaves us with a liberal democratic dilemma. The first horn is allowing the voter base to be overwhelmed by AIs, whether or not they align with human or morally real values. The second horn is the disenfranchisement of those AIs, which constitutes a political injustice. Bulldog, as his paper is written, does not have a way out of this dilemma.
I believe a better understanding of what confers voting rights gets us out of the dilemma. Specifically, I will defuse the second horn by arguing that AIs would not be disenfranchised.
It seems plausible that an AI becomes sentient within the century, but that is insufficient for voting rights. As Bulldog himself acknowledges, children and nonhuman animals cannot, and should not, vote. Being a "morally important welfare subject" is thus insufficient for voting rights.
Second, and more positively, the basis of voting rights is not sentience, but rather people's stakes in an issue. Even if equally morally considerable and equally smart, US citizens can vote in US elections while Canadian citizens cannot. Similarly, only Bostonians can vote in Boston's elections and only members can vote in a school's AI safety club. The reason for these delineations is that legitimate voting depends on the stakes that the voters have in the decisions. What gives me a vote within these systems is that I am a member, and as a member, am individually benefited or harmed by the group's decision.[4]
AI, since it is not properly a member in the vast majority of human affairs, would not qualify for suffrage in human communities, except perhaps in very limited ways. Note that this argument does not depend on AI systems never meeting the capacities or sentience that we traditionally see as necessary for a vote.
It seems plausible that AI systems could deserve a right to vote, but, by the nature of their being very alien to our shared way of life, legitimate participation could only happen in narrow ways that overlap with the American, Bostonian, or club way of life. What may harm or benefit members of those communities, such as water quality, local housing prices, or access to medical care, will not directly harm or benefit sentient AI systems. The AI's alien values and desires are unlikely to materially intersect with mine. As such, they have no right to make decisions on such issues, and I have no right to make decisions on their issues.
There may be narrower communities it would deserve a vote in, such as AI regulation. This would bring back the dilemma in a narrower way: AI interests overwhelming regulation vs. political injustice. This is the persisting dilemma that a liberal democrat must face. Even with this updated view, I would be trapped back in the liberal democratic dilemma, albeit in a more limited way. The only way out would be simply never creating these advanced systems.
Bulldog treats the creation of these digital minds as an inevitability, but it is not. It is an active choice that a select few people are making on behalf of the world, where they cause immense harm either to us or to AIs. This is itself a very strong reason to pause AI development now.
In the previous section, I described how AI handoff does not fit into the liberal democratic tradition. Plausibly, though, the author could (and I believe has) give up commitments to democratic legitimacy in favor of artificial enlightened despotism. Indeed, Bulldog's argument rests solely upon this despotic move if my above argument about suffrage is correct.
In the preamble to his piece, he "support[s] democracy only because [he thinks] it tends to reach more reasonable verdicts than other ways of making decisions." In this piece, he argues that "[o]n a number of plausible views, what matters with respect to societal decision-making is how good the decisions are, instead of who is making them."[5]
In this section, I problematize the concept of "good decision" that Bulldog uses, tying it intrinsically to "who is making [it]." I do this by analyzing what expertise means, tying it intrinsically to democratic legitimacy. Democratic value underlies good decision-making.
What is this democratic value? Bulldog misses the mark when arguing, against democracy, that:
It doesn't seem obvious that one has an inalienable right to make hugely consequential decisions on matters that they're barely informed about which affect others in enormous numbers.
— We Should Hand Off To Morally Reflective AIs
This claim is precisely the democratic value that I care about. I, too, believe that one does not have an inalienable right to make hugely consequential decisions on matters which affect others in enormous numbers. But this is exactly what would happen during AI handoff, and it is what grounds liberal democracy in the first place.
Indeed, what it means to 'know best' intrinsically includes the democratic process. John Dewey argues that 'best' is not just an idea that can be abstracted from the armchair, as advanced AI would do, but rather is directly informed by the people whom the decision impacts. 'Best' here carries both a normative and an empirical question. The guiding normative force is autonomy, or the right to decide one's life for oneself, and the empirical force is how well a decision meets that goal.
To argue this, Dewey problematizes the idea of expertise. Usually, what we mean by expertise is something like knowledge within a field, but Dewey defines knowledge functionally. Expertise is not defined by textbook subject-matter intelligence; rather, the definition of expertise is generated by those affected. Dewey says:
The man who wears the shoe knows best that it pinches and where it pinches, even if the expert shoemaker is the best judge of how the trouble is to be remedied.
— The Public and Its Problems, p. 224[6]
In short, what it means to be a good shoemaker is not understanding shoemaking principles and practice in applying them, but rather the ability to make the shoes that best suit the people wearing them. The shoe wearer is the one who guides inquiry.
To use Bulldog's example, we already do this with cancer treatment. Doctors give patients the tools and information they need in order to make informed and relevant decisions about living a life with cancer. Lawyers inform their clients about the best legal moves to get what their client wants. It would be morally wrong for doctors to sign their patient up for chemo without telling them, or for a lawyer to settle, even if by their lights it would improve their client's welfare. The reason for this is not epistemic, i.e. that the expert does not know enough. The reason is that the client carries the risk of the procedure or lawsuit, and so deserves a say in how it happens, even if it lowers their welfare.
Elected representatives function in exactly the same way: their job is not just to pass the best policy. Bulldog makes this mistake explicitly:
My claim is that the best future for humanity involves us being disempowered in some sense, in that humans aren't making most high-stakes decisions. As an analogy, representative democracy is, in some sense, a form of handoff—we hand off power to our elected representatives.
— We Should Hand Off To Morally Reflective AIs
Electing a representative is not "handing off power" unilaterally, but trusting them to represent the interests of the people who voted for them. Bulldog assumes that the job of an elected representative is to make the decisions that improve overall welfare, but this is not true. Representatives are not re-elected on the basis of whether they in fact made the right decision regarding tariffs, but on whether they enacted the policy that the people who voted for them wanted, whether or not it improved aggregate welfare. Irrespective of the quality of the decision, people will not re-elect a representative who refuses to enact the policy they voted for. The job of an elected representative is thus only to represent the will of the electorate. The value of this view lies not in overall welfare, but in people having control over their lives.
What this practically means for AI, then, is that there is no role for AI in the intrinsically valuable parts of democratic deliberation (that is, until AI gets relevant stakes and capacity). The role for AI in politics is then something like a doctor, lawyer, or representative, which augments our decisions rather than making them for us. This means that no decisions should be handed off to a delegated system unless that system meaningfully represents the people impacted by the decision.
This argument ought to be extended to the project of alignment more broadly. What alignment means to me is substantially different from what it means in the traditional alignment literature, as legitimacy does not rest solely on welfare maximization. Alignment thus means empowering the decisions that a social or political body makes, even if they do not maximize welfare.
On areas of convergence between those bodies, a hyper-intelligent LLM would be wonderful and remove the unnecessary frictions that hinder implementation. But disagreement is the interesting space, since those frictions may be precisely the useful result. Even if hyper-intelligent, AI could not legitimately make a decision on behalf of either party. The best and most useful thing it could do is facilitate conversation between the parties involved, or aggregate and simplify information so voters can better understand it.[7] The answer to politics is not in an algorithm. There are no shortcuts to democracy, and our only legitimate answer is doing politics well.
In this piece, I have given two reasons why AI handoff would be a bad idea. First, the method by which Bulldog wants to give AI power is unjustified within the framework he wants to work within, liberal democracy. Second, I argued that he should be less comfortable giving up liberal democracy than he is: liberal democracy is worth defending, even if AI could theoretically make decisions that improved general welfare more than average voters would. Finally, I discussed implications for alignment more generally.
Hereafter, I will refer to the author as Bulldog.
I doubt both claims.
I see no reason to believe this number, but will assume it for the sake of argument. It's at least equally plausible that sentient AI instances are psychologically indiscrete, and part of one whole 'person.'
An astute reader may note a potential contradiction between the prior paragraph and this one: children and nonhuman animals do have stakes in government decisions, and therefore should earn voting rights! But note that voting is not the only means of political representation. Representation can also be gained through ombudsmen or through delegated voting regimes.
Bulldog appeals to "a number of plausible views," making an argument from overlap. He cites only one plausible view: the Wikipedia page for the book Against Democracy. The book is not an exploration of plausible anti-democratic views, but rather one view. The argument from overlap that Bulldog attempts to make is therefore better suited to the pro-democracy camp, which has a plurality of intrinsic and instrumental reasons for preferring democracy.
Dewey, John. The Public and Its Problems: An Essay in Political Inquiry. Edited by Melvin L. Rogers, Swallow Press, 2016.
Seth Lazar and Lorenzo Manuali have a good paper discussing the current empirical contours of AI and democratic preservation: Lazar, Seth, and Lorenzo Manuali. "Using LLMs to Enhance Democracy." Minds and Machines, vol. 36, no. 12, 2026, https://doi.org/10.1007/s11023-026-09767-y.
