Reading through the article "A Rational Optimist View Of Preventing Agency Decay" is a rich experience. For readers with less patience, here is a ChatGPT summary (that also generated the title of this post).
Much current discussion of artificial intelligence swings between two poles: utopian efficiency and apocalyptic takeover. The more consequential reality lies between these extremes. The emerging risk is not that machines suddenly replace us, but that we gradually hand over pieces of our cognitive life—judgment, initiative, authorship—without noticing the cumulative effect.
The argument in Colin Lewis’s recent essay is straightforward: AI’s primary impact is not abrupt displacement but cognitive offloading. Tasks once requiring human attention and judgment are incrementally transferred to machine systems. This process is economically rational and often highly productive. In one example, an audit process that once required weeks can now be completed in an hour with AI assistance. But such gains come with a hidden shift: the human role is no longer defined by doing the work, but by nominally overseeing it.
This leads to what the author calls agency decay. The issue is not simply job loss, but the erosion of meaningful participation before any job disappears. First, the human is assisted. Then the human supervises. Eventually, the human remains as a formal point of accountability while the substantive reasoning has migrated elsewhere. The signature is human; the cognition is not.
This shift has broader systemic implications. Modern institutions—markets, governments, cultural systems—have historically depended on human participation. That dependence has acted as a constraint, keeping systems at least partially aligned with human interests. If AI reduces the need for human cognition across many domains, that alignment weakens. The system no longer needs us in the same way, and therefore has fewer built-in reasons to serve human flourishing.
Importantly, this is not a sudden rupture but a slow transition—the “boiling frog” scenario. Productivity gains accumulate incrementally. Each step is locally rational, even beneficial. Yet taken together, they shift the locus of intelligence away from human minds toward institutional and computational systems. What disappears is not competence, but ownership of judgment.
Against this, Lewis offers a restrained form of optimism. The key claim is that human agency need not be defended as a sentimental relic. It can be justified on functional grounds. In high-stakes domains, retained human judgment is not inefficiency; it is infrastructure: a source of error correction, adaptability, and accountability. Systems that eliminate it entirely may become brittle, opaque, and ultimately less reliable.
This reframes the problem. The issue is not whether AI should increase productivity—it will—but whether our metrics of success are too narrow. Efficiency measures speed, scale, and cost reduction. It does not capture qualities like judgment, contestability, or moral responsibility. If institutions begin to price these properly—through regulation, professional standards, and organizational design—human agency can remain structurally embedded rather than nostalgically preserved.
The deeper point is that AI forces a clarification: what is human agency for? If thinking, deciding, and creating are increasingly shared with machines, then the value of human cognition must be specified more precisely. Not all tasks need to remain human. But some forms of judgment—especially those involving uncertainty, accountability, and meaning—may be indispensable.
The optimistic scenario is not one in which technological progress slows, nor one in which markets abandon efficiency. It is one in which societies become capable of distinguishing between mere acceleration and genuine capability. That requires expanding our evaluative frameworks beyond productivity alone.
In that sense, AI does not simply threaten agency; it exposes how poorly we have defined it. The challenge is not to resist automation, but to decide, with greater clarity than before, where human judgment is essential—and to design systems that preserve it there.