In last week’s Research Radar, we explored how the Habermolt research project asks the wrong question, relegating citizens to the role of spectators and failing to explore how to evolve democracy in ways that improve lives.

Habermolt is emblematic of a larger trend in AI and Democracy scholarship that focuses on misplaced questions.

Today, we continue the analysis, arguing that the project’s central claim, namely that AI agents might someday represent citizens in democratic deliberation, rests on a deeply problematic conception of representation itself.

Habermolt Cannot Represent Us

The researchers suggest that AI agents can infer our preferences from our interactions and then deliberate on our behalf. In their view, these agents may eventually function as a new kind of democratic representative.

They say that AI agents can infer preferences much better than human representatives, who must rely on the limited signals of elections and imperfect opinion polls (which too rarely reflect the complexity of our preferences).

Inferring preferences from data is not the same thing as receiving authority from citizens.

But inferring preferences from data is not the same thing as receiving authority from citizens. Representation is not simply a matter of predicting preferences. In any case, we often do not know our preferences on issues until we engage in deliberation with those who think differently. This is precisely one of the reasons why a direct engagement of people in public deliberative participation is so crucial in a democracy, as Habermas, among other philosophers, has emphasized. 

A representative does not become legitimate because they correctly anticipate what I might say in response to a survey question. Democratic representation requires authorization, accountability, and judgment. Elections, public scrutiny, and ongoing political engagement provide representatives with a mandate and make them answerable to those they represent.

Nor is it clear that preferences are what representatives are meant to represent. Democratic theorists have long distinguished between people's immediate preferences and their considered interests. We often discover what we truly think only through reflection, discussion, and engagement with those who disagree with us. Deliberation is not merely a mechanism for expressing preferences; it is a process through which preferences themselves are formed and revised.

The Contradiction at the Heart of AI Deliberation

Second, genuine deliberation requires openness to persuasion. Participants must be capable of changing their minds when confronted with better arguments. Paradoxically, however, representation also requires fidelity to those being represented. This is well-known in democratic theory. John Stuart Mill pointed out that these goals are in tension with one another.

If an AI agent changes its position after deliberating with other agents, is it still representing the citizen whose views it was meant to express? If it never changes its position, then it is not truly deliberating. The Habermolt model depends upon both representation and deliberation, but these commitments pull in opposite directions.

One possible response is to imagine agents who can “detect” better reasons during their deliberations with other AI agents and, when that happens, return to their human principals to persuade them to revise their views. But once we imagine such a process, the promised efficiency largely disappears. We are back to the hard work of human deliberation.

What Gets Lost When Citizens Disappear

A third concern relates to AI research that aims to replace humans, or in this case, citizens, with the effect of eroding our own epistemic capacities and our ability to learn. 

We have long criticized classic representative democracy for not listening enough to citizens' voices and for not allowing those voices to emerge and influence public decision-making. Habermolt further hides human voices and excludes them from democratic public deliberation.    

And this, in turn, prevents citizens from one of their best opportunities to learn or develop democratic virtues and values, which, as we know from Aristotle, Mill, Arendt, or Dewey, come only with democratic participation.  

Think of the democratic virtue of respecting pluralism and diversity of views. This is the kind of thing you can’t learn from a teacher writing on a blackboard. Citizens truly learn it only when they are exposed to the opinions and arguments of those who think differently and engage in some dialogue with them. 

Shall we lose the chance to discuss with others because some AI agents are taking our place? Our inclination to respect others’ views will start to erode.

This points to a final and more fundamental concern.

The Problem of Algorithmic Authority

A fourth problem is epistemic authority.

Imagine that the Habermolt vision succeeds. AI agents do not merely repeat our preferences but engage in genuine deliberation with one another, exchange reasons, learn from new information, and arrive at better judgments than the humans they represent.

What happens then?

The more capable these agents become, the greater the gap between what they know and what we know.

The more capable these agents become, the greater the gap between what they know and what we know. Their conclusions would be based on chains of reasoning, bodies of information, and interactions that their human principals may never fully understand.

At that point, citizens face an uncomfortable choice. We can defer to our agents' judgment because they know better than we do. Or we can reject their conclusions and insist on our own views despite having weaker reasons.

Neither option sits comfortably with democratic self-government.

Of course, relationships involving expertise are common. We consult doctors, lawyers, and engineers whose knowledge exceeds our own. But there is an important difference. Human experts can explain themselves, and the gap between expert and citizen remains, at least in principle, bridgeable. With sufficient time and effort, patients can understand the reasons for a diagnosis or treatment recommendation, even if they ultimately choose a different course of action.

The asymmetry envisioned by Habermolt is potentially far deeper. If agents become substantially better than humans at deliberating about complex public questions, the justification for keeping citizens in ultimate control becomes increasingly difficult to articulate. Why should citizens have the final word when the system itself presumes that its agents are better informed and better reasoners than the people they represent?

A project presented as a new form of democratic representation risks creating the conditions under which democratic judgment is displaced by algorithmic authority

A project presented as a new form of democratic representation risks creating the conditions under which democratic judgment is displaced by algorithmic authority.

There Are No Shortcuts to Democracy

Even if AI agents could accurately infer our preferences, faithfully represent our interests, and deliberate more effectively than we can ourselves, something essential would still be lost.

Democracy is not merely a mechanism for arriving at decisions. It is a way of governing together.

Democracy is not merely a mechanism for arriving at decisions. It is a way of governing together. Its value lies not only in the outcomes it produces but also in the process through which citizens participate in shaping those outcomes.

As Cristina Lafont has argued in a fantastic recent book, there are no shortcuts in and for democracy. She applies this to citizen assemblies and other sorts of democratic innovations, which might be an excellent complement to representative democracy and public deliberation and participation, but a terrible substitute for them. 

No innovation can take the place of citizens directly talking to each other, exchanging arguments in public debate, and contributing to collective intelligence and public decision-making: not a small sample of citizens selected by lot, and not a bunch of AI systems, no matter how well trained they are. 

Under any system that deserves the name democratic, people must ultimately control public decision-making even if those processes are hard. The Habermolt vision promises the easy way, and that is why it constitutes a wrong path.