The Cost of Balanced Training-Data Production in an Online Data Market

Augustin Chaintreau; Juba Ziani; Roland Maio

REVIEW 3 major objections 4 minor 62 references

The Cost of Balanced Training-Data Production in an Online Data Market

T0 review · 3 major / 4 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 In a model of an online data market, requiring every seller to produce demographically balanced data can destroy small markets entirely, but once any single group's economic value grows without bound, the relative cost of the requirement…

desk verdict Clear load-bearing algebra error in the participation threshold; the model and the amortization claim are worth engaging, but the paper as written does not support its central theorems. read the letter →

arxiv 2501.19294 v1 pith:CMEJMKNV submitted 2025-01-31 cs.GT

Augustin Chaintreau , Roland Maio , Juba Ziani This is my paper · ORCID

classification cs.GT MSC 91A1091B26

keywords onlinedatamarketsfairnessinterventiondemographicbalancecostofmarketequilibriaNashequilibriumShapleyvaluelearningcurves

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 asks whether a for-profit marketplace can impose a fairness rule on training-data production without destroying itself. In a stylized game-theoretic model, the authors compare two scenarios: sellers freely choose how many samples of each demographic group to produce, and a marketplace that accepts only datasets matching a target demographic balance. The answer is that market size decides everything. In small markets, the rule can backfire completely: sellers who would produce data in the unconstrained market stop producing, so every agent loses its entire baseline utility and the intended beneficiary groups get no data at all. In large markets, if at least one group's extractable economic value grows without bound as buyers enter, the intervention-to-baseline utility ratio tends to 1 for the marketplace and every seller and is at least 1 in the limit for every buyer, so the relative cost of fairness vanishes.

What carries the argument

The analysis runs on two objects. The first is the learning curve $G(x)=(D-\alpha x^{-\beta})_+$, which gives diminishing returns and an accuracy ceiling, so each extra sample contributes less than the last. The second is the potential economic value $C_g$, defined as the reserve price times the number of buyers bidding at least that price; the marketplace maximizes $C_g$ per group, and sellers split revenue by Shapley value proportional to their marginal contribution to $G$. In the baseline, groups decouple and the aggregate sample count for group $g$ is $x_g = (C_g/(c_g \alpha \beta))^{1/(\beta+1)}$. Under the $\gamma$-balance intervention, each seller's choice collapses to one scalar $B$ (total samples), and the coupled production level becomes a power mean of the $C_g$'s weighted by $\gamma^{-\beta}$; comparing utilities then reduces to comparing linear terms in the $C_g$'s with sublinear ones, which is why unbounded growth in any single $C_g$ drives every utility ratio to 1.

What would settle it

Compute, in the quasi-symmetric model with one seller and two groups, the finite-$n$ seller profit ratio for a sequence of buyer values where $\max_p C_g \to \infty$; if the ratio fails to approach 1 while the assumption holds, Theorem 6.1 is false. A structurally sharper check is to search for a parameter regime satisfying the assumption where the intervention-scenario aggregate production $B$ stays bounded as $n\to\infty$, which would contradict Lemma 4.2's growth formula and break the limit.

Watch

Extended reading notes

Core claim

The paper's central claim is that under a quasi-symmetric market model—buyers share one learning curve $G(x)=(D-\alpha x^{-\beta})_+$, sellers share one cost vector $c$—a demographic-balance intervention has a cost that depends entirely on market scale. It proves that for every target vector $\gamma$ there exists a market where the intervention backfires (Theorem 5.1), and that in markets that fully form at baseline only the uniform target is guaranteed never to backfire (Theorems 5.2 and 5.3). The headline result is asymptotic: if some group's potential economic value $C_g = p_g \cdot |\{i: v_{i,g} \ge p_g\}|$ can be made arbitrarily large as the number of buyers $n$ grows, then the utility ratios between intervention and baseline converge to $1$ for the marketplace and for every seller, and to at least $1$ for every buyer (Theorem 6.1, equations (39)–(41)). In words, market growth can amortize the cost of balanced data production until it is a negligible fraction of each agent's utility.

Load-bearing premise

The load-bearing premise is that the marketplace rejects any seller whose dataset is not exactly $\gamma$-demographically balanced, forcing every seller to produce every group in the same proportions; if balance were required only in aggregate, or if transfer learning between groups were allowed, the cost of fairness could be much smaller and the small-market backfire might disappear (as the paper itself notes in its limitations).

Editorial extensions

If this is right

A marketplace that requires balanced data can remain competitive as the market grows: the required balance does not shrink any agent's utility in the limit.
In small markets, imposing any non-uniform balance target risks total market failure, so an ethical marketplace may need to start with the uniform target and wait for demand to scale.
Because the buyer surplus ratio is at least 1 in the limit, the intervention can create a positive externality: some buyers are strictly better off, not merely unharmed.
The result explains a window of viability: ethical data firms may be economically feasible now, when demand is large, even if the same intervention would have been lethal in earlier, smaller markets.
Theorems 5.1–5.4 give quantitative participation thresholds, so the model can be used to test a given market's costs against the backfire region before imposing a target.

Reading between the lines

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

If the balance constraint applied to the aggregate dataset instead of to each seller, sellers could specialize in their cost-advantaged groups; the paper's Section 7 flag suggests the backfire region would shrink, and re-deriving Lemma 4.2 under aggregate-only balance would test this.
With partial transfer learning between groups, the per-group 'learning ante' effectively shrinks; the large-market amortization limit should persist, but the small-market backfire threshold should move, a prediction one could test by introducing a transfer parameter into the learning curve.
The ratio metric suggests a practical regulatory test: measure intervention-to-baseline revenue for a data marketplace as it grows; if the ratio tracks toward 1, the fairness rule is nearly free, and the crossover point $n_0$ from Claim 6.1 could be estimated empirically.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

Hi — quick read of arXiv:2501.19294. The punchline: the algebra error in Claim 4.1 is real and it is load-bearing. The threshold constant k_G has the α exponent and the bracket exponent flipped, so the "if and only if" participation condition is false. For β=1, D=2, α=1, n=1, C_a=10, the true zero-profit cost is c_a ≤ 10, and the paper's formula says c_a ≤ 160. That is not a minor typo: Theorems 4.1, 5.2–5.4 all compare costs to this threshold, so the equilibrium characterizations and backfire results are unsupported as written.

That said, the paper is not empty. The modeling move — endogenizing data production and adding a demographic-balance constraint to the Agarwal et al. marketplace — is a legitimate new contribution. The qualitative claim that fairness costs can backfire in small markets but amortize in large ones is interesting and policy-relevant. The limitations section is unusually honest about the naive intervention and zero-transfer assumptions. Those assumptions are strong, but the authors flag them. The asymptotic amortization theorem (6.1) likely survives a corrected threshold, since the leading positive term is linear in C_a while the negative terms are sublinear; the proof's participation guarantee needs to be redone, but the structure is plausible.

The soft spot is the proof of Claim 4.1, specifically the algebra around inequality (68) in Appendix A.5. The exponent inversion appears to be a simple sign error, but it propagates. I would not desk-reject; I would send this to a referee who can verify the corrected constant, and ask the authors to rework the threshold and rerun the statements that depend on it. If the fix is clean, the paper becomes a decent contribution. If the amortization proof cannot be repaired, the paper still has the small-market backfire result, but that too depends on the corrected threshold.

Verdict: worth a serious referee, with the expectation of major revision. I would not cite it until the constant is fixed.

Referee Report

3 major / 4 minor

Summary. The paper studies a stylized online data market, built on the Agarwal-Dahleh-Sarkar model, in which sellers produce training data endogenously and a marketplace can impose a demographic-balance constraint on each seller's dataset. In a quasi-symmetric setting the authors characterize baseline and intervention equilibria (Section 4), prove that a fairness intervention can backfire and prevent market formation in small/emerging markets (Section 5), and argue that in large markets the cost of fairness, measured as a ratio of intervention to baseline utility, vanishes for the marketplace and sellers and is at least 1 in the limit for buyers (Section 6, Theorem 6.1). The qualitative message is that the economic cost of a balanced-production rule can be maximal in small markets but amortizes with market growth.

Significance. If the results were correct, the paper would make a useful contribution to the economics of fair machine learning and online data markets, with a falsifiable comparative-static prediction: balanced-production interventions are risky in small markets but become asymptotically costless as demand grows. The model is self-contained and does not fit parameters to its conclusions; the proofs are traditional but the derivations are explicit. The authors also candidly state the main modeling limitations (per-seller balance, zero inter-group transfer, single fairness criterion) in Section 7. However, the current manuscript contains a load-bearing algebraic error in the central participation-threshold characterization, and a gap in the asymptotic theorem for zero-weight target groups, so the paper cannot be accepted in its present form.

major comments (3)

[Section 4.1, Eq. (19) and Appendix A.5] The participation constant k_G in Claim 4.1 has inverted exponents and is numerically wrong. Substituting Lemma 4.1's production quantity X_a = (C_a α β / c_a)^{1/(β+1)} into inequality (68), the zero-profit condition C_a G(X_a) - c_a X_a ≥ 0 solves to c_a ≤ C_a D^{(β+1)/β} α^{-1/β} (β^{-β/(β+1)} + β^{1/(β+1)})^{-(β+1)/β}. The paper instead states the equivalent of k_G = D^{(β+1)/β} α^{1/β} (β^{-β/(β+1)} + β^{1/(β+1)})^{(β+1)/β}, i.e., both the α power and the bracket power have the wrong sign. A concrete check: for β=1, α=1, D=2, C_a=10 and m=1, the true threshold is c_a ≤ 10, whereas Eq. (19) gives F_a=160. At c_a=50 the seller's utility is 20 - 2√(10·50) ≈ -24.7, so the seller does not produce, contradicting Claim 4.1. Because the same constant appears in Claim 4.2 and Theorem 4.2, and because the backfire theorems compare costs with F_a, this error propagates into Theorem 4.1, Corollary 4.1, and Theorems 5.2-5.4. The qualitative backfire phenomenon may survive a corrected constant, but the equilibrium characterization as stated is unsound.
[Section 4.2, Theorem 4.2] The 'if and only if' in Theorem 4.2 is not proven. Claim 4.2 establishes only the forward direction: if the sellers produce a positive number of samples at a Nash equilibrium, then the marginal production cost c·γ is at most the threshold F_B(C,γ). No argument is given for the converse, namely that the threshold inequality implies the candidate production level from Lemma 4.2 yields nonnegative seller utility and satisfies the learning-ante condition (25). For the baseline scenario Claim 4.1 supplies both directions with an explicit sufficiency proof; the intervention scenario lacks the corresponding sufficiency step. This leaves the equilibrium characterization in Theorem 4.2 incomplete.
[Section 6 and Appendix A.17, Theorem 6.1] The proof of Theorem 6.1 assumes that every group receives an unbounded amount of data in the intervention scenario as n grows. The argument uses Claim A.3 to conclude that γ_g B([m]) → ∞ for every group g, but this conclusion requires γ_g > 0 for all g. The model allows γ_g = 0 in Definition 3.6. If the target vector has zero weight for some group, the intervention mandates zero samples for that group, so the buyer surplus from that group is zero in the intervention scenario, and the ratio in Eq. (41) cannot be asserted without additional assumptions. The theorem needs either an explicit positivity assumption on the target vector or a separate treatment of zero-weight groups.

minor comments (4)

[Section 3.5] After Definition 3.6, 'demographicaly' should be 'demographically'.
[Section 1] In the contributions bullet list, 'the the cost of fairness can be completely oﬀset' contains a duplicated article; also the same bullet says 'the cost of fairness amortizes' without prior definition of the ratio used later.
[Theorem 5.4] The constant q in inequality (35) is not defined in the theorem statement; it is only introduced inside the proof of Appendix A.12. The statement should define q (the maximum ratio of potential economic values) before using it.
[Section 6] The notation in equations (39)-(41) is slightly inconsistent: U_M^{alt}(p) in Eq. (39) denotes the intervention-scenario revenue, while the same symbol is used with different arguments elsewhere; clarifying the notation would help.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the equilibrium and asymptotic results are derived from stated primitives and do not reduce to their inputs.

full rationale

The paper's derivation chain is self-contained. Equilibrium production quantities follow from first-order conditions on seller utility under Shapley-value revenue division (Proof of Lemma 4.1, Appendix A.3; Proof of Lemma 4.2, Appendix A.7). Participation thresholds are obtained by substituting these production quantities into the nonnegative-utility condition (Proof of Claim 4.1, Appendix A.5; Proof of Claim 4.2, Appendix A.8). The backfire results are established by explicit constructions and threshold comparisons, and the amortization results in Theorem 6.1 follow by comparing the leading linear terms in the potential economic values C_a against sublinear terms in baseline and intervention utilities. No parameter is fitted to the predicted ratio, no target quantity is used to define a premise, and no claimed prediction is equivalent by construction to an input. The only self-authored citations ([33] and [34]) appear in related-work and limitations contexts and are not load-bearing for the central theorems. The skeptical note about inverted exponents in the Claim 4.1 threshold, if correct, would be a mathematical correctness defect in the stated constant, not circularity; it does not show that any result is assumed by definition or derived from itself.

Assumptions & free parameters 0 free parameters · 6 assumptions · 0 invented entities

No fitted constants are introduced; D, alpha, beta, c, and mu are primitives quantified over in theorem statements. The six axioms above are the premises the equilibrium derivations actually invoke.

assumptions (6)

domain assumption Learning curve shape G(x) = (D - alpha x^{-beta})_+ (Definition 3.3)
Adopted from the scaling-law literature; all equilibrium and threshold formulas are derived from this functional form, so it is load-bearing.
domain assumption Zero inter-group transfer (Assumption 3.1)
Only samples of a group's own class contribute to that group's learning curve; the paper calls this an extreme simplification and says it overestimates the cost of fairness.
ad hoc to paper Quasi-symmetric setting (Definition 4.1)
All buyers share one learning curve and all sellers share one cost vector; closed-form equilibria exist only here, and Proposition 4.1 argues the general case has no closed-form solution.
domain assumption Constant marginal production cost per group (Definition 3.2)
Sellers face linear costs c_a per sample; this sample-pricing simplification enters every first-order condition.
standard math Myerson reserve-price mechanism with Shapley revenue division (Section 3.3)
Inherited unchanged from Agarwal et al. [2]; buyers truthfully bid and sellers are paid their Shapley marginal contribution.
ad hoc to paper Each seller's own dataset must be balanced (Section 3.5)
The intervention accepts a seller's data only if that seller's dataset matches the target gamma; this prevents specialization and is the main structural assumption behind the model's coupling. The authors flag it as naive in Section 7.

how reviews work

0 comments

Cite this review

Pith. "Pith review of The Cost of Balanced Training-Data Production in an Online Data Market." pith.science (2026). https://pith.science/paper/CMEJMKNV

@misc{pith2026250119294,
  author       = {Pith},
  title        = {Pith review of: The Cost of Balanced Training-Data Production in an Online Data Market},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CMEJMKNV}},
  note         = {Machine review of arXiv:2501.19294}
}

read the original abstract

Many ethical issues in machine learning are connected to the training data. Online data markets are an important source of training data, facilitating both production and distribution. Recently, a trend has emerged of for-profit "ethical" participants in online data markets. This trend raises a fascinating question: Can online data markets sustainably and efficiently address ethical issues in the broader machine-learning economy? In this work, we study this question in a stylized model of an online data market. We investigate the effects of intervening in the data market to achieve balanced training-data production. The model reveals the crucial role of market conditions. In small and emerging markets, an intervention can drive the data producers out of the market, so that the cost of fairness is maximal. Yet, in large and established markets, the cost of fairness can vanish (as a fraction of overall welfare) as the market grows. Our results suggest that "ethical" online data markets can be economically feasible under favorable market conditions, and motivate more models to consider the role of data production and distribution in mediating the impacts of ethical interventions.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

62 extracted references · 60 canonical work pages

[1]

Niels Henrik Abel. 1826. Démonstration de l’impossibil ité de la résolution al- gébrique des équations générales qui passent le quatrieme d egré. Journal für die reine und angewandte Mathematik 1 (1826), 65–96

work page
[2]

Anish Agarwal, Munther Dahleh, and Tuhin Sarkar. 2019. A marketplace for data: An algorithmic solution. In Proceedings of the 2019 ACM Conference on Economics and Computation . 701–726

work page 2019
[3]

Vítor Albiero, Kai Zhang, and Kevin W Bowyer. 2020. How do es gender balance in training data aﬀect face recognition accuracy?. In 2020 ieee international joint conference on biometrics (ijcb) . IEEE, 1–10

work page 2020
[4]

Dataset Providers Alliance. 2024. Dataset Providers Alliance . https://www.thedpa.ai/

work page 2024
[5]

Julia Angwin, Jeﬀ Larson, Surya Mattu, and Lauren Kirchn er. 2016. Machine Bias: There’s software used across the country to pr edict future criminals. And it’s biased against blacks. ProPublica (2016). https://www.propublica.org/article/machine-bias-ris k-assessments-in-criminal-sentencing

work page 2016
[6]

Santiago Andrés Azcoitia and Nikolaos Laoutaris. 2022. A survey of data mar- ketplaces and their business models. ACM SIGMOD Record 51, 3 (2022), 18–29

work page 2022
[7]

Solon Barocas, Moritz Hardt, and Arvind Narayanan. 2023 . Fairness and machine learning: Limitations and opportunities . MIT press

work page 2023
[8]

Joy Buolamwini and Timnit Gebru. 2018. Gender shades: In tersectional accu- racy disparities in commercial gender classiﬁcation. In Conference on fairness, accountability and transparency. PMLR, 77–91

work page 2018

Show all 62 references

[9]

William Cai, Ro Encarnacion, Bobbie Chern, Sam Corbett- Davies, Miranda Bo- gen, Stevie Bergman, and Sharad Goel. 2022. Adaptive sampli ng strategies to construct equitable training datasets. In Proceedings of the 2022 ACM Conference on Fairness, Accountability, and Transparen...

2022
[10]

Aylin Caliskan, Joanna J Bryson, and Arvind Narayanan. 2017. Semantics de- rived automatically from language corpora contain human-l ike biases. Science 356, 6334 (2017), 183–186

2017
[11]

Nicolas Carlini, Jamie Hayes, Milad Nasr, Matthew Jagi elski, Vikash Sehwag, Florian Tramer, Borja Balle, Daphne Ippolito, and Eric Wall ace. 2023. Extract- ing training data from diﬀusion models. In 32nd USENIX Security Symposium (USENIX Security 23) . 5253–5270

2023
[12]

Raul Castro Fernandez. 2022. Protecting data markets f rom strategic buyers. In Proceedings of the 2022 International Conference on Managem ent of Data . 1755– 1769

2022
[13]

Shuchi Chawla and Meena Jagadeesan. 2022. Individual f airness in advertis- ing auctions through inverse proportionality. In 13th Innovations in Theoretical Computer Science Conference (ITCS 2022). Schloss Dagstuhl-Leibniz-Zentrum für Informatik

2022
[14]

Irene Chen, Fredrik D Johansson, and David Sontag. 2018 . Why is my classiﬁer discriminatory? Advances in neural information processing systems 31 (2018)

2018
[15]

Manu Chopra, Indrani Medhi Thies, Joyojeet Pal, Colin S cott, William Thies, and Vivek Seshadri. 2019. Exploring crowdsourced work in low-resource settings. In Proceedings of the 2019 CHI Conference on Human Factors in Com puting Systems. 1–13

2019
[16]

Sam Corbett-Davies, Johann D Gaebler, Hamed Nilforosh an, Ravi Shroﬀ, and Sharad Goel. 2023. The measure and mismeasure of fairness. The Journal of Machine Learning Research 24, 1 (2023), 14730–14846

2023
[17]

Tore Dalenius. 1977. Towards a methodology for statist ical disclosure control. (1977)

1977
[18]

Deﬁned ˙AI. 2024. Deﬁned ˙AI. https://deﬁned.ai/

2024
[19]

Cynthia Dwork, Moritz Hardt, Toniann Pitassi, Omer Rei ngold, and Richard Zemel. 2012. Fairness through awareness. In Proceedings of the 3rd innovations in theoretical computer science conference . 214–226

2012
[20]

Cynthia Dwork, Nicole Immorlica, Adam Tauman Kalai, an d Max Leiserson
[21]

Hadi Elzayn and Benjamin Fish. 2020. The eﬀects of compe tition and regulation on error inequality in data-driven markets. In Proceedings of the 2020 conference on fairness, accountability, and transparency . 669–679

2020
[22]

Michael Feldman, Sorelle A Friedler, John Moeller, Car los Scheidegger, and Suresh Venkatasubramanian. 2015. Certifying and removing disparate impact. In proceedings of the 21th ACM SIGKDD international conferenc e on knowledge discovery and data mining . 259–268

2015
[23]

Lacuna Fund. 2024. Lacuna Fund. https://lacunafund.org/about/

2024
[24]

Arpita Ghosh and Aaron Roth. 2011. Selling privacy at au ction. In Proceedings of the 12th ACM conference on Electronic commerce . 199–208

2011
[25]

Matthew Gwilliam, Srinidhi Hegde, Lade Tinubu, and Ale x Hanson. 2021. Rethinking common assumptions to mitigate racial bias in fa ce recognition datasets. In Proceedings of the IEEE/CVF International Conference on Com puter Vision. 4123–4132

2021
[26]

Moritz Hardt, Eric Price, and Nati Srebro. 2016. Equali ty of opportunity in su- pervised learning. Advances in neural information processing systems 29 (2016)

2016
[27]

Jordan Hoﬀmann, Sebastian Borgeaud, Arthur Mensch, El ena Buchatskaya, Trevor Cai, Eliza Rutherford, Diego de Las Casas, Lisa Anne H endricks, Jo- hannes Welbl, Aidan Clark, et al. 2022. Training compute-op timal large lan- guage models. In Proceedings of the 36th Internationa...

2022
[28]

Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B Brown , Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeﬀrey Wu, and Dario A modei. 2020. Scaling laws for neural language models. arXiv preprint arXiv:2001.08361 (2020)

2020 arXiv
[29]

Karya. 2024. Karya. https://www.karya.in/

2024
[30]

Javen Kennedy, Pranav Subramaniam, Sainyam Galhotra, and Raul Castro Fer- nandez. 2022. Revisiting online data markets in 2022: A sell er and buyer per- spective. ACM SIGMOD Record 51, 3 (2022), 30–37

2022
[31]

Jon Kleinberg, Sendhil Mullainathan, and Manish Ragha van. 2017. Inherent Trade-Oﬀs in the Fair Determination of Risk Scores. In 8th Innovations in Theoretical Computer Science Conference (ITCS 2017) . Schloss Dagstuhl-Leibniz- Zentrum fuer Informatik

2017
[32]

Vlasis Koutsos, Dimitrios Papadopoulos, Dimitris Cha tzopoulos, Sasu Tarkoma, and Pan Hui. 2021. Agora: A privacy-aware data marketplace. IEEE Transactions on Dependable and Secure Computing 19, 6 (2021), 3728–3740

2021
[33]

Guocheng Liao, Yu Su, Juba Ziani, Adam Wierman, and Jian wei Huang. 2022. The privacy paradox and optimal bias-variance trade-oﬀs in data acquisition. ACM SIGMETRICS Performance Evaluation Review 49, 2 (2022), 6–8

2022
[34]

Roland Maio and Augustin Chaintreau. 2020. Incentives needed for low-cost fair lateral data reuse. In Proceedings of the 2020 ACM-IMS on Foundations of Data Science Conference. 71–82

2020
[35]

Aditya Krishna Menon and Robert C Williamson. 2018. The cost of fairness in binary classiﬁcation. In Conference on Fairness, accountability and transparency . PMLR, 107–118

2018
[36]

Mozilla. 2024. Common Voice. https://commonvoice.mozilla.org/en

2024
[37]

Felix B Mueller, Rebekka Görge, Anna K Bernzen, Janna C P irk, and Maximil- ian Poretschkin. 2024. LLMs and Memorization: On Quality an d Speciﬁcity of Copyright Compliance. arXiv preprint arXiv:2405.18492 (2024)

2024 arXiv
[38]

Roger B Myerson. 1981. Optimal auction design. Mathematics of operations research 6, 1 (1981), 58–73

1981
[39]

Aakash Varma Nadimpalli and Ajita Rattani. 2022. GBDF: gender balanced deep- fake dataset towards fair deepfake detection. In International Conference on Pat- tern Recognition. Springer, 320–337

2022
[40]

Ziad Obermeyer, Brian Powers, Christine Vogeli, and Se ndhil Mullainathan
[42]

OSTP. 2023. Equity and Law Enforcement Data Collection, Use, and Transparency. https://web.archive.org/web/20250118013103/https://www.whitehouse.gov/wp-content/uploads/2023/05/NSTC-

2023
[43]

Katie Paul and Anna Tong. 2024. Inside Big Tech’s un- derground race to buy AI training data. Reuters (2024). https://www.reuters.com/technology/inside-big-techs -underground-race-buy-ai-training-data-2024-

2024
[44]

Billy Perrigo. 2023. Exclusive OpenAI used Kenyan work ers on less than $2 per hour to make ChatGPT less toxic. Time Magazine 18 (2023), 2023

2023
[45]

HLR (Harvard Law Review). 2017. State v. Loomis: Wiscon sin Supreme Court Requires Warning Before Use of Algorithmic Risk Assessment s in Sentencing. Harvard Law Review 130 (2017), 1530–1537

2017
[46]

Fabian Schomm, Florian Stahl, and Gottfried Vossen. 20 13. Marketplaces for data: an initial survey. ACM SIGMOD Record 42, 1 (2013), 15–26

2013
[47]

Lloyd S Shapley. 1953. A value for n-person games. Contribution to the Theory of Games 2 (1953)

1953
[48]

Markus Spiekermann. 2019. Data marketplaces: Trends a nd monetisation of data goods. Intereconomics 54, 4 (2019), 208–216

2019
[49]

Florian Stahl, Fabian Schomm, Lara Vomfell, and Gottfr ied Vossen. 2015. Mar- ketplaces for digital data: Quo vadis? Technical Report. ERCIS Working Paper

2015
[50]

Florian Stahl, Fabian Schomm, and Gottfried Vossen. 20 14. The data marketplace survey revisited. Technical Report. ERCIS Working Paper
[51]

Fairly Trained. 2024. Fairly Trained. https://www.fairlytrained.org/ Augustin Chaintreau, Roland Maio, and Juba Ziani

2024
[52]

BL Van Der Waerden and Fred Blum. 1949. Modern algebra. Volume I. Frederick Ungar

1949
[53]

Tom Viering and Marco Loog. 2022. The shape of learning c urves: a review. IEEE Transactions on Pattern Analysis and Machine Intelligence 45, 6 (2022), 7799– 7819

2022
[54]

Kaiyu Yang, Klint Qinami, Li Fei-Fei, Jia Deng, and Olga Russakovsky. 2020. To- wards fairer datasets: Filtering and balancing the distrib ution of the people sub- tree in the imagenet hierarchy. In Proceedings of the 2020 conference on fairness, accountability, and transparen...

2020
[55]

Rich Zemel, Yu Wu, Kevin Swersky, Toni Pitassi, and Cynt hia Dwork. 2013. Learning fair representations. In International conference on machine learning . PMLR, 325–333. A PROOFS A.1 Proof of Proposition ( 4.1) Proposition 4.1. There does not exist a general closed-form soluti...

2013
[58]

To do so, we must evaluate G ((/u1D70C/u1D454 /u1D705/u1D454 /u1D6FC/u1D6FD ) 1 /u1D6FD+1 ) (72) which depends on whether Inequality ( 69) holds

We must show that the sellers will obtain non-negative utility by producing a positive number of samples, i.e., Inequality ( 66) holds. To do so, we must evaluate G ((/u1D70C/u1D454 /u1D705/u1D454 /u1D6FC/u1D6FD ) 1 /u1D6FD+1 ) (72) which depends on whether Inequality ( 69) ho...
[61]

We conclude that the data market does not form in t he intervention scenario

follows. We conclude that the data market does not form in t he intervention scenario. □ A.10 Proof of Theorem ( 5.2) Theorem 5.2. Let /u1D441buyers and /u1D440sellers be a fully-forming data market. If the marketplace chooses the uniform interventio n, i.e., /u1D6FE= /u1D462,...
[62]

if /u1D70C/u1D454= 0, then /u1D715 /u1D715/u1D70C/u1D454 /u1D453(/u1D70C) > 0; and 2) if /u1D70C/u1D454= /u1D450/u1D707, then /u1D715 /u1D715/u1D70C/u1D454 /u1D453(/u1D70C) < 0. And observe that in both cases, we have /u1D715 /u1D715/u1D70C/u1D454 /u1D453(/u1D70C) = − /u1D450 ...
[85]

is by deﬁnition of the seller’s utility; Equa- tion ( 86) is by deﬁnition of the payment division function; Equa- tion (87) is by Fact (A.3) since the sellers all play the same strategy; Equation (88) is by deﬁnition of the allocation function; Equation (89) is by quasi-symmet...
[99]

is by deﬁnition of the seller’s utility; Equa- tion (100) is by deﬁnition of the payment division function; Equa- tion (101) is by Fact (A.3) since the sellers all play the same strategy; Equation (102) is by deﬁnition of the allocation function; Equation (103) is by quasi-sym...
[205]

□ A.14 Proof of Claim ( A.1) Claim A.1

will be satisﬁed for all /u1D441> /u1D4410. □ A.14 Proof of Claim ( A.1) Claim A.1. If max/u1D45D/u1D454 /u1D70C/u1D454→ ∞ as /u1D441→ ∞ , then there exists an /u1D4410 such that /u1D441> /u1D4410 implies that for all /u1D457, /u1D465( /u1D457) /u1D454 > 0. P/r.sc/o.sc/o.sc/f....
[2018]

In Conference on fairness, accountability and transparency

Decoupled classiﬁers for group-fair and eﬃcient mach ine learning. In Conference on fairness, accountability and transparency . PMLR, 119–133
[2019]

Science 366, 6464 (2019), 447–453

Dissecting racial bias in an algorithm used to manage t he health of popu- lations. Science 366, 6464 (2019), 447–453

2019

Pith tools

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

[1] [1]

Niels Henrik Abel. 1826. Démonstration de l’impossibil ité de la résolution al- gébrique des équations générales qui passent le quatrieme d egré. Journal für die reine und angewandte Mathematik 1 (1826), 65–96

work page

[2] [2]

Anish Agarwal, Munther Dahleh, and Tuhin Sarkar. 2019. A marketplace for data: An algorithmic solution. In Proceedings of the 2019 ACM Conference on Economics and Computation . 701–726

work page 2019

[3] [3]

Vítor Albiero, Kai Zhang, and Kevin W Bowyer. 2020. How do es gender balance in training data aﬀect face recognition accuracy?. In 2020 ieee international joint conference on biometrics (ijcb) . IEEE, 1–10

work page 2020

[4] [4]

Dataset Providers Alliance. 2024. Dataset Providers Alliance . https://www.thedpa.ai/

work page 2024

[5] [5]

Julia Angwin, Jeﬀ Larson, Surya Mattu, and Lauren Kirchn er. 2016. Machine Bias: There’s software used across the country to pr edict future criminals. And it’s biased against blacks. ProPublica (2016). https://www.propublica.org/article/machine-bias-ris k-assessments-in-criminal-sentencing

work page 2016

[6] [6]

Santiago Andrés Azcoitia and Nikolaos Laoutaris. 2022. A survey of data mar- ketplaces and their business models. ACM SIGMOD Record 51, 3 (2022), 18–29

work page 2022

[7] [7]

Solon Barocas, Moritz Hardt, and Arvind Narayanan. 2023 . Fairness and machine learning: Limitations and opportunities . MIT press

work page 2023

[8] [8]

Joy Buolamwini and Timnit Gebru. 2018. Gender shades: In tersectional accu- racy disparities in commercial gender classiﬁcation. In Conference on fairness, accountability and transparency. PMLR, 77–91

work page 2018