Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

Selective Response Strategies for GenAI

T0 review · 4 major / 5 minor · reviewed 2026-08-09 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read A GenAI service can increase its long-term revenue and user welfare by deliberately withholding or degrading answers on emerging topics, because the users it turns away generate training data on human forums.

desk verdict Genuinely new idea and a solid dynamic model, but the regulatory centerpiece (Theorem 5.1) is false as stated, so the paper needs major revision before it can be trusted. read the letter →

arxiv 2502.00729 v2 pith:NDRCCX32 submitted 2025-02-02 cs.AI cs.GTcs.SI

Boaz Taitler , Omer Ben-Porat This is my paper · ORCID
classification cs.AIcs.GTcs.SI
keywords selectiveresponsegenerativeAIdataaccumulationuserwelfarerevenuemaximizationonlineQ&Aforumsgametheory
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tries to establish that a generative-AI platform should sometimes not answer: on emerging topics it withholds answers or offers deliberately lower-quality ones, steering users to human Q&A forums. The paper argues this short-term sacrifice compounds: each denied user who posts on a forum creates a unit of training data, which improves GenAI accuracy, which attracts more users next round, which generates still more data. In the model, reducing the answer level in any single round strictly increases cumulative data in all later rounds and weakly increases the share of users who choose GenAI; there are instances where this dominates always-responding on both revenue and welfare. The paper also gives an approximately optimal revenue-maximizing algorithm and regulatory conditions under which one forced round of selective response raises welfare with bounded revenue loss.

What carries the argument

The engine is the one-round data-update map: starting with cumulative data $D$ and response level $x$, next-round data is $D + (1 - p(D,x))$, where $p(D,x)$ is the softmax share of users who receive an answer from GenAI. The paper proves that under Assumption 2.1 this map is strictly increasing in $D$; since a lower $x$ in a round means more users go to the human forum, the map turns one denied query into strictly more data in every later round. Iterating that monotonicity is Theorem 4.1, and all later results—the dominance comparison, the approximation guarantees, and the regulator's thresholds—ride on it.

What would settle it

Instrument a deployed GenAI service that withholds answers on a new topic and measure the fraction of denied users who actually post on the partner forum within a week; if that fraction is a constant $q<1$, the recurrence becomes $D_{t+1}=D_t+q(1-p_t)$, the strict data gain in Theorem 4.1 shrinks proportionally, and for small $q$ one can exhibit parameter ranges in which no selective strategy beats always-responding.

Watch

Extended reading notes

Core claim

The core claim is formalized as Theorem 4.1: fix any strategy $x$ and any round $\tau$; a $\tau$-selective modification that lowers the answered fraction in round $\tau$ yields strictly more cumulative data $D_t$ in every later round $t>\tau$, and weakly higher user proportions $p_t$, with equality only in degenerate zero-response states. Because $D_t$ drives GenAI quality through the accuracy function $a(D_t)$, and user choices respond to quality through the softmax, this data advantage is self-reinforcing. The theorem is the reason selective response can dominate the always-responding strategy on revenue and welfare simultaneously (Observation 3.1), why the price of always responding is unbounded in revenue and arbitrarily close to 2 in welfare, and why both the revenue-maximization and welfare-constrained algorithms work.

Load-bearing premise

The model assumes that every user who is denied an answer posts their question on the human forum, contributing one unit of data; if denied users instead leave or simply abandon the question, the compounding data engine that makes selective response beneficial disappears.

Editorial extensions

If this is right

  • Reducing the answer level in any single round strictly increases cumulative data in all later rounds and weakly raises future user proportions, so a sufficiently patient GenAI can trade a little current engagement for a permanently larger future user base.
  • The price of always responding is unbounded in revenue: for any $M$ there is an instance where the best selective strategy earns more than $M$ times the always-responding revenue.
  • Welfare under the best selective strategy can be made arbitrarily close to double the always-responding welfare.
  • An approximately optimal revenue-maximizing strategy can be found by dynamic programming on a discretized data state, with additive error $\epsilon L_r T^2$ and runtime $O(T^2|A|/\epsilon)$; a grid over $[0,1]$ extends the guarantee to continuous strategies.
  • A regulator who sees only past actions can force a one-round answer reduction that raises total welfare whenever current answered-utility is below one threshold and the platform commits to future utility above a second threshold, with the revenue loss bounded.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A testable corollary the paper does not spell out: deployment logs of a selective-response service should show answer rates starting near zero on a new topic, then jumping upward as forum data accumulates; the timing of that jump should track the accuracy function's Lipschitz constant.
  • The model equates saying nothing with giving a lower-quality answer; separating the two would change the revenue calculus, because a visibly wrong answer can damage trust in a way silence does not, and an optimal policy would likely favor explicit disclaimers over incorrect content.
  • The same control problem arises inside a single organization: an AI assistant that defers hard queries to a human annotation queue is effectively choosing a selective-response schedule, so the paper's monotonicity result suggests how to tune deferral thresholds under a data-accumulation objective.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper studies a T-round model in which a GenAI platform chooses a selective response level x_t in [0,1] each round, users choose between GenAI and a human forum via a softmax rule, and users who do not receive an answer post on the forum and thereby generate training data for GenAI. The authors claim that selective response can Pareto-dominate always responding (Observation 3.1), that a one-round reduction in response strictly increases all future data and weakly increases future user shares (Theorem 4.1), and they propose approximation algorithms for revenue maximization (Theorems 4.4 and 4.6) and for welfare-constrained revenue maximization (Theorem 4.7). They then state a threshold theorem (Theorem 5.1) about the welfare effect of a single-round selective modification, and use it to derive regulatory sufficient conditions (Corollary 6.1) and a revenue-gap bound (Corollary 6.2). The paper concludes with a discussion of modeling limitations and simulations over cutoff strategies.

Significance. The model is original and the core monotonicity result, Theorem 4.1, is a nontrivial formal observation that could be useful for later work on GenAI-data ecosystems. The paper is also commendably explicit about its key assumption that unanswered users post on the forum, and the ASR algorithm is presented with concrete approximation and runtime guarantees. However, the advertised regulatory and constrained-optimization contributions rest on results that are false or invalid as stated: Theorem 5.1(2) is contradicted by zero-response rounds, Corollary 6.2 has a sign error, and the reduction behind Theorem 4.7 uses negative rewards in a framework that requires nonnegative rewards. These are load-bearing errors, not presentation issues, and they undermine the paper's central claims about welfare regulation and constrained revenue maximization.

major comments (4)
  1. [Section 5, Theorem 5.1(2) and Appendix D] Part 2 of Theorem 5.1 is false for any positive threshold B. Take any strategy x with x_t = 0 for some t > \tau, as in Example 2.2, and any \tau-selective modification x^\tau. Since x^\tau_t = x_t = 0, we have w^g_t(x^\tau) = a(D_t(x^\tau)) * 0 = 0 < B for every B > 0, but w_t(x^\tau) = w_t(x) = w_s, contradicting the strict inequality. The proof in Appendix D does not repair this: its derivative argument establishes negativity of dh/dy only for negative accuracy values (the auxiliary function is analyzed as y to -infinity), whereas accuracy is confined to [0,1], and the proof never actually constructs the threshold B. The only way to avoid the counterexample is B <= 0, but then Part 1's hypothesis w^g_\tau(x) < B is impossible since w^g_\tau(x) >= 0, making Corollary 6.1 vacuous. Consequently, the regulatory sufficient conditions in Section 6 do not follow from the stated theorem.
  2. [Section 6.2, Corollary 6.2 and Appendix E.1] The displayed bound in Corollary 6.2 contains a sign error. The proof's derivation produces L_r(p_\tau(x) - p_\tau(x^\tau)) times the sum of gamma^t from t = \tau+1 to T, which is nonnegative and is bounded above by L_r gamma^\tau (p_\tau(x) - p_\tau(x^\tau))/(1 - gamma). The corollary instead states the second term as L_r gamma^\tau (p_\tau(x^\tau) - p_\tau(x))/(1 - gamma), which is negative for any \tau-selective modification. As printed, the right-hand side is negative, while U(x^\tau) - U(x) can be positive; for instance, with T = 2, a(D) = 0.4D, w_s = 0.5, beta = 10, r(p) = p^2, gamma = 1, and x = (1,1), the \tau = 1 modification x^\tau = (0,1) satisfies Assumption 2.1 and gives U(x^\tau) > U(x). The revenue-gap bound in Section 6.2 is therefore false as stated.
  3. [Section 4.3 and Appendix C.3, Theorem 4.7] The reduction of the welfare-constrained problem to PARS-MDP is invalid as written. The second reward function is defined as W(s^d_t, y) = floor(y q(d,y)(a(d) - w_s))_epsilon, which takes negative values whenever a(d) < w_s, but the PARS-MDP framework used in the proof (and in reference [3]) requires nonnegative rewards R_A, R_P : S x A to R_{\ge 0}. Likewise, the budget B is defined through W_1 = W - T w_s, which can be negative, while the PARS-MDP budget is required to be nonnegative. A constant shift of the welfare reward would be needed to make the reduction valid, but no such shift is introduced or analyzed. Hence the constraint encoding and the guarantees of Theorem 4.7 are not established.
  4. [Section 6.1, Corollary 6.1] Corollary 6.1 is stated without a proof, and it does not follow from Theorem 5.1 as stated. Under the commitment condition, the 'gray' rounds covered by Part 2 are impossible: for any t > \tau with w^g_t(x) > 0, the commitment gives w^g_t(x) > C >= B, and since w^g_t(x^\tau) >= w^g_t(x), Part 2's hypothesis w^g_t(x^\tau) < B cannot hold; for w^g_t(x) = 0, we have x_t = 0 and the two welfare values coincide. Thus the intended 'positive area exceeds negative area' argument is not captured by the theorem, and the corollary needs a separate proof. The appendix does not supply one, so the Section 6.1 welfare-improvement claim is currently unsupported.
minor comments (5)
  1. [Abstract and Introduction] The paper repeatedly claims to derive 'sufficient and necessary conditions' for selective response to improve welfare, but Section 6 contains only sufficient conditions and no necessary condition is stated or proved.
  2. [Appendix C.2, Theorem 4.4 proof] In the proof of Theorem 4.4, U_t(x) is defined as the sum from i=t to T of gamma^{i-t} p_i(x), but the revenue objective is U(x) = sum_t gamma^t r(p_t(x)); the proof should use r(p_i) for consistency.
  3. [Section 6.2, Corollary 6.2] Corollary 6.2 does not state an assumption that gamma < 1, yet the displayed bound contains a factor 1/(1-gamma) and is undefined at gamma = 1, a case that the paper considers elsewhere.
  4. [Appendix B.1, Proposition 3.2 proof] In the proof of Proposition 3.2, the sigmoid function is defined using q(T-1,1) and q((T-1)/2,1), but the subsequent displayed expressions sometimes write q(T-1) without the second argument, which makes the construction hard to follow.
  5. [Appendix C.3, Step 4] The text refers to the 'PARSE-MDP problem' in one place and to 'PARS-MDP' elsewhere; the terminology should be made consistent with reference [3].

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: central results are derived from the model's definitions and assumptions; the only self-citation is to a general prior-work algorithm and does not carry the main claim.

full rationale

The paper's central selective-response claim (Theorem 4.1) is proved internally from the model's primitives, not from a fitted value or from the authors' prior conclusions. Section 2 defines D_t(x)=D_{t-1}(x)+(1-p_{t-1}(x)), p_t(x)=x_t σ_t(x), and wg_t(x)=a(D_t(x))·x_t; the proof of Theorem 4.1 constructs f(D,x)=D-Q(D,x), proves d f/dD>0 under Assumption 2.1, and iterates the data-dominance relation. No parameter is fitted to data and no 'prediction' is identified with an input by construction. Observation 3.1, Proposition 3.2, and Proposition 3.3 are supported by explicit instances and self-contained bounds in Appendix B. The ASR guarantees (Theorems 4.4 and 4.6) are derived within the paper via Lemmas C.3-C.10. The only self-citation entering a proof is [3] (PARS-MDP) in the welfare-constrained algorithm of Section 4.3 / Appendix C.3; that cited work is a general MDP reward-shaping theorem with its own stated assumptions and does not assume the selective-response conclusion, so it is independent support for a secondary algorithmic result rather than a circular load-bearing step. The paper also candidly discloses limitations (Section 7: reliance on Assumption 2.1, the open closed-form for threshold B, and the restricted continuous-extension technique), none of which indicate circularity. The externally supplied counterexample to Theorem 5.1, if correct, concerns internal correctness of a welfare comparison, not a reduction of the result to its inputs; correctness risk is outside the circularity pass. Accordingly, no circular step is identified and the score is 0.

Assumptions & free parameters 0 free parameters · 5 assumptions · 0 invented entities

The model introduces no fitted free parameters; the accuracy function, revenue function, discount factor, softmax temperature, and forum utility are exogenous inputs. The central theorems rely on stated behavioral assumptions (softmax choice, all unanswered users posting on the forum) and on a technical Lipschitz condition. The welfare-constrained revenue result additionally invokes a prior algorithm by the same authors.

assumptions (5)
  • domain assumption Users choose between GenAI and Forum according to a softmax function σ_t(x) = e^{β w^g_t} / (e^{β w^g_t} + e^{β w_s})
    Section 2 defines user decisions as softmax; this is a modeling choice, not derived from empirical data.
  • domain assumption All users who do not receive an answer from GenAI post their question on Forum, contributing 1 unit of data
    Section 2, paragraph 'Users': 'We assume the former.' This is load-bearing for the data accumulation dynamic.
  • ad hoc to paper Accuracy function a satisfies Assumption 2.1: a is strictly increasing and L_a-Lipschitz with L_a ≤ 4/β
    Assumption 2.1 is introduced to guarantee monotonicity of data generation (Proposition 4.2); it is a technical restriction chosen for the proof.
  • domain assumption Revenue function r is non-decreasing and L_r-Lipschitz
    Section 2 defines r; used throughout for revenue bounds.
  • standard math The PARS-MDP algorithm of Ben-Porat et al. [3] solves Problem (P5) in the stated time
    Theorem 4.7 depends on this prior result by the same authors; the proof is not reproduced in this paper.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Selective Response Strategies for GenAI." pith.science (2026). https://pith.science/paper/NDRCCX32

@misc{pith2026250200729,
  author       = {Pith},
  title        = {Pith review of: Selective Response Strategies for GenAI},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NDRCCX32}},
  note         = {Machine review of arXiv:2502.00729}
}
read the original abstract

The rise of Generative AI (GenAI) has significantly impacted human-based forums like Stack Overflow, which are essential for generating high-quality data. This creates a negative feedback loop, hindering the development of GenAI systems, which rely on such data to provide accurate responses. In this paper, we provide a possible remedy: A novel strategy we call selective response. Selective response implies that GenAI could strategically provide inaccurate (or conservative) responses to queries involving emerging topics and novel technologies, thereby driving users to use human-based forums like Stack Overflow. We show that selective response can potentially have a compounding effect on the data generation process, increasing both GenAI's revenue and user welfare in the long term. From an algorithmic perspective, we propose an approximately optimal approach to maximize GenAI's revenue under social welfare constraints. From a regulatory perspective, we derive sufficient and necessary conditions for selective response to improve welfare improvements.

Figures

Figures reproduced from arXiv: 2502.00729 by the authors.

Figure 1
Figure 1. A visualization for Example 2.2. The blue (circle) curve shows the proportion of [PITH_FULL_IMAGE:figures/full_fig_p007_1.png] view at source ↗
Figure 2
Figure 2. Illustrating Theorem 5.1. The left figure illustrates GenAI’s expected utility vs [PITH_FULL_IMAGE:figures/full_fig_p013_2.png] view at source ↗
Figure 3
Figure 3. Example of a constructed graph with a discretization factor of [PITH_FULL_IMAGE:figures/full_fig_p035_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Revenue and welfare difference between the GenAI’s optimal cutoff selective [PITH_FULL_IMAGE:figures/full_fig_p053_4.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Data Sharing with a Generative AI Competitor

    cs.GT 2025-05 conditional novelty 5.0 of 10

    In a two-stage data-sharing game, the unique equilibrium is either that the firm shares just enough data to stop the platform buying expert data, or that the firm shares an amount that maximizes its payoff while the p...

Reference graph

Works this paper leans on

63 extracted references · 54 canonical work pages · cited by 1 Pith paper

  1. [3]

    Ben-Porat, Y

    O. Ben-Porat, Y. Mansour, M. Moshkovitz, and B. Taitler. Principal-agent reward shaping in mdps. In Proceedings of the AAAI Conference on Artificial Intelligence , volume 38, pages 9502–9510, 2024

    work page 2024

  2. [1]

    Babichenko, I

    Y. Babichenko, I. Talgam-Cohen, H. Xu, and K. Zabarnyi. Algorithmic cheap talk. In Proceedings of the 25th ACM Conference on Economics and Computation , EC ’24, page 5–6. Association for Computing Machinery, 2024

    work page 2024

  3. [2]

    Bailey, D

    M. Bailey, D. Johnston, T. Kuchler, J. Stroebel, and A. Wong. Peer effects in product adoption. American Economic Journal: Applied Economics , 14(3):488–526, 2022

    work page 2022

  4. [4]

    Bergemann and A

    D. Bergemann and A. Bonatti. Data, competition, and digital platforms. American Economic Review, 114(8):2553–2595, 2024. 16

    work page 2024

  5. [5]

    Bergemann and S

    D. Bergemann and S. Morris. Information design: A unified perspective. Journal of Economic Literature, 57(1):44–95, 2019

    work page 2019

  6. [6]

    Bergemann, B

    D. Bergemann, B. Brooks, and S. Morris. The limits of price discrimination. American Economic Review, 105(3):921–957, 2015

    work page 2015

  7. [7]

    Burtch, D

    G. Burtch, D. Lee, and Z. Chen. The consequences of generative ai for online knowledge communities. Scientific Reports, 14(1):10413, 2024

    work page 2024

  8. [8]

    Y. Chow, G. Tennenholtz, I. Gur, V. Zhuang, B. Dai, A. Kumar, R. Agarwal, S. Thia- garajan, C. Boutilier, and A. Faust. Inference-aware fine-tuning for best-of-n sampling in large language models. In The Thirteenth International Conference on Learning Representations, 2025

    work page 2025

Show all 63 references
  1. [9]

    Conitzer, R

    V. Conitzer, R. Freedman, J. Heitzig, W. H. Holliday, B. M. Jacobs, N. Lambert, M. Moss´ e, E. Pacuit, S. Russell, H. Schoelkopf, E. Tewolde, and W. S. Zwicker. Social choice for AI alignment: Dealing with diverse human feedback. CoRR, abs/2404.10271,

    arXiv

  2. [10]

    J. W. Crandall, M. Oudah, Tennom, F. Ishowo-Oloko, S. Abdallah, J.-F. Bonnefon, M. Cebrian, A. Shariff, M. A. Goodrich, and I. Rahwan. Cooperating with machines. Nature communications, 9(1):233, 2018

    2018

  3. [11]

    V. P. Crawford and J. Sobel. Strategic information transmission. Econometrica: Journal of the Econometric Society , pages 1431–1451, 1982

    1982

  4. [12]

    S. Dean, E. Dong, M. Jagadeesan, and L. Leqi. Accounting for ai and users shaping one another: The role of mathematical models. arXiv preprint arXiv:2404.12366 , 2024

    2024 arXiv

  5. [13]

    R. M. del Rio-Chanona, N. Laurentsyeva, and J. Wachs. Large language models reduce public knowledge sharing on online q&a platforms. PNAS nexus , 3(9):pgae400, 2024

    2024

  6. [14]

    S. A. Esmaeili, K. Bhawalkar, Z. Feng, D. Wang, and H. Xu. How to strategize human content creation in the era of genai? arXiv preprint arXiv:2406.05187 , 2024

    2024 arXiv

  7. [15]

    Feldman, R

    M. Feldman, R. Meir, and M. Tennenholtz. Competition in the presence of social networks: how many service providers maximize welfare? In International Conference on Web and Internet Economics , pages 174–187. Springer, 2013

    2013

  8. [16]

    Frieder, L

    S. Frieder, L. Pinchetti, R.-R. Griffiths, T. Salvatori, T. Lukasiewicz, P. Petersen, and J. Berner. Mathematical capabilities of chatgpt. Advances in Neural Information Processing Systems, 36, 2024

    2024

  9. [17]

    M. R. Garey and D. S. Johnson. Computers and intractability , volume 174. freeman San Francisco, 1979. 17

    1979

  10. [18]

    G. Hardin. The tragedy of the commons: the population problem has no technical solution; it requires a fundamental extension in morality. science, 162(3859):1243–1248, 1968

    1968

  11. [19]

    Hemmer, L

    P. Hemmer, L. Thede, M. V¨ ossing, J. Jakubik, and N. K¨ uhl. Learning to defer with limited expert predictions. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 37, pages 6002–6011, 2023

    2023

  12. [20]

    Jagadeesan, M

    M. Jagadeesan, M. I. Jordan, and N. Haghtalab. Competition, alignment, and equilibria in digital marketplaces. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 37, pages 5689–5696, 2023

    2023

  13. [21]

    Z. Ji, N. Lee, R. Frieske, T. Yu, D. Su, Y. Xu, E. Ishii, Y. J. Bang, A. Madotto, and P. Fung. Survey of hallucination in natural language generation. ACM Computing Surveys, 55(12):1–38, 2023

    2023

  14. [22]

    Zhang, Y

    junyou li, Q. Zhang, Y. Yu, Q. FU, and D. Ye. More agents is all you need.Transactions on Machine Learning Research, 2024. ISSN 2835-8856. URL https://openreview.n et/forum?id=bgzUSZ8aeg

    2024

  15. [23]

    A. T. Kalai and S. S. Vempala. Calibrated language models must hallucinate. In Proceedings of the 56th Annual ACM Symposium on Theory of Computing , pages 160–171, 2024

    2024

  16. [24]

    Karle, M

    H. Karle, M. Peitz, and M. Reisinger. Segmentation versus agglomeration: Competition between platforms with competitive sellers. Journal of Political Economy , 128(6): 2329–2374, 2020

    2020

  17. [25]

    Kasneci, K

    E. Kasneci, K. Seßler, S. K¨ uchemann, M. Bannert, D. Dementieva, F. Fischer, U. Gasser, G. Groh, S. G¨ unnemann, E. H¨ ullermeier, et al. Chatgpt for good? on opportunities and challenges of large language models for education. Learning and individual differences , 103:102274, 2023

    2023

  18. [26]

    M. L. Katz and C. Shapiro. Network externalities, competition, and compatibility. The American economic review, 75(3):424–440, 1985

    1985

  19. [27]

    Keinan and O

    G. Keinan and O. Ben-Porat. Strategic content creation in the age of genai: To share or not to share? arXiv preprint arXiv:2505.16358 , 2025

    2025

  20. [28]

    Koco´ n, I

    J. Koco´ n, I. Cichecki, O. Kaszyca, M. Kochanek, D. Szyd lo, J. Baran, J. Bielaniewicz, M. Gruza, A. Janz, K. Kanclerz, et al. Chatgpt: Jack of all trades, master of none. Information Fusion, page 101861, 2023

    2023

  21. [29]

    Koutsoupias and C

    E. Koutsoupias and C. Papadimitriou. Worst-case equilibria. In Annual symposium on theoretical aspects of computer science, pages 404–413. Springer, 1999. 18

    1999

  22. [30]

    Laufer, J

    B. Laufer, J. Kleinberg, and H. Heidari. Fine-tuning games: Bargaining and adaptation for general-purpose models. In Proceedings of the ACM on Web Conference 2024 , pages 66–76, 2024

    2024

  23. [31]

    K. Li, O. Patel, F. Vi´ egas, H. Pfister, and M. Wattenberg. Inference-time intervention: Eliciting truthful answers from a language model. Advances in Neural Information Processing Systems, 36, 2024

    2024

  24. [32]

    Li and K

    X. Li and K. Kim. Impacts of generative ai on user contributions: evidence from a coding q &a platform. Marketing Letters, pages 1–15, 2024

    2024

  25. [33]

    J. Liu, C. S. Xia, Y. Wang, and L. Zhang. Is your code generated by chatgpt really correct? rigorous evaluation of large language models for code generation. Advances in Neural Information Processing Systems , 36, 2024

    2024

  26. [34]

    Y. L. Lo, C. S. de Witt, S. Sokota, J. N. Foerster, and S. Whiteson. Cheap talk discovery and utilization in multi-agent reinforcement learning. In The Eleventh International Conference on Learning Representations, 2023

    2023

  27. [35]

    C. Lu, T. Willi, A. Letcher, and J. N. Foerster. Adversarial cheap talk. In International Conference on Machine Learning , pages 22917–22941. PMLR, 2023

    2023

  28. [36]

    D. P. McIntyre and A. Srinivasan. Networks, platforms, and strategy: Emerging views and next steps. Strategic management journal , 38(1):141–160, 2017

    2017

  29. [37]

    M. M. Mello and N. Guha. Chatgpt and physicians’ malpractice risk. JAMA Health Forum, 4(5):e231938, May 2023. doi: 10.1001/jamahealthforum.2023.1938. URL https://jamanetwork.com/journals/jama-health-forum/fullarticle/2805334

    2023

  30. [38]

    P. R. Milgrom. Good news and bad news: Representation theorems and applications. The Bell Journal of Economics , pages 380–391, 1981

    1981

  31. [39]

    Mozannar and D

    H. Mozannar and D. Sontag. Consistent estimators for learning to defer to an expert. In International conference on machine learning , pages 7076–7087. PMLR, 2020

    2020

  32. [40]

    Raghavan

    M. Raghavan. Competition and diversity in generative ai. arXiv preprint arXiv:2412.08610, 2024

    2024 arXiv

  33. [41]

    Rietveld and M

    J. Rietveld and M. A. Schilling. Platform competition: A systematic and interdisci- plinary review of the literature. Journal of Management , 47(6):1528–1563, 2021

    2021

  34. [42]

    Roughgarden

    T. Roughgarden. Selfish routing and the price of anarchy . MIT press, 2005

    2005

  35. [43]

    R. Shin. Humiliated lawyers fined $5,000 for submitting chatgpt hallucinations in court: ‘i heard about this new site, which i falsely assumed was, like, a super search engine’, June 2023. URL https://fortune.com/2023/06/23/lawyers-fined-fil ing-chatgpt-hallucinations-in-court/ . 19

    2023

  36. [44]

    H. Sun, Y. Chen, S. Wang, W. Chen, and X. Deng. Mechanism design for LLM fine- tuning with multiple reward models. In Pluralistic Alignment Workshop at NeurIPS 2024, 2024

    2024

  37. [45]

    Taitler and O

    B. Taitler and O. Ben-Porat. Braess’s paradox of generative ai. In Proceedings of the AAAI Conference on Artificial Intelligence , volume 39, pages 14139–14147, 2025

    2025

  38. [46]

    G. Tullock. Efficient rent seeking. In J. M. Buchanan, R. D. Tollison, and G. Tullock, editors, Toward a Theory of the Rent-Seeking Society , pages 97–112. Texas A and M University Press, College Station, 1980

    1980

  39. [47]

    F. Yao, C. Li, D. Nekipelov, H. Wang, and H. Xu. Human vs. generative ai in content creation competition: symbiosis or conflict? In Proceedings of the 41st International Conference on Machine Learning , ICML’24. JMLR.org, 2024. 20 A Definitions and Notations We first define th...

    2024

  40. [49]

    • U (x) > 2.483

    Pareto dominance This is shown in Example 2.2, for which it holds that • U (¯x) < 2.356. • U (x) > 2.483. • W (¯x) < 5.73. • W (x) > 6.2

  41. [50]

    We calculate the revenue and the social welfare induced by ¯x by calculating the proportions for every t ∈ [T ]

    Decreases revenue and increases welfareLet T = 5 and consider the instance a(D) = 1 − e−0.4D, γ = 1, β = 3, ws = 0.7 and r(p) = p. We calculate the revenue and the social welfare induced by ¯x by calculating the proportions for every t ∈ [T ]. Therefore, the induced revenue is...

  42. [51]

    21 We denote x the strategy that satisfies xt = ( 1 t = T 0 Otherwise

    Increases revenue and decreases welfareLet T = 5 and consider the instance a(D) = 1 − e−0.4D, γ = 1, β = 3, ws = 0.1 and r(p) is the step function defined as r(p) = ( 1 p ≥ q(4, 1) 0 Otherwise . 21 We denote x the strategy that satisfies xt = ( 1 t = T 0 Otherwise . Notice tha...

  43. [52]

    23 Therefore, by the intermediate value theorem, there exists T0 such that h(T0) = 1 2−ε

    lim T →∞ h(T ) → 0.5. 23 Therefore, by the intermediate value theorem, there exists T0 such that h(T0) = 1 2−ε . Furthermore, dh dT = − β T 2 1 1 + e β T − β T 1 1 + e β T 1 − 1 1 + e β T ! β T 2 + 1 1 + e β T 1 − 1 1 + e β T ! β T 2 = − β T 1 1 + e β T 1 − 1 1 + e β T ! β T 2...

  44. [53]

    x = (0.04, 0.97, 1). Notice that we do not consider schemes where x3 = 0 since for any such scheme, the scheme which is identical at round t = 1, t = 2 and plays x3 = 1 induces higher revenue; therefore, the revenue difference between x and the other schemes is as follows: • U...

  45. [54]

    For a deterministic MDP, a policy is equivalent to a path 35 τ , which in our case is a sequence of T edges starting from state s0 1 and leading to a state in ST +1

    We define policy π : S → A to be the mapping between each state and the action GenAI should take in that state. For a deterministic MDP, a policy is equivalent to a path 35 τ , which in our case is a sequence of T edges starting from state s0 1 and leading to a state in ST +1....

  46. [55]

    ≥ Ut(x⋆) − εM T2, • V W (πx⋆ , s0

  47. [56]

    We finished Step 2 and now move on to develop the machinery to find the selective response strategy that gives us the guarantees of our theorem

    ≥ Wt(x⋆) − (La + 1)εT 2. We finished Step 2 and now move on to develop the machinery to find the selective response strategy that gives us the guarantees of our theorem. Step 4. We define the Weight-Constrained Shortest Path (WCSSP) [17]. Given a weighted graph G = (V, E) with...

  48. [57]

    The PARSE-MDP problem is defined over an MDP with two reward functions RA, RP : S × A → R≥0 and a budget B ∈ R≥0

    with a deterministic transition function. The PARSE-MDP problem is defined over an MDP with two reward functions RA, RP : S × A → R≥0 and a budget B ∈ R≥0. The goal is to construct a new reward function RB : S × A → R≥0 such that the total rewards over the whole MDP is less th...

  49. [58]

    Therefore, it holds that U (˜x) > VG(π ˜x, s0

    = TX t=1 ⌊r(yq(dt, ˜xt))⌋ε ≤ TX t=1 r(yq(dt, ˜xt)) ≤ TX t=1 r(yq(Dt(˜x), ˜xt)) = TX t=1 r(pt(˜x)) = U (˜x). Therefore, it holds that U (˜x) > VG(π ˜x, s0

  50. [59]

    We move on to evaluate the welfare

    > U(x⋆) − εM T2. We move on to evaluate the welfare. Notice that the welfare is not monotonic in pt. Instead of using our previous technique, we use Lemma C.11 and get that W (˜x) ≥ V W (˜x, s0

  51. [60]

    This completes the proof of Theorem 4.7

    − εT 2(La + 1) ≥ W 1 − εT 2(La + 1) − εT 2(La + 1) = W 1 − 2εT 2(La + 1). This completes the proof of Theorem 4.7. Proof of Lemma C.11. We begin by showing that there cannot be a large gap between the data accumulated in the original problem and the data according to our MDP f...

  52. [61]

    From Proposition C.1, for every t > τit holds that dt(xτ ) > dt(x) and therefore wg t (xτ ) ≥ wg t (x) ≥ C

    If wg t (x) ≥ C. From Proposition C.1, for every t > τit holds that dt(xτ ) > dt(x) and therefore wg t (xτ ) ≥ wg t (x) ≥ C. Notice that dh(y, x) dy = x2β 1 1 + eβ(ws−xy) 1 1 + e−β(ws−xy) (y − ws) + x 1 1 + eβ(ws−xy) ≥ x2β 1 1 + eβ(ws−xy) 1 1 + e−β(ws−xy) (xy − ws) + x 1 1 + e...

  53. [62]

    For the next two properties, we show that there exists y0 ∈ R such that for every x >0 it holds that dh(y0,x) dy < 0. First, we rewrite dh(y,x) dy : dh(y, x) dy = x2β 1 1 + eβ(ws−xy) 1 1 + e−β(ws−xy) (y − ws) + x 1 1 + eβ(ws−xy) = x 1 1 + eβ(ws−xy) 1 + xβ 1 1 + e−β(ws−xy) (y −...

  54. [63]

    This completes the proof of Theorem 5.1

    The result h(a(Dt(xτ )), xτ τ ) > h(a(Dt(x)), xτ ) follows immediately from the previous arguement. This completes the proof of Theorem 5.1. Proof of Lemma D.1. Fix x ∈ [0, 1] and we denote ˜q(y) = eβy eβy +eβws . Therefore, g(y) can be written as g(y) = xq(y)y + (1 − xq(y))ws...

  55. [2024]

    URL https://doi.org/10.48550/arXiv.2 404.10271

    doi: 10.48550/ARXIV.2404.10271. URL https://doi.org/10.48550/arXiv.2 404.10271

    doi arXiv

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.