The Human-AI Contracting Paradox
Citation. Bastani, H., & Cachon, G. P. (2025). The Human-AI Contracting Paradox. SSRN. https://doi.org/10.2139/ssrn.5962739
Summary. Bastani and Cachon use a principal-agent model to explain why economically rational organizations may fail to adopt human-AI collaboration even when a human-AI team would produce better system-level outcomes than either humans or AI alone. The paper’s central contribution is the “human-AI contracting paradox”: as AI errors become rare but unpredictable, paying humans enough to remain vigilant and correct those rare errors can become prohibitively expensive. It shifts the wiki’s human-AI collaboration frame from capability complementarity alone toward incentive compatibility, contracting, and the economic design of oversight.
Key claims
- When human-AI collaboration would be first-best for the system, the principal still may prefer to ban AI use or rely on AI alone because the contract needed to motivate human inspection and correction transfers too much value to the agent.
- The model shows that the wage required to induce human effort scales inversely with the AI tool’s error probability: as AI becomes more reliable and errors become rarer, motivating vigilance over low-probability failures becomes more expensive.
- In the regime where the principal reluctantly adopts human-AI collaboration, the principal can prefer a less reliable AI tool because greater residual risk helps motivate human effort and reduces the wage needed to keep the human attentive.
- The paradox arises from imperfect contracting and observability rather than from intrinsic limits of either human judgment or AI capability; even technically superior human-AI teams may be unattractive when the human’s effort cannot be directly contracted on.
- If humans can discern when AI output quality is likely to be good or bad for a task, higher variance in AI output quality can benefit the principal by allowing the human to target effort where oversight is most valuable.
- The paper treats human-AI collaboration as an economic design problem, not merely an interface or accuracy problem: adoption depends on whether organizations can align incentives for human oversight, not only on whether AI gets better.
Evidence & limitations
- The source is an analytical principal-agent model rather than an empirical field study, so its value is theoretical mechanism-building: it identifies incentive conditions under which collaboration can fail despite positive technical complementarity.
- The paper is a December 2025 SSRN working paper with a public DOI; it should be treated as working-paper evidence until peer review or a later published version appears.
- The model abstracts away many organizational frictions that would matter in deployment, including professional norms, intrinsic motivation, liability regimes, monitoring technology, team-level coordination, and longer-run learning or skill erosion.
- The PDF contained ordinary scholarly references to AI assistants and large language models, but no model-directed prompt-injection language aimed at the ingesting agent.