Generative AI can quietly hollow out public knowledge archives by resolving problems privately and discouraging contributors, creating self-reinforcing low-archive traps; nudging users to share AI-assisted solutions can offset private diversion but cannot restore participation-driven declines without policies to sustain contributors.

Main Finding

Generative AI can shrink the stock of public, reusable knowledge on digital platforms through two distinct margins: (1) a flow margin — AI privately resolves questions that would otherwise be posted, reducing posted volume; and (2) a resolution margin — AI raises contributors’ outside options, leading contributors to exit (pool thinning) and causing congestion that lowers the probability that posted questions get answered. These margins interact via a self‑undermining feedback (richer archives improve AI, which diverts more postings and raises outside options), producing possible multiple steady states and a structural minimum-viable archive below which the platform can collapse. Policy tools differ by margin: encouraging public sharing of AI-assisted solutions mitigates flow losses but cannot repair participation-driven degradation of resolution; sustaining contributor engagement (incentives, retention) is required to address the resolution margin. A simple empirical diagnostic follows: in a congested regime, a joint decline in posted volume and conditional resolution signals supply-side pool thinning, whereas volume decline with stable/increasing resolution implies private diversion (flow) dominates.

Key Points

• Two margins of archive decline:
  • Flow margin (private diversion): AI resolves more problems privately before they are posted, reducing the flow of knowledge-enhancing queries that reach the platform.
  • Resolution margin (pool thinning / congestion): AI improves outside options, causing contributors to exit; fewer contributors → lower matching probability → smaller chance a posted query is resolved.
• Composition/selection effect: When AI disproportionately diverts routine (easy) questions, the remaining posted pool becomes richer in knowledge-enhancing questions, which can raise the willingness to answer (may partly offset other effects).
• Race of forces determining resolution (Proposition 1): the net change in resolution probability is the outcome of three forces — composition enrichment (↑ resolution), posted-flow reduction (↓ congestion → ↑ resolution), and pool thinning (↓ resolution). Pool thinning can dominate and reduce resolution even when selection suggests the opposite.
• Dynamic crowd-out and decomposition (Proposition 2, Corollary 1): the decline in archive growth decomposes into the flow margin (fewer H-type queries posted) and the resolution margin (lower σ, the lifetime resolution probability of posted queries).
• Self‑undermining feedback and tipping (Proposition 3): as the archive changes it shifts outside options, which can create multiple steady states including a low-archive trap (a minimum viable archive). AI compresses the viable region, making collapse more likely.
• Policy instrument distinction (Proposition 4): encouraging public sharing of AI-assisted answers converts private resolutions into public artifacts and offsets the flow margin, but cannot restore resolution lost from contributor exit. There exists a threshold of sharing uptake above which the minimum-viable archive disappears.
• Domain heterogeneity: the routine-task share π (fraction of routine/easy tasks) scales both margins — high-π domains are doubly vulnerable to AI-induced archive decline.

Data & Methods

• Methodology: formal dynamic partial-equilibrium model of public-good provision on a digital Q&A platform. Analytical derivation of equilibrium objects, comparative statics, and steady-state dynamics; proofs establish lemmas and propositions. Appendices extend robustness (e.g., T>1, ability-dependent private resolution, alternative matching functions).
• Core model ingredients and objects:
  • Time-discrete dynamics for public archive Kt with effective depreciation λ; steady states solve h(K) = λK, where h(K) is expected new public knowledge created at K.
  • Population: unit mass of agents with ability α ∼ Ψ; each period receives a query: type L (routine) with prob. π or type H (knowledge-enhancing) with prob. 1 − π. Residual difficulty r drawn from Gθ.
  • Private resolution: succeeds if r ≤ r̄eθ(K); probability aeθ(K) = Gθ(r̄eθ(K)); AI raises these probabilities.
  • Posting decision: if private resolution fails, agent posts if subjective benefit U eθ(K) = σe(K)·Vθ exceeds posting cost d ∼ Γθ; posted flow qeθ(K) = Γθ(U eθ(K)) · (1 − aeθ(K)).
  • Matching/congestion: with Ψ(α*) participating agents and posted flow Qe(K), match probability for a posted query is µe(K) = min{1, Ψ/Qe}. Each participant can handle ≤1 query per period (T=1 baseline).
  • Answering decision: matched agent answers iff idiosyncratic cost c ≤ c = max{0, β Δ̄ + u − C(K)}; lifetime resolution σe(K) = µe(K)·F(c).
  • Participation cutoff α solves Se(K) = we(α, K; π) where Se is expected per-period payoff from volunteering and we is outside-option payoff; AI shifts we upward and lowers α* (fewer participants).
• Empirical motivation and cited evidence:
  • Observed declines in public-platform activity after LLM rollouts: Del Rio‑Chanona et al. (2024) on Stack Overflow, Padilla et al. (2025) on search traffic reductions to knowledge sites, Lyu et al. (2025) on Wikipedia views/edits declines.
• Robustness and extensions: model examines ability-dependent private resolution, alternative matching functions (search frictions), multi-period query lifetimes T > 1; main comparative statics persist.

Implications for AI Economics

• Distinct mechanisms require distinct policy responses:
  • Flow-driven decline (private diversion): policies that encourage or nudge users to publish AI-assisted answers (e.g., “share to public knowledgebase” buttons, easy export of chat answers to forums, provenance/attribution tooling) can convert private resolutions into public artifacts without weakening AI capability.
  • Resolution-driven decline (pool thinning): must be addressed by supply-side interventions that maintain or subsidize contributor engagement (monetary rewards, stronger reputation/pay structures, lowered answering costs via tooling, retention programs). Public sharing alone cannot restore matching capacity.
• Diagnostic for empirical work and platforms:
  • Track both posted volume and conditional resolution (answer rate given a post). In congested platforms, if posted volume and conditional resolution fall together → pool thinning is present. If posted volume falls but conditional resolution is stable or rising → private diversion dominates.
  • Monitor contributor counts/activity (Ψ), response latencies, match probabilities, and archive quality K over time to detect approach to the minimum-viable archive/tipping region.
• Cross-domain vulnerability: platforms serving domains with higher routine-task shares (π) are more exposed; cross-platform or cross-topic heterogeneity in π can be used in difference‑in‑differences or event-study designs around AI rollouts to identify margins.
• Relation to other policy ideas: complements other proposals (e.g., Acemoglu, Kong, and Ozdaglar’s garbling idea) — public sharing preserves AI capability while addressing flow losses; garbling targets incentives for human effort. Both may be part of a policy toolkit depending on welfare objectives.
• Research and measurement agenda:
  • Quantify relative magnitudes of the two margins empirically (requires platform logs: posting rates, answer rates, active answerers, AI usage).
  • Structural estimation to measure π, matching curvature, and the size of the self‑undermining feedback to assess tipping risk.
  • Field experiments: nudges to share AI-assisted answers (treatment) vs. contributor incentives (treatment) to test differential impacts on posted flow vs. resolution.
• Limitations to bear in mind:
  • Partial-equilibrium framework: no social-welfare optimization or general-equilibrium feedbacks outside the platform.
  • Reduced-form outside-option modeling (we) abstracts from explicit time allocation; quantifying π’s effect requires empirical calibration or richer time-allocation extensions.
  • Baseline assumes T = 1 (one-period query lifetime) and risk-neutral agents; qualitative insights extend but quantitative implications depend on extensions.

Overall, the paper provides a clear theoretical decomposition of how generative AI can erode digital public goods via two separable margins, gives empirically testable diagnostics to identify which margin dominates on particular platforms, and shows that remedies must be targeted to the operative margin — encouraging public sharing helps recover flow losses but restoring contributor supply requires direct interventions.