acceptodds
Under review as a conference paper at ICLR 2027

Can Verification Sustain Skill?

Abstract

A common way to make AI delegation reliable is to have AI generate an output that a human verifies and corrects. A static view of this workflow treats verification competence as fixed. Over time, however, practice and non-use can change this competence. We model a worker whose skill determines manual and verification reliabilities and thereby shapes workflow-specific performance losses and completion costs. We study how skill evolves when the worker chooses among manual work, verified delegation, and pure delegation to maximize current utility by trading off performance loss against completion cost. Current skill determines workflow choice, which shapes future skill. Under verified delegation, reviewing provides verification practice, whereas generative practice occurs only when the worker detects and repairs an AI error. We characterize the global dynamics and identify three verification-related regimes: collapse to zero skill, an interior skill level sustained by verified delegation, and competence sustained at a manual–verified switching boundary. Crucially, verified delegation can remain optimal over a skill range while no positive skill level is sustainable. Adding verified delegation can improve current utility yet eliminate an otherwise high-skill basin. Holding learning dynamics fixed, workflow incentives alone can produce all three regimes. Better AI can further reduce retained human skill by reducing manual practice and the errors that trigger repair-based learning. The regimes persist over nontrivial learning-parameter ranges and under additional learning from correct AI outputs, while task-level randomness can drive trajectories near a stable verified equilibrium into pure delegation. Thus, keeping a human in the loop need not preserve the capability required for effective future oversight.

Primary area:
optimization
Keywords:
verification, skill retention, AI delegation, switched learning dynamics

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_yAGFOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper studies whether human verification of AI-generated outputs can sustain human skill over time. The authors model a worker who chooses between three options: doing task manually, using AI and checking its outputs or fully delegating the task to AI. Each choice affects future skill through learning or forgetting. The authors show that verification does not always prevent skill loss and workers may gradually lose skill even while checking and correcting AI outputs. It may be likely that skill may collapse, stabilize through verification or be maintained by switching between manual work and verification depending on the conditions. The authors also show that better AI capabilities can lead to greater skill loss by reducing manual practice and opportunities to correct AI mistakes and that random events can push workers from stable verification into full delegation. The paper provides theoretical proofs, numerical examples, sensitivity analyses, and stochastic simulations. It also includes an illustrative mapping from educational data to model parameters, but does not empirically validate the long-term predictions.

Critical strengths
  1. This is an important problem about long-term human-AI interaction and its impact. The authors focus on how repeated AI use changes human skill over time instead of focusing on the immediate benefits of AI assistance. This is particularly relevant as human verification is often proposed as a sanity check for reliable AI use. The paper argues that keeping humans in the loop for checking AI outputs does not necessarily retain their skills.

  2. The theoretical analysis identifies conditions in which verification can sustain skill, need for manual practice or collapse of a skill. The interesting observation from the work is that even when verification could sustain skill if continued, workers may switch to full delegation before reaching that stable level. The paper also shows that changing workflow incentives alone can lead to different long-term outcomes without changing the learning dynamics.

Critical weaknesses
  1. The model uses a single skill variable for both doing the task manually and verifying AI outputs. I think this is a major limitation since the main question of the paper is whether verification can sustain skill. These two skills can develop differently. For example, someone who reviews a lot of AI-generated code may become slower at writing code but better at finding bugs. The current model cannot capture this. Also, since verification and manual work share the same skill, any forgetting from not doing manual work also reduces verification ability.

  2. The model doesn't factor in that working with AI is a skill as well which people can get better at over time. Verification can also be done through automated ways such as tests, external tools and not just manual debugging which the current model fails to account.

  3. I think the results depend a lot on the learning and forgetting assumptions. Manual work always gives at least as much learning as verification, learning slows down at higher skill levels and forgetting can be larger than what is gained from verification. Under these assumptions, it is expected that even highly skilled workers lose skill while verifying. Section 4 shows that collapse can be avoided with stronger verification learning or weaker forgetting but it is not clear which parameter values are realistic.

  4. The result that better AI cannot improve human skill also seems to follow from the model design. So, in this model, the worker learns from verification by a fixed amount that does not depend on better AI and correct AI outputs doesn't necessarily improve learning. Appendix E.2 shows that learning can improve with correct AI outputs under a fixed AI quality. However, it is unclear if the findings hold once when both additional learning and workflow switching are considered.

  5. The workers always pick the option with the highest immediate utility and do not think about their future skill. For instance, verification could keep skill at a stable level, but the worker switches to full delegation before getting there and the skill collapses. Real workers may keep verifying to stay competent or reduce their AI use slowly instead of switching all at once. So, I think the collapse mostly comes from this assumption. The proofs may be correct under the assumptions, but I am not convinced that the conclusions apply to real AI-assisted work.

Questions for authors
  1. How would the conclusions change if doing the task manually and verifying AI outputs are treated as different skills? Could verification remain effective or improve even as doing task manually skill declines?

  2. What supports the assumed balance between verification learning, manual practice, and forgetting? Section 4 shows that changing these rates can prevent collapse so how should we understand which parameter settings are realistic?

  3. Under the additional learning mechanism in Appendix E.2, could stronger AI improve long-term human skill when workflow switching is considered? Would this change the conclusion that better AI cannot increase retained skill?

  4. How sensitive are the collapse results to workers always maximizing immediate utility? Would occasional manual practice or continued verification change the conclusions?

  5. The simulations show that randomness can push workers from verification into full delegation after which they cannot recover. Could randomness also push workers toward verification or manual work and how would allowing these transitions affect the conclusions?

Reviewer_ancvOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes a theoretical framework to study whether human verification of AI-generated outputs can sustain human skills required for effective supervision over AI. The model considers a human worker who repeatedly chooses among manual work, verified delegation, and pure delegation based on current performance and costs, and choices affect future human skill through practice, error correction, and decay. This paper characterizes three possible long-run outcomes: complete skill collapse, an interior skill level sustained by verification, and competence sustained through switching between manual work and verification. It shows that verified delegation may improve current utility without preserving human skill, and that better AI cannot increase human skill level. These results are illustrated through theoretical analysis and simulations.

Critical strengths
  1. The paper tries to address a timely and important question of whether human-in-the-loop AI workflows can preserve the expertise needed for effective supervision over AI.
Critical weaknesses
  1. The paper’s main conclusions are heavily driven by restrictive and insufficiently justified assumptions about human skill development trajectory. In Section 2, three human choices are specified: 1) manual work is assumed to always improve skills, 2) pure delegation can only lead to skill decay, 3) verification provides practice but only improves learning when AI error is detected and corrected. Though making those assumptions makes theoretical analysis tractable, such restrictions are not a reflection of reality. Human skill acquisition and retention are more complex than the model assumes. Learning curves can stall or temporarily decline during practice rather than improving monotonically [1], and heavy reliance on automation can leave core skills largely intact while degrading specific cognitive components [2]. Reduced practice under delegation therefore need not imply convergence toward zero competence. And AI might facilitate learning in some settings [3] while having negligible [4,5] or negative effects in others [6,7]. The AI impact on human skill development is still evolving, and there are many heterogeneous findings across tasks and experimental designs. However, this paper makes strong explicit assumptions about how different AI-assisted workflows and human choices affect human learning, resulting in some of the conclusions being closely tied to the modelling choices. Thus, I am not convinced that the main findings represent general properties of human–AI collaboration rather than consequences of a specific and restrictive learning model.

  2. The paper lacks empirical validation of the model’s assumptions and predicted dynamics. The main results are established theoretically and demonstrated through simulations, but these simulations are generated from the same assumed model assumptions and mechanisms and therefore do not provide independent evidence that the proposed dynamics accurately describe human behavior. The paper does not include any human-annotated or longitudinal dataset tracking human skill development, workflow choices, subsequent independent performance, and any other metrics of interest. Nor does it conduct a user study to validate the proposed mechanisms or test the model's dynamic predictions. (Note: I notice that Appendix F maps an existing educational experiment onto selected model parameters, but the authors acknowledge that this is only illustrative and not an empirical validation.)

I found the abstract and introduction hard to follow. The paper introduces many technical concepts, theoretical regimes, and results before explaining the underlying model or providing sufficient intuition. Many concepts are only formally defined later in the paper, making the abstract and introduction feel confusing and scattered. I think the paper would benefit from a clearer organization of the text.

References:

[1]: Gray, Wayne D., and John K. Lindstedt. "Plateaus, dips, and leaps: Where to look for inventions and discoveries during skilled performance." Cognitive science 41, no. 7 (2017): 1838-1870.

[2]: Casner, Stephen M., Richard W. Geven, Matthias P. Recker, and Jonathan W. Schooler. "The retention of manual flying skills in the automated cockpit." Human factors 56, no. 8 (2014): 1506-1516.

[3]: Buçinca, Zana, Siddharth Swaroop, Amanda E. Paluch, Finale Doshi-Velez, and Krzysztof Z. Gajos. "Contrastive explanations that anticipate human misconceptions can improve human decision-making skills." In Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems, pp. 1-25. 2025.

[4]: Karny, Sheer, Lukas William Mayer, Jackie Ayoub, Miao Song, Haotian Su, Danyang Tian, Ehsan Moradi-Pari, and Mark Steyvers. "Learning with AI Assistance: A path to better task performance or dependence?." In Proceedings of the ACM Collective Intelligence Conference, pp. 10-17. 2024.

[5]: Yizhou Tian, Harry, Hasan Amin, and Ming Yin. "Understanding the effects of AI-assisted critical thinking on human-AI decision making." In Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems, pp. 1-30. 2026.

[6]: Wu, Shang, Catarina G. Belem, Shuyuan Fu, Mark Steyvers, and Padhraic Smyth. "How AI assistance affects human skill development: A study of learning with logic puzzles." ACM HCOMP 2026.

[7]: Liu, Grace, Brian Christian, Tsvetomira Dumbalska, Michiel A. Bakker, and Rachit Dubey. "AI assistance reduces persistence and hurts independent performance." COLM 2026.

Questions for authors

See the concerns raised in the Weaknesses section above. Given my substantial concerns about the model assumptions, I do not have additional questions at this stage. I would first like to see the authors provide stronger justification for these assumptions and/or additional empirical evidence supporting them before discussing other aspects of the paper.

2 comments

Thank you for the response and the revised manuscript. However, my main concerns remain. The core learning assumptions are unchanged in the revised version, and the relaxations discussed in the rebuttal such as plateaus and a positive default skill level are not analyzed. I therefore maintain my score.

Thank you for your follow-up. To address the two specific relaxations you identified, we have added Appendix G.4, “Robustness to residual skill and manual-learning plateaus,” to the revised manuscript. We retain endogenous workflow choice and numerically examine how the outcomes change when the baseline learning assumptions are relaxed.

1. Positive residual skill under non-use. We replace the non-use signal −Δθ-\Delta\theta with −Δ(θ−θd)-\Delta(\theta-\theta_d), allowing skill to approach a positive default level θd\theta_d under continued pure delegation. At (θa,γ,θ0)=(0.5,0.4,0.8)(\theta_a,\gamma,\theta_0)=(0.5,0.4,0.8), verification is initially optimal but gives way to pure delegation when θd=0.02\theta_d=0.02 or 0.050.05; the respective skills at T=300T=300 are 0.020000.02000 and 0.050000.05000. At θd=0.10\theta_d=0.10, verification remains optimal through T=300T=300, with skill near 0.5880.588. Table 7 in Appendix G.4 reports these comparisons. Thus, verification abandonment and substantial skill loss do not require convergence to zero, although a higher default level prevents the transition in this numerical comparison.

2. Manual-learning plateaus and declines. We modify only the manual-learning drift to permit zero or negative latent-skill change over a specified interval, leaving workflow utilities and the other learning branches unchanged. Unlike the expected-performance plateau discussed in our first response, this modification directly changes latent skill formation. At (θa,γ,θ0)=(0.3,0.3,0.65)(\theta_a,\gamma,\theta_0)=(0.3,0.3,0.65), the baseline (ζ=1\zeta=1) reaches skill near 0.9350.935 through manual–verified switching by T=300T=300. A manual-learning plateau (ζ=0\zeta=0) instead yields 0.6500.650, while a negative manual-learning signal (ζ=−0.2\zeta=-0.2) yields approximately 0.5920.592. Table 8 in Appendix G.4 reports these results. The high-skill switching outcome is therefore sensitive to manual-learning assumptions.

Importantly, the verification-to-pure-delegation collapse trajectory at (θa,γ,θ0)=(0.5,0.32,0.8)(\theta_a,\gamma,\theta_0)=(0.5,0.32,0.8) persists under the plateau modification (ζ=0\zeta=0), since manual work is never selected along that trajectory. This isolates a collapse mechanism that does not depend on positive manual learning on this path.

All comparisons use the original utility-maximizing workflow rule, with outcomes computed at T=300T=300 using step h=0.01h=0.01; repeating them with h=0.005h=0.005 changes terminal skills by less than 2×10−42\times10^{-4}.

These numerical comparisons show that verification abandonment and substantial skill loss persist with positive residual skill and, along a path that never selects manual work, under a manual-learning plateau. They also identify cases where a higher default level or weaker manual learning changes the selected workflow regime; they do not establish global robustness or empirical applicability.

Reviewer_m8hbOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper models a worker whose skill determines manual and verification reliabilities and thereby shapes workflow-specific performance losses and completion costs. The author study how skill evolves when the worker chooses among manual work, verified delegation, and pure delegation to maximize current utility by trading off performance loss against completion cost. Current skill determines workflow choice, which shapes future skill. Under verified delegation, reviewing provides verification practice, whereas generative practice occurs only when the worker detects and repairs an AI error. The authors characterize the global dynamics and identify three verification-related regimes: collapse to zero skill, an interior skill level sustained by verified delegation, and competence sustained at a manual--verified switching boundary. Crucially, verified delegation can remain optimal over a skill range while no positive skill level is sustainable. Adding verified delegation can improve current utility yet eliminate an otherwise high-skill basin. Holding learning dynamics fixed, workflow incentives alone can produce all three regimes. Better AI can further reduce retained human skill by reducing manual practice and the errors that trigger repair-based learning. The regimes persist over nontrivial learning-parameter ranges and under additional learning from correct AI outputs, while task-level randomness can drive trajectories near a stable verified equilibrium into pure delegation. Thus, keeping a human in the loop need not preserve the capability required for effective future oversight.

Critical strengths

None

Critical weaknesses
  1. The novelty of the proposed method is not clearly articulated. The authors should explicitly clarify the differences between their approach and existing state-of-the-art (SOTA) methods in terms of technical design and key contributions, highlighting the specific innovations beyond prior work.

  2. The experimental section is too limited to support the claims made in the manuscript and does not provide sufficient evidence to establish the effectiveness of the proposed method. The authors should substantially expand the experiments and provide more comprehensive analyses to validate the proposed approach.

  3. The manuscript lacks both quantitative and qualitative comparisons with existing SOTA methods. The authors should include comparisons with competitive baselines using appropriate evaluation metrics, as well as visual results to demonstrate the strengths and limitations of the proposed method. These evaluations are essential to establish its effectiveness and competitiveness.

Questions for authors
  1. The novelty of the proposed method is not clearly articulated. The authors should explicitly clarify the differences between their approach and existing state-of-the-art (SOTA) methods in terms of technical design and key contributions, highlighting the specific innovations beyond prior work.

  2. The experimental section is too limited to support the claims made in the manuscript and does not provide sufficient evidence to establish the effectiveness of the proposed method. The authors should substantially expand the experiments and provide more comprehensive analyses to validate the proposed approach.

  3. The manuscript lacks both quantitative and qualitative comparisons with existing SOTA methods. The authors should include comparisons with competitive baselines using appropriate evaluation metrics, as well as visual results to demonstrate the strengths and limitations of the proposed method. These evaluations are essential to establish its effectiveness and competitiveness.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.