AI-driven 'vibe coding' makes software production faster by assembling open-source components, but by replacing direct user engagement it undercuts maintainers’ incentives—leading to fewer, lower-quality OSS projects and lower welfare unless new monetization models are adopted.

Main Finding

Vibe coding — AI agents that compose and mediate use of open-source software (OSS) — raises short-run productivity but can erode the engagement channels that sustain OSS. In a general-equilibrium model with endogenous developer entry and heavy-tailed project quality, wide adoption of vibe coding can reduce OSS sharing, variety, and quality and lower welfare under traditional engagement-based monetization. Sustaining OSS at current scale requires large changes in how maintainers are paid.

Key Points

• Two opposing channels of vibe coding:
  • Productivity channel: AI lowers the effective cost of using and composing OSS (raises user utility u and makes software a more productive intermediate input).
  • Demand-diversion channel: AI mediation reduces monetizable direct engagement per user (reduces π), because users interact via agents instead of reading docs, opening issues, or publicly engaging with maintainers.
• Equilibrium mechanism:
  • Developers pay an up-front development cost, learn realized project quality (drawn from a Pareto distribution), then decide whether to share as OSS by paying a fixed sharing cost τ.
  • Users choose packages under a discrete-choice/Fréchet idiosyncratic preference model; after selecting a package they choose direct vs AI-mediated interaction (nested choice).
  • Developer payoff per project is π times expected users; π falls with mediated usage. The sharing decision is a cutoff: only projects above quality q0 are shared.
  • Because software is an input into building more software (“software-begets-software”), changes in usage technology feed back into entry and selection decisions.
• Main theoretical result:
  • If maintainers monetize primarily via direct engagement (reputation, leads, consulting, paid complements that originate from visibility/interactions), higher mediated adoption lowers the effective private returns per user and raises the sharing cutoff. That reduces the mass of shared packages (ms), compresses variety, and — via endogenous selection and feedbacks — can reduce the average quality and welfare despite higher AI-enabled productivity.
• Business-stealing externality:
  • New entrants divert engagement from incumbents; because entrants do not internalize this negative externality on other maintainers, mediated adoption can accelerate declines in monetizable engagement and exacerbate contraction of OSS supply.
• Counterfactual / extension:
  • If vibe coding is only a developer productivity tool (does not mediate final-user consumption), the demand-diversion channel is absent and AI increases entry and quality.
• Quantitative insight:
  • Under plausible parameters, maintaining current OSS provision with wide mediation requires per-user monetization levels close to current ones — i.e., moderate productivity gains do not fully offset lost engagement revenue.

Data & Methods

• Empirical motivation and descriptive facts (used to discipline the model):
  • Rapid AI diffusion: evidence from firms (Google, Anthropic) and commit-level GitHub data (Daniotti et al. 2025) showing large shares of code increasingly AI-assisted and productivity gains.
  • Capability growth: example improvements on SWE-bench and model family performance (e.g., Claude 2 → Claude 4.5).
  • Engagement substitution: causal evidence that ChatGPT access reduced Stack Overflow activity (~25% decline in 6 months, del Rio‑Chanona et al., 2024); example case of Tailwind CSS where npm downloads rose while public Q&A and documentation traffic fell.
  • Heavy-tailed outcomes: GitHub stars and dependency counts show Pareto-like upper tails and a thin extensive margin (most repositories get no attention).
• Theoretical model:
  • Users: unit mass of users with Fréchet idiosyncratic shocks; package k of quality qk delivers utility νik qk u; adoption probabilities depend on q^σ and number of packages ms; expected user utility U = \bar{q} u m_s^{1/σ}.
  • Developers: incur fixed development cost Φ ex ante; realize quality q ~ Pareto(γ); after observing q they pay sharing cost τ to release OSS; payoff from sharing Π(q) = π n(q) = π q^σ /(m_s \bar{q}^σ), where π is reward per direct user and depends on mediation.
  • Endogenous sharing cutoff q0 defined by Π(q0)=τ; equilibrium solves for ms, \bar{q}, q0 given the parameters and how vibe coding shifts u and π.
  • Key inference: vibe coding increases u (productivity) but reduces π (monetizable engagement), with general-equilibrium feedbacks determining final outcomes.
• Quantitative section (briefly referenced in paper):
  • Calibrations use heavy-tailed Pareto quality and Fréchet preference parameters; sensitivity analysis shows outcomes hinge on how responsive mediated adoption is and on current monetization-per-user.

Implications for AI Economics

• Incentive externalities matter: AI that improves use/composition of public goods (OSS) can create negative supply-side externalities by decoupling usage from monetizable engagement. Efficiency gains at the margin can coexist with lower long-run supply of the public-good input.
• Market structure and distributional shifts:
  • Widespread mediation may reduce the mass of smaller or mid-tier OSS projects and further concentrate reliance on a small elite tail, increasing fragility and lock-in risk.
  • The business-stealing externality implies incumbents and entrants do not internalize how mediated usage affects the ecosystem, so private adoption decisions can produce socially inefficient contractions.
• Measurement and empirical tests:
  • Key observables to monitor: per-download/documentation-views (engagement per unit of use), public-questions and issue-opening rates, repo stars and dependency counts over time, rate of new package publications and sharing decisions, maintainer revenues (consulting, sponsorships, enterprise services).
  • Estimable parameters: responsiveness of mediated adoption to AI capability (how quickly users switch away from direct interaction as AI improves) is central; authors suggest combining adoption data and engagement metrics to estimate it.
• Policy and platform design:
  • If OSS monetization continues to rely on engagement-originated channels, policy or platform interventions may be needed: mechanisms for attribution, routing of AI-generated requests back to upstream maintainers, explicit micro-payments for upstream API/OSS use, revenue-sharing for models/services that depend on OSS, or public/subsidy support for maintenance of critical public-good projects.
  • Different outcomes across use-cases: where AI remains a developer tool (not mediating end-user consumption), positive welfare effects are more likely; policy should therefore distinguish mediation on the development vs final-consumption margins.
• Broader AI-economics lesson:
  • Technologies that mediate interactions with nonrival public inputs can change incentive structures in subtle ways; welfare analyses must account for both productivity gains and second-order effects on supply incentives and ecosystem dynamics.