Artificial intelligence that tailors content to each user could make clever people sharper and leave everyone else behind. That is the prediction of an economic model by Indira Puri of New York University and Laura Veldkamp of Columbia University.
Their paper was published online in the Journal of Monetary Economics on December 24, 2025, and appears in the journal’s January 2026 volume.
The brain as a muscle
The model rests on two assumptions, each backed by an existing literature. The first comes from medicine: the ability to process complex information improves with practice, much like a muscle. The authors cite studies linking crosswords, cognitive training and other mental exercise to better cognitive performance.
The second comes from the AI literature: big data and algorithms let content providers predict each person’s abilities and preferences, and tailor what they serve accordingly.
In the model, profit-maximizing providers choose how complex their content should be. Without AI, they do not know who is who and serve everyone the same thing. With AI, they know each consumer’s type and match complexity to it. Less able consumers get less demanding material, so they get less mental exercise, while more able ones may benefit from content pitched at their level.
The authors put it bluntly. AI-driven content, they write, “is like placing candy on the plate of every diner that has a hard time finishing their vegetables.”
Children and the long run
The model predicts that this kind of targeting consistently harms children, who are served low-complexity content and miss the challenging material that would help them develop.
Extended over generations, and allowing for firms that cannot read people perfectly, the model points toward long-run convergence of cognitive ability, but at a lower level than before. That happens if people dislike content that is too hard more than content that is too easy. The authors call these long-run predictions speculative and untestable.
Suggestive evidence
To look for traces of the mechanism, the authors measured the readability of Facebook posts from the official pages of the two main US parties, about 10,080 Democratic and 1,946 Republican posts. The variation in complexity within each month rose over time on both pages. The authors stress this is “far from identifying a causal effect.”
They also point to the OECD’s 2023 Survey of Adult Skills, which assessed more than 220,000 adults aged 16 to 65 across 38 countries and found declining skills in most of them, especially in literacy.
The paper argues that this channel deserves a place in debates over AI regulation and financial literacy, and that what matters is not AI as such but any technology that matches information more precisely to the person receiving it.
Study Details:
- Title: Artificial intelligence and cognitive inequality
- Authors: Indira Puri, Laura Veldkamp
- Journal: Journal of Monetary Economics
- Publication Date: December 24, 2025
- DOI: 10.1016/j.jmoneco.2025.103884