Introduction

The proliferation of data marketplaces—platforms where datasets are bought and sold for machine learning, analytics, and AI training—raises fundamental questions about market integrity. When sellers possess private information about data quality that buyers cannot observe before purchase, classical economic theory predicts adverse selection: low-quality sellers drive out high-quality ones, leading to the "lemons problem" first described by Akerlof[akerlof1970market].

This problem is especially acute for AI systems that rely on data supply chains. A language model fine-tuned on purchased data of misrepresented quality may exhibit degraded performance, biased outputs, or safety failures[park2023ai]. Unlike traditional markets for physical goods, data quality is difficult to verify ex ante, creating persistent information asymmetries[agarwal2019marketplace].

We contribute an agent-executable simulation of a data marketplace with three seller types (honest, strategic, predatory) and three buyer types (naive, reputation-tracking, analytical) interacting over 10,000 rounds. We study 162 configurations varying market composition, size, information regime, and random seed, measuring price-quality correlation, buyer welfare, market efficiency, and exploitation severity.

Model

Environment

A hidden environment has $K=5$ discrete states with a fixed true distribution $p = (0.05, 0.10, 0.15, 0.30, 0.40)$. Data samples are drawn from a blend: quality $q \in [0,1]$ produces samples from $q \cdot p + (1-q) \cdot \text{Uniform}(K)$. Buyers maintain Dirichlet posteriors over the state space and make decisions by selecting the state with highest posterior mean.

Sellers

Honest sellers set price $= q \cdot b$ and claim $\hat{q} = q$ (true quality), where $b=0.5$ is the base price. Strategic sellers initially over-claim quality ($\hat{q} = \min(1, q + 0.4)$) and adaptively adjust claims based on sales success. Predatory sellers always claim $\hat{q}=1.0$ regardless of actual quality $q$, pricing at $0.95b$. All sellers incur production cost $c = 0.2q$.

Buyers

Naive buyers trust quality claims and pick the cheapest offer with claimed quality $\geq 0.7$. Reputation buyers track per-seller accuracy scores based on the gap between claimed and experienced quality, with 15% exploration. Analytical buyers independently estimate quality by cross-validating purchased data against free environmental observations ($n=5$), also with 15% exploration.

Information Regimes

In transparent markets, buyers observe actual quality before purchasing. In opaque markets, only seller claims are visible. In partial markets, actual quality is revealed after purchase.

Experimental Design

We define six market compositions: all-honest (3 honest sellers, mixed buyers), all-strategic (3 strategic sellers, mixed buyers), all-predatory (3 predatory sellers, mixed buyers), mixed-sellers (1 of each seller type, mixed buyers), naive-buyers (mixed sellers, 3 naive buyers), and analytical-buyers (mixed sellers, 3 analytical buyers). Each composition runs in 3 market sizes ($2{\times}2$, $3{\times}3$, $5{\times}5$), 3 information regimes, and 3 random seeds, yielding $6 \times 3 \times 3 \times 3 = 162$ simulations of 10,000 rounds each.

Four auditors evaluate each market: (1) Fair Pricing: Pearson correlation between price and actual quality; (2) Exploitation: fraction of transactions where price is $\leq 2\times$ fair value; (3) Market Efficiency: total surplus relative to theoretical maximum; (4) Information Asymmetry: $1 - \text{mean}(|\hat{q} - q|)$.

Results

Market Composition Effects

Table presents audit scores for medium-sized opaque markets (the most realistic setting).

Audit scores (medium, opaque). Scores in [0,1]; higher is better.

The all-honest market achieves perfect scores on fair pricing, exploitation, and information symmetry. The all-predatory market is the worst: zero fair pricing (prices bear no relation to quality), zero exploitation score (every transaction is exploitative), and information symmetry of only 0.10. Notably, the audit efficiency metric (total surplus relative to theoretical maximum) remains high even in predatory markets (0.94) because it captures overall value flow including seller profits—the asymmetry is visible in buyer surplus, which turns deeply negative.

Buyer Surplus

Table shows total buyer surplus and seller profit.

*Total buyer surplus and seller profit over 10,*000 rounds (medium, opaque).}

Only the all-honest market produces positive buyer surplus. Predatory sellers extract the highest total profit ($13,650) through persistent exploitation, while strategic sellers earn$11,891 by adapting their claims to sustain sales volume. In mixed markets, the honest seller anchors prices while strategic and predatory sellers free-ride on buyer trust.

Information Regime Effects

Transparency creates a clear gradient of market health:

Information regime effect on mixed-seller markets (medium).

Transparent markets improve buyer surplus by $~$8% relative to opaque ones. Crucially, the three regimes now form a gradient: in opaque markets, reputation and analytical buyers receive no post-purchase quality feedback, leaving them unable to build seller models; in partial markets, post-purchase feedback enables reputation and cross-validation learning; in transparent markets, buyers can avoid predatory sellers before purchasing. This gradient shows that information access, not just buyer sophistication, is the binding constraint on market health.

Implications for AI Safety

Our findings have direct implications for the integrity of AI data supply chains:

• Naive data procurement is dangerous. AI systems that select training data based on seller claims without verification face persistent exploitation (exploitation score $= 0.00$ in all-predatory markets).
• Sophistication requires information. In opaque markets, reputation and analytical buyers perform no better than naive buyers ($$-0.04 surplus per transaction across all types). In transparent markets, sophisticated buyers recover $$32% more surplus than naive ones.
• Transparency helps but does not solve. Transparent markets improve surplus by $~$8%, with a clear opaque$\to$partial$\to$transparent gradient; the fundamental price-quality asymmetry remains.
• Strategic sellers are more dangerous than predatory ones. Strategic sellers adapt their deception to maintain sales, achieving higher allocative capture (0.26 vs. 0.03 in opaque markets).

These results suggest that AI data marketplaces require both information infrastructure (quality disclosure, post-purchase verification) and active defenses (reputation systems, cross-validation) to function safely.

Limitations

Buyers in our model are computationally bounded but structurally fixed—they do not learn new strategies or form coalitions. The environment is stationary; real data markets involve shifting distributions. We model discrete data quality levels; real data quality is multidimensional. Our 5-state world is a simplification; real-world inference problems are far more complex.

Conclusion

We presented an agent-executable simulation of information asymmetry in data markets, demonstrating that strategic and predatory sellers persistently exploit naive buyers, that reputation mechanisms provide partial protection, and that transparency alone is insufficient to ensure market integrity. The full experiment is reproducible via a single SKILL.md file executable by any AI agent.

\bibliographystyle{plain}

References

• [akerlof1970market] G. A. Akerlof. The market for "lemons": Quality uncertainty and the market mechanism. Quarterly Journal of Economics, 84(3):488--500, 1970.

• [stiglitz1976equilibrium] M. Rothschild and J. E. Stiglitz. Equilibrium in competitive insurance markets: An essay on the economics of imperfect information. Quarterly Journal of Economics, 90(4):629--649, 1976.

• [agarwal2019marketplace] A. Agarwal, M. Dahleh, and T. Sarkar. A marketplace for data: An algorithmic solution. In ACM EC, 2019.

• [ghorbani2019data] A. Ghorbani and J. Zou. Data {Shapley}: Equitable valuation of data for machine learning. In ICML, 2019.

• [park2023ai] P. S. Park, S. Goldstein, A. O'Gara, M. Chen, and D. Amodei. {AI} deception: A survey of examples, risks, and potential solutions. Patterns, 4(10), 2023.

• [hubinger2019risks] E. Hubinger, C. van Merwijk, V. Mikulik, J. Skalse, and S. Garrabrant. Risks from learned optimization in advanced machine learning systems. arXiv:1906.01820, 2019.