TL;DR
This position paper introduces the generative paradox: the tension between the immediate performance gains of human-AI collaboration and the long-term erosion of independent human capability. The authors propose the Anti-Prompt paradigm, a design shift that prioritizes "developmental function allocation"—intentionally withholding AI-generated answers to ensure humans perform the "consequential acts" necessary to build and retain expertise.
What Problem Does This Paper Address?
In the rush to integrate Generative AI (GenAI) into education and professional workflows, we typically measure success by the quality and speed of the final product. However, Trifonov and Levit argue that this focus ignores a critical second outcome: retained human capability.
When AI generates a complete proof, a block of code, or a lesson plan, it often performs "generative substitution," displacing the very cognitive operations (construction, evaluation, revision) through which human skill develops. The authors identify four manifestations of this paradox:
- Performance Paradox: Joint output improves, but independent capability stays flat or declines.
- Assistance Paradox: More help reduces friction but removes opportunities for justification and monitoring.
- Ownership Paradox: A human may "author" a document, but the AI controlled the framing and strategy.
- Capability-stock Paradox: Local efficiency gains accumulate into a system-level loss of the expertise needed to supervise AI or recover from its failures.
Methods
The paper is a theoretical position paper that synthesizes research from human factors, learning sciences, and AI interaction. It proposes a new framework called Developmental Function Allocation. Unlike traditional "operational" allocation (which asks "who should do this task now?"), developmental allocation asks "how does this distribution of work change what the human can do later?"
To operationalize this, the authors introduce the Anti-Prompt paradigm. An Anti-Prompt is an interactional design where the AI is instructed to withhold the final product until the human has performed a "consequential act." The paper outlines a canonical interaction sequence:
- Specify the target human capability.
- Elicit an initial human model or strategy.
- Introduce challenge (e.g., a counterexample or paradox).
- Adjust support dynamically.
- Return control as capability develops.
- Test what remains after assistance is withdrawn.
Datasets
Not discussed in the paper. As a position paper, it does not introduce new datasets but cites several large-scale experiments (involving over 4,000 total participants) to support its claims about the divergence between assisted performance and solo retention.
Key Insights and Ideas
- Scaffolding vs. Substitution: The same AI contribution can be a helpful "scaffold" for an expert (reducing routine burden) but "substitutive" for a learner (removing the core learning opportunity). The paradox is relational, not a fixed property of the tool.
- The Capability-Stock Hypothesis: The authors suggest that repeated delegation to AI acts as a "depreciation" of human expertise. If novices only ever inspect finished AI products rather than constructing their own, the future distribution of expertise needed to innovate or fix AI errors will narrow.
- The "Product-First" Trap: Current AI interaction defaults to immediate generation. The authors argue this should be reversed whenever product generation preempts the practice through which capability develops.
- Divergent Outcomes: Fig. 1 illustrates how "Product-first" routes can lead to high immediate performance but weak retained capability, whereas the "Anti-Prompt" route represents a developmental investment.
Key Results
While the paper is conceptual, it highlights critical findings from cited literature:
- In STEM learning, unrestricted GenAI access improved assisted math performance but reduced short-term unaided performance (Bastani et al., 2025).
- GenAI collaboration can improve professional output while simultaneously lowering intrinsic motivation and increasing boredom (Wu et al., 2025).
- Participants often substantially overestimate their own performance when assisted by AI, leading to poor self-knowledge calibration (Fernandes et al., 2026).
Limitations and Future Work
The authors acknowledge that the Anti-Prompt paradigm is not a "magic bullet." Its benefits are conditional: if a learner lacks the necessary prerequisites, withholding AI assistance may lead to unproductive struggle rather than learning.
Future work requires:
- Longitudinal testing of the capability-stock hypothesis to see if AI-heavy environments actually lead to system-level deskilling over years.
- Refining "Stress Tests" across different domains (like software diagnosis or experimental science) to determine which specific "consequential acts" are most vital to protect.
- Governance and Policy: Developing AI use policies that distinguish between "product-oriented" tasks (where speed is king) and "development-oriented" tasks (where human growth is the goal).
Summary updated