On page 549 of “Algorithmic Accountability and Public Reason”, Binns mentions that the more algorithms are responsible for decision-making, the more we need a system of accountability. The author argues that the political theory of public reason can help decide a framework to do so. Public reason is the universal rules that can be justified on public grounds and shared by reasonable people, without appealing to beliefs that are religious, metaphysical, moral, or political in nature. That is, the aforementioned fields will have certain content that cannot be the basis for determining universal laws. The author suggests this theory is useful in minimizing the effects of algorithmic decision-making in that it will limit decision-makers to act on very specific standards that are acceptable to all reasonable people.

Given this context, I propose the question: who are these reasonable people? This question is important because the people responsible for public reasoning make universal rules that fit everyone (reasonable/not) the same way. One possible answer is that reasonable people are those who agree on universal laws that are ethical in the context of human rights like liberty, opportunity, and equal wealth distribution. But, if we take the example of India, gender equality is not an objective truth, which implies that every region has its own separate truths. This is because culture, public opinion, and upbringing are driving factors of objective truths, and consequently determine how people think. If I am born into a place that only thinks in a certain way, say that men and women are not equal, then my thinking would be conditioned in the same direction, and I would not be able to fathom a framework where there is such equality. Further, within a region, there may be subregions that think differently from each other, for example, abortion laws in the United States. The objective truth that a mother gets to choose what she wants to do with her body is not agreed upon everywhere: some states/regions have banned it and would punish her for the crime, others don’t, and some even support it.

It is not that these places are not reasonable, but that there are different kinds of reasonable thoughts. However, the problem is that we want to hold the algorithms accountable in the same way everywhere, but to do so, we would now have to exclude a place that does not follow the same reasonable thought process. We cannot account for every region’s opinions and customs when creating a universally accountable algorithm system, and it is unclear where to draw the boundaries of such regions, since two regions might differ in one thought but agree on another. We don’t have a universally agreed-upon definition of who counts as a reasonable person in the context of public reason, since regions differ in their basic moral values. Therefore, the practical implementation and scalability of such an accountability system for algorithms would not hold.

NEW DRAFT

In "Algorithmic Accountability and Public Reason," Reuben Binns argues that as algorithms take on more decision-making, we need a framework to hold them accountable, and that the political theory of public reason can provide one. Public reason is the set of justifications that reasonable people can share without appealing to beliefs that are religious, metaphysical, moral, or political in nature. The appeal for algorithmic accountability is clear: limit decision-makers to standards acceptable to all reasonable people, and you have a principled basis for what an algorithm is allowed to do.

Binns anticipates two objections to this proposal - that public reason is already operative through democratic law-making, and that opaque models resist accountability at all - and answers both. I want to press on a third objection he doesn't address, one that is specific to the algorithmic part of the proposal rather than the public reason part. Public reason was designed for human deliberation - legislatures, courts, slow institutions where a judge in one jurisdiction and a judge in another can both invoke it and reach decisions that feel locally legitimate. The standard gets renegotiated case by case, by situated reasoners who can see each other's objections. However, algorithms don't work that way. They're encoded once and deployed at scale, and whatever "reasonable person" standard you build in gets shipped, frozen, to every context the system touches.

This is where Binns' framework runs into trouble. He acknowledges that the precise content of public reason is supposed to "emerge from a process of reflective equilibrium between equal citizens," and explicitly declines to "presuppose any particular form" of it. That move is defensible in a philosophy paper - you can leave the substantive content to be worked out later. An encoded algorithm cannot punt. Somebody on the engineering team has to pick the standard before deployment, and that pick travels.

Take gender equality: there is no cross-jurisdictional consensus on what it requires, or even whether it is the right frame. The Rawlsian move is to say that anyone who rejects basic gender equality is not reasonable, full stop. That is a defensible philosophical position, but as an engineering specification, it means the algorithm encodes one region's reasonable-person standard and treats every other region's as a deviation to be corrected. Or take abortion in the United States - not across countries, but within one. There is no shared standard about what a reasonable person believes about bodily autonomy. A hiring algorithm, a content moderation algorithm, or a healthcare triage algorithm that encodes one position is not neutral; it is locally legitimate in some states and locally illegitimate in others, and there is no version of "reasonable" that fixes this.

Binns is most vulnerable to this worry in his discussion of how public reason should reassert universal principles against biases inherited from training data. He uses the example of landlords denying housing on religious grounds: if those biases would not survive scrutiny under public reason, neither should their algorithmic descendants. The argument assumes a "we" who can identify which biases violate public reason. At deployment scale, that "we" is the team writing the spec. Whose reflective equilibrium gets encoded?

You could object that human institutions handle this through federalism - different jurisdictions, different rules. But the whole appeal of algorithmic systems is scale and consistency. A nationally deployed model is not going to ship fifty variants. And even if it did, where do the boundaries go? Two regions might agree on bodily autonomy and disagree on speech, or agree on speech and disagree on equality. Reasonable-person standards don't decompose neatly along jurisdictional lines.

The deeper issue is that conditioning shapes what feels intuitively reasonable to the people writing the spec. I don't mean people are trapped by their upbringing (reformers exist everywhere, and minds change), but I mean that any team encoding a "reasonable person" into a system is going to encode the version that feels obvious to them, and that version will carry cultural priors they can't fully see. In human deliberation, those priors get surfaced through disagreement. In a deployed model, they get surfaced through harm.

The closest thing to a solution in the existing literature is what Iyad Rahwan calls "society-in-the-loop" — the idea that human-in-the-loop oversight should be extended from individual operators to entire affected communities, so that the values an algorithm enforces remain renegotiable by the people it acts on. A related line of work on "contestable AI by design" pushes this further: rather than trying to encode the right values up front, the system is built so that any decision can be challenged, reviewed by a human, and used to update the model. The EU's GDPR Article 22 and the AI Act's Article 14 codify versions of this — a right to human review of significant automated decisions, and a requirement of meaningful human oversight for high-risk systems.

I find this direction more promising than substantive public reason, for exactly the reason this post has been pressing. Contestability is procedural. It doesn't require the engineering team to pick the correct reasonable-person standard before deployment; it requires them to build a system that can be argued with after deployment, by the people the decision lands on. The "reasonable person" is no longer a frozen spec — they're a participant in an ongoing process, which is what public reason was always meant to describe.

But I don't think contestability fully escapes the problem either. Two issues. First, contestability is expensive at scale: a system that processes millions of decisions a day cannot offer meaningful human review of each one, and the literature is honest that most current implementations are what compliance researchers call "theatrical" — nominally present but structurally incapable of changing outcomes. Second, contestability still requires someone to adjudicate the contest. If a hiring algorithm's decision is appealed, a human reviews it — but that human is operating inside the same institution, with the same cultural priors, that wrote the original spec. You've moved the reasonable-person problem one level up, not solved it. The appeals process inherits whatever the deploying organization counts as reasonable.

The honest version of my position, then, is this: contestability is a better framework than substantive public reason for the deployment problem, because it treats disagreement as ongoing rather than resolvable. But it works only when the contesting party has real power — institutional, legal, or collective — to force a different outcome. Without that, "human in the loop" becomes a legitimacy stamp on the same encoded values, and the cross-jurisdictional problem returns in a slightly different shape.

So the project of algorithmic accountability is not dead, but the version that runs on a substantive reasonable-person standard is. What's worth building toward is contestability with teeth - enforceable rights to challenge decisions, jurisdictional pluralism in how appeals are adjudicated, and an honest acknowledgment that any deployed system is a frozen snapshot of one community's reflective equilibrium. The EU AI Act's Article 14 is essentially a bet on this direction, and the next few years will tell us whether "meaningful human oversight" can be made real or whether it collapses into the theatrical compliance the literature already worries about. Public reason might still be useful at this layer, but as a standard for what counts as a legitimate appeals process, not as a value set to be encoded in the model. The reasonable person was never supposed to be a spec; they were supposed to be in the room.