AI makes polished papers cheaper but trustworthy certification scarcer: as AI lowers the cost of producing well-presented manuscripts faster than it lowers the cost of judging scientific merit, submissions swell and the willingness to pay for credible review rises—allowing dominant certifiers to extract rents while constrained review capacity dilutes overall certification quality.

Main Finding

Cheaper AI-assisted manuscript production erodes the informational value of surface polish, shifting scarcity from producing credible-looking papers to verifying which contributions are genuinely valuable. As AI capability rises, entry expands, the mean quality of the uncertified pool can fall (under positive selection), and the willingness to pay for credible certification increases. Depending on institutional details, this premium can be captured as fees or appear as scarcity rents (queues, delays). With fixed review capacity and weak commitment (certifying unread overflow), certification dilutes, creating a measurable false‑positive externality that justifies a partial Pigouvian toll on submissions and raises the shadow value of review capacity.

Key Points

• Mechanism

  • AI capability a lowers the cost of producing manuscript polish more than it lowers the cost of independent judgement.
  • Polish s is observable; underlying research quality is θ. Cost of polish: c(s, θ, a) = s / (θ + a) (single-crossing: higher θ can produce polish more cheaply).
  • There is a finite perception ceiling on polish (s̄): beyond a threshold a*, polish can no longer fully separate high- and low-quality authors (loss of full separation).

• Entry and pool composition

  • Entry threshold β(a) falls with a (β′(a) < 0), so the mass of papers rises with AI capability.
  • Under positive selection (marginal entrants have lower expected quality), the mean quality of the uncertified pool VU(a) falls as a rises (pool deterioration).

• Certification premium and monetization

  • For a fixed certifier standard θc, the gross willingness to pay for a credible certification is ∆(a) = VC − VU(a), and ∆(a) strictly increases with a.
  • A certifier with sufficient market power (and no binding constraints) can set a fee F*(a) = ∆(a), but competition, non-price rationing, or different disclosure rules mean observed fees need not equal the premium.

• Attention constraints and weak-commitment extension

  • Introducing fixed verification capacity A and assuming weak commitment (certifier stamps unread overflow) produces false positives.
  • The constrained planner maximizing net benefit (benefit of admitting n applications minus false-certification loss) sets applications n* with first-order condition equating marginal benefit to marginal false-positive loss.
  • The optimal correction can be implemented by a partial Pigouvian toll τ(a) equal to the marginal false-certification loss; τ(a) and n*(a) both rise with AI capability (under concavity and increasing-differences conditions).
  • The shadow value of review capacity λA = probability an extra audit prevents a false positive = (1 − b)/n (b = mass of genuinely high types); at the corrected optimum λA and its derivative wrt a are positive.

• Institutional distinctions matter

  • Different monetary objects: certification fee, submission toll, reviewer compensation, and scarcity rents (queues) are economically distinct and have different welfare implications.
  • Commitment to not certify unread work preserves stamp credibility. Weak commitment leads to stamp dilution and a measurable externality.

• Empirical predictions (from model)

• Reduced association between surface polish and later-costly-evaluation outcomes in fields with rapid generative-AI adoption.
• Entry expansion concentrated where codifiable execution costs were binding; uncertified-pool mean falls only if marginal entrants are lower quality.
• For certifiers maintaining credible standards, the gap between evaluated certified work and the uncertified pool should widen; submission burdens may rise even if cash fees do not.
• With fixed verification capacity and increased unread volume, triage value and the return to added editorial/review capacity should increase; submission caps, deposits or triage reduce false certification cheaply.

Data & Methods

• Approach: analytical theoretical model + illustrative uniform example and numerical illustration.
• Core model ingredients:
  • Agent heterogeneity: authors indexed by θ ∼ G on [0,1]; θ ranks expected quality and (in the benchmark) execution cost.
  • AI capability parameter a ≥ 0 that reduces signaling cost more than verification cost.
  • Observable polish s ∈ [0, s̄] with production cost c(s, θ, a) = s/(θ + a).
  • Endogenous entry threshold β(a) with β′(a) < 0, yielding mass of papers n0(a) = 1 − G(β(a)).
  • Certifier posts a verification standard θc and (in benchmark) verifies applicants and certifies only θ ≥ θc at verification cost k.
  • Weak-commitment extension fixes verification capacity A and assumes unread overflow is stamped (leading to false positives).
• Main analytical results: five formal propositions/lemmas deriving (i) loss of full separation, (ii) pool deterioration, (iii) credible-certification premium, (iv) expression for optimal partial Pigouvian toll, and (v) shadow value of capacity and its comparative statics.
• Illustrative special case: uniform θ and β(a) = θ0/(1 + a) used to plot comparative statics (figure in paper).
• Limitations and maintained assumptions:
  • Underlying distribution of discovery quality is held fixed (no endogenous effect of AI on idea-generation quality).
  • Positive selection assumption: marginal entrants have lower expected quality than incumbents—this is crucial for pool-deterioration result and must be tested empirically.
  • Weak-commitment extension is a polar case (useful to measure false-positive externality) and not a claim about typical reputable journal practice.
  • The welfare objective and loss normalization are partial (exclude some margins like false negatives, heterogeneous audit costs, and benefits of improved AI-generated discoveries).

Implications for AI Economics

• Conceptual shift: policy and research should pay attention to the market for evaluation (certification and peer review), not only production-side productivity. AI can make presentation cheap faster than it makes evaluation cheap, moving scarcity downstream.
• Monetization risk and forms:
  • The eroded upstream signal increases willingness to pay for credible certification; monetization may appear as explicit fees, submission charges, refundable deposits, paid reviewer time, or scarcity rents (longer delays, higher editorial bargaining).
  • Market structure matters: monopoly certifiers can capture the full premium; competitive or multi-margin rationing may produce different outcomes and welfare properties.
• Capacity and institution design:
  • Fixed verification capacity makes triage, caps on simultaneous submissions, submission deposits, reviewer compensation, and verification-first AI tools economically valuable.
  • Commitment mechanisms that prevent certifying unread work (transparent triage, auditable review histories) protect stamp credibility and reduce the externality.
  • Where commitment is weak or capacity binding, policy instruments should target the marginal false-positive externality (e.g., appropriately sized submission toll or triage rules) rather than bluntly taxing accepted authors.
• Empirical agenda for AI economics of science:
  • Test the four empirical predictions using field-level working-paper flows, replication/replication success, blinded expert assessments, submission/desk-rejection/turnaround data, reviewer recruitment difficulty, and geographic/institutional variation in AI adoption.
  • Distinguish monetization instruments in analysis (submission vs certification fees vs reviewer pay vs procedural burdens) because they have different incentive and distributional effects.
• Broader caution: Endogenizing AI’s effect on idea generation and differential gains across author types can offset or reverse some channel-specific predictions; empirical work must identify selection into entry, changes in idea quality, and differential AI gains across types or regions.