A paper published this week on arXiv hit with unusual clarity: Rismanchian et al. found that students using generative AI on math problems finished faster and retained less. They called it "faster completion, less learning." The same week, a separate arXiv paper introduced SOLAR, a self-optimizing autonomous agent designed for lifelong learning and continual adaptation in AI systems. The pairing is uncomfortable. We are building machines that learn continuously and persistently while simultaneously building tools that teach humans to learn less.

The Shortcut Economy and Its Costs

The Rismanchian study is specific about the mechanism. Students using AI didn't just skip steps. They reduced their engagement with the underlying conceptual structure of problems. The speed gain was real. The knowledge built was not. This maps onto a broader pattern visible across media and culture right now: the optimization of completion rates at the expense of depth. Meta's new Forum app, described as a space for "deeper discussions," is also a product shaped by engagement metrics. Depth is the stated goal. Retention is the actual measure. The two are not the same thing, and the infrastructure almost always optimizes for the latter.

SOLAR's Inversion: Machines That Never Stop Learning

The SOLAR paper proposes an AI architecture explicitly designed to overcome the catastrophic forgetting that plagues current LLMs, where learning new things erases old competencies. The irony is almost too neat. Researchers are building AI systems with robust lifelong learning capabilities at the precise moment when AI tools are measurably degrading human lifelong learning. A 2025 paper in Computers and Education by Mollick and Mollick noted that the skills most at risk from AI shortcutting are the foundational ones: the mental models built from struggle that enable transfer learning in novel situations. What SOLAR is engineered to preserve in machines, the current deployment model of AI tutors is quietly eroding in people.