Source-linked AI summary
The generative paradox: the Anti-Prompt paradigm for human capability
Margarita Trifonov Vadim E. Levit
TL;DR
Generative AI can improve joint products without establishing what people can do after assistance changes, making retained capability a design concern. The paper proposes developmental function allocation and an Anti-Prompt interaction order that leaves consequential acts open to human construction. It argues that product validity, responsibility for reasoning, and later capability require distinct evidence.
Problem
In education, judging AI mainly by human-AI product quality leaves unclear what people can understand, challenge, repair, and adapt after assistance changes or disappears.
Source: §1 · p. 1Method
Results
The paper distinguishes four forms of the generative paradox and argues that improved joint performance alone does not establish retained human capability.
Source: §1.1 · p. 2, §9 · p. 12Takeaways & Limitations
Product validity, who supplied consequential reasoning, and capability after support changes require different evidence.
Source: §8 · p. 11The proposed Anti-Prompt instructions alone do not establish that human construction or retained capability occurred.
Source: §3 · p. 6
Abstract
from arXiv · showhide
Generative AI in education is usually judged by the output of human-AI teams. That focus overlooks a second outcome: what people can still understand, challenge, repair, and adapt when assistance changes or disappears. We call this tension the generative paradox and distinguish four manifestations: performance, assistance, ownership, and capability stock. We propose developmental function allocation, which complements operational function allocation and meaningful human control by treating retained human capability as a design objective alongside current joint performance. The Anti-Prompt paradigm reverses the default interaction order: when immediate generation would preempt a consequential human act, AI should first create the conditions for human construction. We distinguish the validity of a product, responsibility for its construction, and evidence of capability after support changes. Representative Anti-Prompts, a canonical interaction sequence, and stress tests spanning mathematical reasoning, experimental science, software diagnosis, and lesson design make these distinctions concrete. We conclude with falsifiable predictions about retained capability and a system-level hypothesis about the reproduction and distribution of expertise.
1. The generative paradox
Generative AI can improve human-AI products while leaving uncertain what people can do independently after assistance changes. The paper frames this as generative substitution when AI displaces the human operation intended to develop or maintain a capability.
1. The generative paradox
- The design question is which consequential act remains open to the learner and what later evidence shows capability after support changes.The distinction between performance during instruction and durable learning is established in learning research.Source: §1 · p. 1
- Immediate output, later solo performance, motivation, and self-knowledge can diverge across AI-assisted tasks.Across four experiments with 3,562 participants, professional-output gains did not consistently persist in solo work and coincided with lower intrinsic motivation and greater boredom; in two reasoning studies with 698 participants, accuracy improved while participants overestimated performance.Source: §1 · pp. 1–2
- STEM findings vary by design: unrestricted math access reduced short-term unaided performance, a guarded tutor largely mitigated the loss, and a physics tutor produced substantial learning gains.Developers learning an unfamiliar programming library with assigned AI assistance showed weaker subsequent conceptual understanding, code reading, and debugging on average; exploratory analyses associated more cognitively engaged interaction with better learning.Source: §1 · p. 2
- Generative substitution occurs when AI displaces the human construction, evaluation, or revision intended to develop or maintain a target capability.Whether delegation is substitutive depends on the person and capability: routine calculation may be incidental for one learner, while proof strategy may be the target operation for another.Source: §1 · p. 2
The better AI becomes at producing intellectual products, the more
The generative paradox comprises dissociations between current products and human capability across performance, assistance, ownership, and expertise distribution. These outcomes require appropriate comparisons and cannot be inferred from product quality alone.
1.1 Four manifestations of the generative paradox
- Performance and assistance paradoxes arise when better joint output or progress coexists with unchanged or weaker independent capability, or when help removes opportunities to practice consequential cognitive operations.Tests compare immediate output with delayed reconstruction and transfer, and compare assistance conditions by which cognitive opportunities remain available.Source: §1.1 · pp. 2–3
- Ownership can diverge from control: people may author an artifact while AI frames the problem or supplies strategy, whereas AI execution can coexist with human capacity to challenge or recover.Responsibility profiles include framing, selection, execution, justification, monitoring, and authority to override and recover.Source: §1.1 · pp. 2–3
- Capability-stock paradoxes concern how locally efficient delegation may narrow future expertise needed for oversight, recovery, teaching, or improvement of AI-supported practices.The proposed indicators track opportunity, recovery, and expertise distribution longitudinally.Source: §1.1 · pp. 2–3
- These are observable dissociations, not logical contradictions; output accuracy, time saved, satisfaction, or AI-content share alone can miss control and retained capability.Fig. 1 presents hypothesized routes rather than universal laws; product-first assistance can support learning, and an Anti-Prompt intervention can fail.Source: §1.1 · pp. 2–3, Figure 1 · p. 4
2. Developmental function allocation
Developmental function allocation extends function-allocation and meaningful-control perspectives by treating retained capability as an outcome of present work distribution. It specifies capability dimensions and connects prospective allocation decisions to later evidence after support changes.
2. Developmental function allocation
- Developmental function allocation asks how today's distribution of work changes what people can later understand, challenge, recover, adapt, and teach.It connects human-factors allocation with research on learning, agency, deskilling, and meaningful human control.Source: §2 · pp. 4–5
- Retained capability is a profile spanning understanding, justification, monitoring and error detection, recovery, and adaptation or transfer, assessed alongside authority and opportunity to exercise it.Competence does not follow from formal authority, and error detection does not entail permission to override.Source: §2 · p. 5
- Anti-Prompt complements existing companion and Agentivism accounts by linking a specified consequential act, assistance timing and content, and later capability evidence.The contribution is this integrated design-and-evaluation focus, not the invention of learner agency, withholding, or post-support testing; a single prompt cannot guarantee the interaction follows it.Source: §2 · p. 5
- Operationalization specifies the target act and prerequisites before outcomes, then records responsibility across interaction functions and distinguishes an invitation, an attempt, and a justified contribution.Alternative valid solution paths remain possible, and authority and opportunity are recorded separately from capability dimensions.Source: §2 · p. 5
3. The Anti-Prompt paradigm
The Anti-Prompt paradigm reverses the interaction order when immediate generation would preempt a consequential human act, preserving an opportunity for human construction and later testing what remains.
3. The Anti-Prompt paradigm
- When AI production would preempt a capability-building act, Anti-Prompt reverses the default order so the person can construct or justify that act before disclosure.The interaction then tests capability after support changes.Source: §3 · p. 6
- The examples protect different acts: the proof prompt elicits and tests selected inferences, while the pendulum prompt asks learners to predict, control variables, address confounds, and adapt a design.The proof prompt is novice-facing; the prompts themselves do not establish that capability was retained.Source: §3 · p. 6
- A backward-hurdle variant works from a sufficient penultimate claim toward candidate subgoals, then requires forward reconstruction without equivalent AI assistance.For induction, reconstruction must justify each step rather than treat a desired subgoal as an established premise.Source: §3 · p. 6
- The canonical sequence identifies a consequential act, elicits a human model, introduces challenge, adjusts support, returns control, and tests retained capability.It is a heuristic, not a fixed protocol, and can branch to prerequisite instruction, direct explanation, accessibility support, or durable hybrid work.Source: §3 · p. 6, Figure 2 · p. 7
- Premature production can displace generation, retrieval, selection, justification, or monitoring, so the aim is the least substitutive assistance consistent with progress, access, and responsible performance.Withholding is not intrinsically beneficial; novices may need explicit modeling, while support can become redundant for experts.Source: §3 · p. 7
- The same correct proof can arise from different interactions, so product quality alone cannot show who controlled the reasoning or what each learner can later justify or adapt.The comparison must record what the learner was invited to do and what assistance supplied.Source: §3 · p. 8
4. When paradox makes revision necessary
A pedagogical paradox can challenge a learner’s model, but the useful work is diagnosing the failed assumption and revising the model, not experiencing surprise alone.
4. When paradox makes revision necessary
- A pedagogical paradox is a counterintuitive result or boundary case that exposes limits in a learner’s current model, distinct from the generative paradox of performance and retained capability diverging.The generative paradox concerns immediate joint performance and retained human capability.Source: §4 · p. 8
- The Koch snowflake challenges the assumption that perimeter and area move together: its perimeter grows without bound while its enclosed area converges to a finite value.Explaining the contrast requires a more precise account of their different convergence behavior.Source: §4 · p. 8
- An Anti-Prompt elicits a provisional model, introduces the smallest anomaly it cannot explain, and asks the learner to diagnose the failed assumption before supplying the decisive distinction.The checkpoint commits to an expectation, exposes a minimally sufficient anomaly, diagnoses the assumption, and tests reconstruction against a neighboring case; surprise alone is insufficient.Source: §4 · pp. 8–9
5. Stress tests across STEM practice
The stress tests frame capability as more than a correct product: it includes explaining, adapting, and making consequential decisions when support or conditions change.
5. Stress tests across STEM practice
- In the proof example, the protected act may be identifying the contradiction or justifying the parity step, rather than reproducing the entire proof unaided.The novice-facing prompt supplies some architectural cues; an advanced version can withhold the lowest-terms and parity cues.Source: §5 · p. 9
- In software incident response, a successful patch can restore service without preserving causal diagnosis; stronger evidence includes explaining the failure chain and recovering from a related failure under changed conditions.The consequential act includes tracing dependencies and selecting a safe recovery strategy.Source: §5 · p. 9
- In lesson design, polished materials do not establish ownership; stronger evidence includes justified rejection, counterfactual adaptation, and enactment under changed conditions.The protected decisions include the target idea, anticipated student thinking, representation sequence, and evidence of learning.Source: §5 · p. 9
6. The capability-stock hypothesis
Repeated allocation decisions may reshape the future distribution of expertise, creating a system-level capability-stock hypothesis. The hypothesis is unvalidated and calls for longitudinal testing rather than inference from output alone.
6. The capability-stock hypothesis
- The capability-stock hypothesis proposes that repeated allocation decisions change the future distribution of expertise.It treats practice opportunities as inflows to future capability and nonuse, attrition, and organizational exit as sources of depreciation.Source: §6 · p. 9
- Possible indicators include independently capable practitioners, recovery after automation failure, training throughput, unaided alternatives, and succession or apprenticeship rates.The passage presents these as possible indicators, not as a validated stock measure.Source: §6 · p. 9
- An observational analysis associated AI adoption in science with higher individual output, smaller teams, fewer junior researchers, and a narrower collective research focus.These associations do not establish deskilling, but are consistent with a possible bottleneck in expertise reproduction and distribution.Source: §6 · p. 9
- The capability-stock hypothesis should be tested longitudinally rather than inferred from output alone.Local efficiency may diverge from long-term resilience.Source: §6 · p. 9
- Restricting novices to inspecting finished AI products while experts retain consequential decisions may concentrate rather than democratize expertise-development opportunities.A capability-preserving institution should consider access to formative practice, not merely access to AI.Source: §6 · p. 10
7. What would count as a decisive test?
The proposed tests distinguish immediate product success from later capability, while isolating allocation choices and their conditions. The predictions are conditional and require comparison with strong capability-preserving alternatives.
7. What would count as a decisive test?
- A first comparison should distinguish direct generation, generated solutions with structured self-explanation, Socratic assistance, and Anti-Prompt allocation.Outcomes should include immediate performance, delayed reconstruction, near and far transfer, unfamiliar-error detection, confidence calibration, and recovery; actual act-level allocations matter more than condition labels.Source: §7 · p. 10
- A focused study would randomize learners to attempt a target inference before disclosure or receive it before explaining, matching task, model, explanatory material, checker, and support budget.It would record actual assistance and time, then assess a prevalidated unaided task variant after a prespecified delay.Source: §7 · p. 10
- The primary analysis should follow assigned conditions, include incomplete products, report attrition, and jointly report product success, later capability, and instructional costs.The test concerns what the allocation changes, not which condition produces more attractive dialogue.Source: §7 · p. 10
- Component studies should vary construction protection, responsibility visibility, support withdrawal or transformation, and forward versus backward hurdle sequencing.Pre-specified outcomes, blinded scoring where feasible, and independent coding of responsibility profiles can help distinguish intervention effects from predictive analyses; mediation alone would not establish a causal mechanism.Source: §7 · p. 10
- The proposed central hypothesis predicts better later performance on a target act when learners with prerequisites construct or justify it before disclosure than under matched early disclosure.The predicted benefit is smaller or absent when the act is already mastered or indispensable prerequisites are missing.Source: §7 · p. 10
- The paradigm must be compared with strong Socratic, independent-first, or other capability-preserving approaches, and equivalent outcomes do not establish implementation superiority.Predictions should be tested against a prespecified smallest educationally meaningful advantage, with uncertainty reported.Source: §7 · p. 11
8. Assessment and governance
Assessment must distinguish product validity, who supplied consequential reasoning, and capability remaining after support changes. Governance should make allocation and responsibility visible while preserving accessibility, privacy, transparency, and human agency.
8. Assessment and governance
- Product validity, responsibility for consequential reasoning, and capability after support changes require different evidence.A proof certificate supports a formal claim under checking assumptions but does not establish a person's understanding, originality, or contribution to a decisive step.Source: §8 · p. 11
- Assessment can sample delayed reconstruction, unfamiliar errors, recovery from misleading suggestions, and adaptation to changed constraints.Fresh task variants may reduce exact-answer rehearsal, but do not guarantee new reasoning; variants need verified solutions, appropriate difficulty, and alignment with the protected act.Source: §8 · p. 11
- No single indicator is decisive: reconstruction can be rehearsed, confidence miscalibrated, and interaction traces strategically shaped.Removing equivalent reasoning assistance need not remove accessibility accommodations or ordinary tools that do not perform the target act.Source: §8 · p. 11
- Policies should distinguish product-oriented from development-oriented AI use and specify when direct generation, capability-preserving allocation, or accessibility and safety considerations apply.They should also clarify who retains responsibility when AI guidance fails.Source: §8 · p. 11
- Interfaces could display current allocations and let users reframe or override them, while research records capture only assistance needed to interpret the target act.Privacy, transparency, and human agency remain central constraints.Source: §8 · p. 11
- The aim is a sufficiently broad distribution of human capability for informed delegation, independent challenge, recovery, teaching, and innovation—not preservation of effort or manual performance for its own sake.Consequential work may be distributed across people and tools when responsibility remains visible and recoverable.Source: §8 · p. 11
9. Conclusion
The conclusion argues that output quality alone cannot resolve the tension between present performance and future capability. It calls for designing and valuing both the joint product and retained capability after assistance changes.
9. Conclusion
- Generative AI can improve present performance while displacing practice through which future human capability develops.Developmental function allocation treats retained capability as a design outcome alongside current joint performance.Source: §9 · p. 12
- When immediate generation would substitute for a consequential human act, the Anti-Prompt paradigm reverses the default order and first creates conditions for human construction.The most valuable AI contribution in some contexts may be a question, obstacle, counterexample, or paradox that leaves the consequential act open to the person.Source: §9 · p. 12
- Human-AI interaction has two outcomes to design for and value: the joint product and retained human capability after assistance changes.The stated scope includes AI-enabled learning environments and professional practices they prepare learners to enter.Source: §9 · p. 12