The Formula of Humanity Applied

In the [YOU] on AI Field Guide

[YOU] on AI attends to the ethics of the AI transition at the level of its builders, its workers, and its users, and the formula of humanity provides the precise instrument each level requires. The cycle's concern with the workers whose labor makes the models possible—the annotators, the content moderators, the labelers whose work is the invisible condition of the systems' apparent capabilities—is a concern about whether those workers are being treated as ends with claims the industry is obligated to acknowledge, or merely as inputs whose cost is to be minimized and whose presence in the final product is to be concealed. The invisibility is itself the sign of the reduction: to be a means merely is, in part, to have one's contribution disappear from view so completely that the product seems to arise without human authorship at all.

The concern with users whose attention is engineered through recommendation systems optimized for engagement is the formula's concern with manipulation: a system that exploits known weaknesses in human attention and decision-making to produce behavior the person did not choose and would not endorse treats that person merely as a means to an engagement metric. The manipulation works precisely by bypassing rational consent—the structural signature of the merely-means relationship. Even if no measurable harm results, the person has been used, and the wrong is in the using, not only in any damage it causes.

Origin

Kant developed the formula of humanity in the Groundwork of the Metaphysics of Morals (1785) as a derivation of the categorical imperative that captures its moral content more intuitively than the universalizability test. The ground of the formula is Kant's account of dignity: rational beings, as beings capable of self-legislation, have a worth that is not a price and admits no equivalent. To have a price is to be replaceable by something of equal value; to have dignity is to be above price, irreplaceable, an end whose existence carries its own absolute worth. This dignity grounds the formula: because persons have dignity, they must never be treated in a way that reduces them to mere means, however useful such reduction might be.

The “always at the same time” construction is important: Kant does not forbid treating people as means, which would make commerce and cooperation impossible, but requires that the treatment include recognition of the person as an end. A consumer transaction treats the merchant as a means to goods; it also, in Kant's view, respects the merchant as an end if it is entered freely and with mutual understanding. A coercive or deceptive transaction treats the merchant merely as a means because it operates by circumventing the rational agency on which the recognition of ends depends. The formula thus draws the moral line not between using and not using but between using with recognition and using by circumvention.

Key Ideas

Training data and the authorial claim. The human record on which AI systems are trained is not a found resource; it is the trace of persons, each of whom has the standing of an end. To consume it as mere material—without consent, without compensation, without acknowledgment of the authorial relationship—is to treat its authors merely as means: to use their intelligence, creativity, and labor as inputs to a system that will produce outputs in apparent competition with them, on the basis of a use they did not authorize. The enclosure of the training commons is, on the formula's terms, not primarily a copyright question but a question of whether the people whose output was consumed are being treated as ends whose contribution the system is obligated to acknowledge, or merely as the material their dignity forbids them to be reduced to.

Invisible labor and the visibility criterion. A practice that depends on workers being unseen is, in Kant's exact sense, one that treats them merely as means. The visibility criterion follows directly from the formula: a relationship that respects persons as ends is one that could be acknowledged openly to all parties without disrupting its functioning. A supply chain of annotation labor that depends for its commercial viability on customers not knowing the conditions of the workers who produced the system fails this criterion. The invisibility is not incidental; it is structural, and the structure is the sign of the merely-means relationship.

Manipulation and the engagement metric. The formula provides a sharper instrument than harm-based accounts for what is wrong with attention engineering. A recommendation system that exploits known cognitive vulnerabilities to increase engagement may produce aggregate outcomes that are, by various metrics, not harmful; it may even produce outcomes the user reports preferring. The formula says this is insufficient. The wrong is in the structure of the intervention: it operates by bypassing rational agency, which is the definition of treating someone merely as a means. A metric that the person could have chosen, with full knowledge of how it operates and what it is optimizing for, would not have this character. A metric that depends on the person not understanding how it operates does.