Digital-labor platforms can complete large batches of tasks while paying a vanishing fraction of total labor cost by exploiting workers' uncertainty; however, a small, targeted coalition of low-cost workers enforcing a price floor can force the platform to pay proportional (linear) total wages, while random organizing is largely ineffective.

Main Finding

A platform that sequentially posts take-it-or-leave-it prices can exploit workers’ uncertainty (left tails of reservation-price distributions) to substantially suppress total wages while keeping task completion wait time linear in the number of tasks M. Under natural distributional assumptions a simple quantile-based pricing rule lets the platform pay sublinear total wages (e.g., O(log M) or even constant) instead of Θ(M). Random (horizontal) collective action by a fixed fraction of workers is ineffective at undoing this suppression, but a small, targeted (vertical) coalition of low-cost workers that commits to a price floor can force total platform spending back to Θ(M).

Key Points

• Model setup

  • Principal wants to procure M tasks (categories) by repeatedly posting a single price; workers arrive sequentially.
  • Each worker has a valuation (reservation price) for each category drawn independently from category distributions Di on [0,1]. A worker accepts if some category’s valuation ≤ posted price.
  • Price stays posted until a worker accepts; on acceptance that category is removed. Wait time is the total number of rejections (deferred events) across the process.
  • Principal targets linear expected wait time: E[WAIT(p)] = O(M).

• Principal’s objective and wage suppression

  • Principal minimizes total payments COST(p) = sum of posted prices at success events (equivalently total wages paid).
  • If each category has a non-negligible price floor p, then to guarantee coverage in finite time the principal must spend Θ(p M) = Θ(M).
  • If no floor exists (workers accept arbitrarily low prices and distributions have mass near zero), a principal with access to quantile information can adopt a quantile-based pricing strategy that achieves linear wait time but pays only sublinear total wages.
  • Specific regimes:
    • Moderately dispersed distributions (e.g., uniform): platform can drive total wages down to sublinear regimes (e.g., O(√M) or O(log M) depending on tail).
    • Heavy left-skew (mass concentrated near zero, e.g., Beta(a<1,1) tails): platform can sometimes obtain constant total payment in M.
  • Intuition: by “waiting out” higher-cost workers and posting low quantile prices, the principal exploits the positive probability of encountering low-reservation workers; as M grows, this can dramatically reduce aggregate payments relative to Θ(M).

• Collective action

  • Horizontal (random) coordination: recruiting a random α-fraction of workers to impose a price floor does not change the suppression bounds unless α→1; random partial participation is largely ineffective.
  • Vertical (targeted) coordination: recruiting a specific segment (e.g., the low-cost tail within some categories) and committing them to a price floor is effective even with a small coalition. The paper gives a closed-form condition (in terms of α and budget) for when a targeted intervention restores total payments to Θ(M) while preserving O(M) wait time.

• Complementary experiments

  • Synthetic Monte Carlo simulations across distributions and regimes confirm theoretical insights and showcase when collective action helps.

Data & Methods

• Theoretical model: a posted-price procurement model with M categories, sequential arrivals, i.i.d. valuation vectors drawn from product distribution D = D1×...×DM, valuations ∈[0,1].
• Pricing strategy analyzed: a simple quantile-based rule (principal is assumed to have oracle access to quantiles of Di) that guarantees linear expected wait time while minimizing cost.
• Analytical characterization: proofs link the left-tail growth (behavior of Di near 0) to asymptotic scaling of total cost with M (Θ(M) if a positive floor exists; sublinear e.g., O(log M), O(1) in vulnerable tails).
• Collective-action analysis: formal comparison between random (horizontal) and targeted (vertical) coalitions; derivation of budget/size conditions for vertical coalitions to be effective.
• Simulations: Monte Carlo experiments (60 runs), code available at the authors’ GitHub; experiments run in Python 3.9 to validate and illustrate regimes (uniform, Beta, Dirac/floored distributions, etc.).

Implications for AI Economics

• For AI data procurement (annotation, red-teaming, dataset labeling), platform pricing can systematically depress aggregate wages without slowing throughput, if workers’ reservation-price distributions have nontrivial left tails. Observed low pay on crowdsourcing platforms can be a consequence of strategic platform pricing exploiting worker uncertainty.
• Policy levers:
  • Minimum wage or price-floor regulation on platforms is effective: any credible positive floor prevents the sublinear-pay regimes and forces spending linear in M.
  • Transparency and information sharing about reservation prices or past payments can reduce exploitable uncertainty and make wage suppression harder.
  • Support for organizing targeted collectives (vertical strategies) is more effective than scattershot/random recruitment: organizing specific low-cost segments to commit to price floors yields much larger wage-restoring effects per participant.
• Platform design considerations:
  • Platforms or buyers with the ability to “wait out” high-priced workers (no time-pressure penalties) can extract large monopsony rents. Time constraints, batching rules, or matching mechanisms that prevent indefinite waiting can limit this exploitation.
  • Competition among principals (multiple buyers) or platform fees/commissions (excluded in the model) could materially change incentives; future mechanism design should account for multi-principal dynamics.
• Empirical suggestions:
  • Estimating reservation-price distributions—especially left-tail behavior—is critical for assessing monopsony power and designing interventions.
  • Experiments that simulate quantile-informed pricing (or measure its feasibility) can help understand how much platform knowledge is needed to implement suppression in practice.

Limitations / open questions - The model assumes oracle/quanti le access to distributional information and i.i.d. draws; learning dynamics and information acquisition costs are not modeled. - Single-principal setting; competition among multiple firms/platforms, platform take rates, multi-round worker learning and reputation effects are left for future work. - Practical costs and enforceability of collective commitments, and dynamic worker responses, deserve empirical and theoretical follow-up.