If AI takes over the jobs that quietly kept companies answerable to people, who's still watching the store?

This explores whether human jobs inside a system quietly do alignment work, so that replacing them with AI removes a check nobody designed on purpose. The clearest case in the corpus is the 'gradual disempowerment' argument Does incremental AI replacement erode human influence over society?. Economies, governments and cultural institutions stay roughly aligned with what people want partly because they depend on workers who care how things turn out. That dependence is implicit alignment: a firm or agency can only drift so far from human preferences before the people running it push back or leave. As AI takes over that labor, the check weakens without anyone deciding to remove it. Because institutions depend on each other, their drift can compound until it becomes hard to reverse. The replacement is also uneven. Firms with more exposure to AI swap out online freelance workers faster and more cheaply than others do Do firms substitute labor for AI at different rates?. So the erosion probably shows up first in pockets, not evenly across the economy.

The same pattern appears at the level of individual models, under different names. A review of more than 400 alignment papers found that the field almost always treats alignment as changing the AI's behavior. Very little work looks at how humans adapt to AI, and the review warns that this gap slowly wears down people's ability to oversee these systems Why does alignment research ignore how humans adapt to AI?. Jan Leike makes a related point from inside a lab. Today's alignment successes are 'easy mode' because humans can still read and audit what models do. Once models act in ways people can't follow, the hard problem comes back Can we solve AI alignment before models become uninterpretable?. Put these together and labor and oversight look like versions of the same thing: humans who stay close enough to the work to notice when it goes wrong.

A more philosophical thread explains why this closeness might matter. One argument, drawing on the philosopher Charles Peirce's theory of signs, holds that goals written only in symbols can't be guaranteed to match real values. They need contact with the world and with other people Can AI systems achieve real alignment without world contact?. Human workers are one of the main ways AI systems get that contact. Work on self-improvement reaches a similar conclusion from another direction: reliable improvement needs checks from outside the system, because a model can't fully verify itself What actually constrains AI systems from learning misalignment?. Another argument says AI should follow the norms of particular social roles, such as doctor or teacher, instead of averaged-out preferences Should AI alignment target preferences or social role norms?. That raises an open question the corpus doesn't settle: who keeps a role's standards alive once people no longer fill it?

One note complicates this. Simply having humans take part doesn't produce alignment. In a 12-person study, people who helped design their own AI preference agents felt well represented. Independent testing found the agents were only partly aligned, and more generic than the people's own answers. Taking part created the feeling of alignment without the substance Does co-design participation hide misalignment in preference agents?. So keeping humans involved isn't enough on its own terms. What protects alignment is humans whose involvement can actually change outcomes and catch mistakes. Losing labor matters because it usually removes that, but a human in the loop as a token gesture doesn't bring it back.

A caveat: only one note in the collection takes on the labor question directly. The rest connect to it from oversight, grounding and verification. The idea that work itself functions as an alignment mechanism is still an open claim here, not an established result.