Why the curve isn't a simple U
A U-shaped story would say: a little delegation hurts learning, more delegation helps again, so the relationship bends back upward. The mathematical model of human skill and AI delegation does not produce that shape. Instead it treats skill and delegation as two quantities that evolve together: delegation rises when the AI outperforms the person, skill grows with practice and fades without it, and both are driven by the same goal of reducing task error [2].
Run that feedback loop forward and you get two stable outcomes rather than a curve with a dip and a recovery. One is a high-skill equilibrium, essentially human-only learning where the person keeps practicing. The other is a stable low-skill equilibrium of persistent reliance. Between them sits a sharp basin boundary — a dividing line such that, once the dynamics carry a learner past it, the pull toward the low-skill side becomes effectively irreversible [2].
What moves the tipping point
The model's answer to "what shifts the turning point" is AI quality. The authors characterize how better AI deforms the basin boundary — that is, it changes where the line between the two equilibria falls and how easy it is to cross [2]. The uncomfortable implication is that a stronger assistant can make the low-skill outcome more attractive to slide into, because the short-run payoff of delegating looks better even as the long-run skill trajectory worsens.
That split between short and long run is central. The model shows AI assistance can strictly improve performance in the near term while producing persistent performance loss over the long term compared with never using AI, driven by a negative feedback between delegation and practice: the more you delegate, the less you practice, and the less you practice, the more delegating pays off [2]. The effects held up under noise and asymmetric trust updates in the model, which suggests the mechanism is not a fragile artifact of perfect assumptions [2].
What this means for keeping learning deep
If the risk is a one-way slide rather than a temporary dip, then timing matters more than dosage. The model's sharp basin boundary implies that early decisions are disproportionately consequential, because once the dynamics settle into persistent reliance, climbing back out is hard [2]. That reframes the practical question from "how much AI should I use" to "when in the learning process am I using it, and does it replace the practice that builds skill."
Work on AI-driven adaptive learning frames the same tension in design terms, describing the delegation of decision-making responsibility to a machine as a turning point and arguing that AI should function as a support while humans supply the depth of the learning experience [1]. That is a design aspiration rather than a tested result — the work does not establish a U-shaped relationship or pin down what shifts a turning point [1]. Read alongside the model, it points the same direction: the depth of learning depends on whether delegation is arranged to preserve the practice that builds skill, not on finding a magic middle amount of AI use.
Sources used in this answer
AI-Driven Adaptive Learning: Architecture, Governance, and a Road Map for Scalable Personalization
Discusses delegating decision-making to machines as a turning point and argues AI should support rather than replace the human depth of a learning experience, without establishing a U-shaped relationship or identifying what shifts a turning point.
The Geometry of Learning Under AI Delegation
Models human skill and AI delegation as a coupled dynamical system and finds two stable equilibria — high-skill human learning and persistent low-skill reliance — separated by a sharp, effectively irreversible basin boundary that AI quality deforms, with AI able to improve short-run performance while causing long-run skill loss.