When AI Improves Answers but Slows Knowledge Creation: Matching and Dynamic Knowledge Creation in Digital Public Goods
summary
The gist
Generative AI helps users solve problems more efficiently but without leaving a public trace; this paper builds a dynamic model of public good provision where agents contribute by solving problems…
In short
The paper models how AI affects public knowledge archives by analyzing two margins: flow margin (private diversion of queries) and resolution margin (thinning contributor pools). It shows that AI can create multiple steady states, including a minimum viable archive. Public logging targets the flow margin, but sustaining growth requires addressing pool thinning through other mechanisms.
Key concepts
- Flow Margin
- This is the decline in the number of knowledge-enhancing queries posted to a platform that actually reach it. AI reduces this because it solves problems privately before they are publicly posted, diverting potential contributions away from the public archive.
- Resolution Margin
- This is the drop in the likelihood that a question posted on the platform gets answered. AI impacts this by raising outside options for solvers and thinning the pool of active contributors, making it harder to resolve posted problems.
- Self-Undermining Feedback Loop
- A cycle where a richer public archive attracts more outside options, which in turn thins the contributor pool. This loop slows down archive growth and can lead to multiple steady states, including a low structural minimum viable archive.
Terminology used across episodes
This episode discusses
- When AI Improves Answers but Slows Knowledge Creation: Matching and Dynamic Knowledge Creation in Digital Public Goods · Paper Radio
The paper
When AI Improves Answers but Slows Knowledge Creation: Matching and Dynamic Knowledge Creation in Digital Public Goods · Read on arXiv
Department of Economics, Fairfield University
Generative AI helps users solve problems more efficiently, but can also replace exchanges between people with private conversations with chatbots. When more problems are solved privately, fewer problem-solving episodes are converted into reusable public knowledge for others. We build a dynamic model of public good provision where agents post and answer questions on a public platform. AI can shrink the platform's archive through two margins. Private diversion can reduce the posted volume of knowledge-enhancing questions (the flow margin), while contributor exit can lower conditional resolution (the resolution margin). Low-archive traps can arise with or without AI, because a thin archive makes public answering too costly to sustain; AI can narrow the viable region. The model gives conditional empirical predictions that distinguish demand-side diversion from supply-side pool thinning under explicit identifying restrictions. Logging private solutions adds archive inflow, while rewards for answering strengthen contribution incentives.
Transcript
Introduction to the show: ident: Economics. The day in economics: theory, mechanisms and the evidence beneath them.
Ruth: I'm Ruth.
Warren: And I'm Warren, here with Iris.
Ruth: We're looking at the paper: "When AI Improves Answers but Slows Knowledge Creation".
Warren: Here's the gist of the paper. Generative AI helps users solve problems more efficiently but without leaving a public trace; this paper builds a dynamic model of public good provision where agents contribute by solving problems posted on a platform and the accumulated solutions form a depreciating public archive, revealing how AI reduces archive creation through two margins: the flow margin and the resolution margin The Two Margins of Archive Creation AI reduces archive creation through two margins that require different instruments: the flow margin and the resolution margin The flow margin is the decline in the posted volume of knowledge-enhancing queries that reach the platform, primarily driven by private diversion where AI resolves problems privately before they reach the platform The resolution margin is the decline in the probability that posted questions get answered, primarily driven by contributor pool thinning as AI raises outside options and thins contributors These two margins interact through a self-undermining feedback that can generate low-archive traps Composition of Posted Pool and Resolution Race Lemma 1 establishes the composition effect, showing that the expected knowledge increment is increasing in the ratio qeH/qeL, meaning c∗,AI(K) ≥ c∗,HO(K) if the posted pool is weakly more H-rich under AI Proposition 1 shows that resolution falls under AI whenever pool thinning dominates the combined effect of the composition shift and posted-flow reduction The sign of σAI - σHO is determined by a race among three forces: composition enrichment, posted-flow reduction, and pool thinning Dynamic Crowd-Out and Multiple Steady States Proposition 2 establishes that when AI reduces knowledge creation at the human-only steady state, the AI economy admits a steady state KAI in [0, KHO) The condition for this is equivalent to qHO H(KHO)σHO(KHO) > qAI H(KHO)σAI(K), which decomposes the decline into flow margin and resolution margin components Proposition 3 shows that a self-undermining feedback can generate multiple steady states, including a structural minimum viable archive below which the platform cannot sustain itself The Role of Public Logging Public logging acts as a conversion lever targeting the flow margin by converting private AI-assisted resolutions into public knowledge with rate η Proposition 4 shows that for any η > 0, average creation in the AI economy is greater than in the human-only economy However, conversion alone cannot repair congestion or participation-driven deterioration in conditional resolution Cross-Platform Vulnerability and Limits High routine-task share π makes platforms more exposed because AI diverts more routine queries privately and raises the outside-option shift for contributors The model suggests that high-π domains are predicted to be more exposed than low-π domains, consistent with empirical evidence Conversion offsets private diversion but does not repair congestion or participation deterioration in conditional resolution The Self-Undermining Feedback Loop A self-undermining feedback loop arises because a richer archive raises outside options, thinning the contributor pool and slowing archive growth This mechanism creates a hump shape in the average creation rate φAI(K) that generates multiple steady states, including an unstable threshold KAIU and a locally stable KAIH The structural minimum viable archive is environment-independent The Importance of Instrument Separation The two margins respond to different instruments; public logging targets the flow margin, while sustaining conditional resolution requires maintaining contributor engagement through complementary instruments like reputation systems The model isolates platform-level mechanisms that aggregate frameworks do not separate The paper concludes that the two margins require different instruments, and logging alone may be insufficient when the contributor pool has collapsed Model Architecture and Extensions The model abstracts beyond Q&A platforms to study stochastic arrival of problems at a knowledge frontier, a matching technology between problems and solvers, an outside option competing for solvers’ attention, and AI entering by improving that outside option The architecture applies with different incentive structures to academic research or private R&D The model does not claim that AI reduces overall innovation or economywide knowledge, and the conversion parameter η should be interpreted as conversion into public artifacts Parametric Example Results
Title and authors: Ruth: So we're looking at this paper titled "When AI Improves Answers but Slows Knowledge Creation: Matching and Dynamic Knowledge Creation in Digital Public Goods." It seems to be diving into how generative AI helps solve problems faster, but in doing so, it might actually make the public knowledge base smaller.
Warren: That sounds like a tension we see everywhere, right? Faster answers versus long-term learning. The authors are building a dynamic model of public good provision where people contribute solutions to a platform and those solutions form an archive that depreciates over time.
Iris: They focus on how AI reduces archive creation through two specific margins: the flow margin and the resolution margin. It’s not just one thing causing the decline in public knowledge.
Ruth: Right, let's unpack that a bit more. The paper says there's a flow margin, which is basically when the volume of new knowledge-enhancing questions posted to platforms drops because AI solves those problems privately before they ever make it onto the platform.
Warren: And then there’s the resolution margin. That’s about how likely we are to get an answer for a question that *is* posted, and that probability declines because AI raises the outside options for contributors, which thins out our pool of active contributors on the platform.
Iris: Exactly. The paper says these two margins interact in a way that creates a self-undermining feedback loop, which can lead to what they call low-archive traps where things just get worse and worse.
Ruth: That sounds like a classic congestion problem, but framed around knowledge creation rather than just traffic jams. How does the paper break down what causes that interaction?
Warren: The paper shows that the composition of the posted pool matters. They have this lemma establishing a composition effect, meaning how rich or poor in knowledge-enhancing questions our platform is affects how much AI can help, which they write as c star equals c star AI K if the posted pool is weakly more H rich under AI K.
Iris: That suggests that if the pool of questions already has a lot of high-quality stuff, like H rich content, then AI actually helps more with knowledge increment than it would in a weaker pool.
Ruth: Interesting. So it’s not just about how much we post, but what we post. Then there's this proposition one which says resolution falls under AI whenever the pool thinning effect outweighs the combined effect of that composition shift and the drop in posted flow.
Title and authors: Warren: That sounds like a race between three forces: composition enrichment, that flow reduction from AI solving things privately, and pool thinning as more people look elsewhere for answers. They determine the sign of sigma AI minus sigma HO, which is important for understanding when AI really takes over.
Iris: Then they move into dynamic crowd-out and multiple steady states. Proposition two shows that when AI reduces knowledge creation at the human-only steady state, the AI economy can settle into a new steady state KAI that’s smaller than the human-only archive KHO, provided a certain condition holds.
Ruth: That condition decomposes the decline into those flow margin and resolution margin components we talked about earlier. So it’s not just one factor causing the slowdown; it’s this specific balance between private solving and contributor thinning that dictates where things land.
Warren: And proposition three points out that this self-undermining feedback loop can create multiple steady states, including a structural minimum viable archive level below which the platform can't even sustain itself. That sounds like a real cliff for any platform trying to grow knowledge publicly.
Iris: It’s not just about the archive shrinking; it’s about the system potentially becoming unstable and falling into some kind of trap where it can't recover naturally.
Ruth: So, what is this paper suggesting we actually *do*? They introduce public logging as a way to target that flow margin directly. It acts as a conversion lever, turning those private AI-assisted solutions into public knowledge with a rate eta.
Warren: Proposition four shows that for any positive conversion rate eta, the average creation in the economy is greater than it would be if we just relied on the human-only economy. That sounds encouraging if we can get that conversion working.
Iris: But they also point out a limitation there, which is important: conversion alone cannot fix congestion or deterioration driven by participation issues in conditional resolution, meaning you can’t just log everything and expect the contributor pool to stay healthy on its own.
Ruth: That leads into the cross-platform vulnerability issue. They look at routine tasks with high routine-task share pi, and they suggest these platforms are more exposed because AI diverts more of those routine queries privately and also raises outside options for contributors there.
Warren: So, high-pi domains are predicted to be more vulnerable than low-pi ones, which aligns with what we see empirically about different types of knowledge tasks. Conversion helps offset the private diversion, but it doesn't fix the participation issues in resolution.
Title and authors: Iris: The paper highlights that the two margins respond to different tools; public logging targets flow, while keeping contributors engaged requires other things, like reputation systems for maintaining that conditional resolution.
Ruth: So what we’re seeing here is that simply having AI solve problems privately isn't the end of the story; it depends entirely on what kind of knowledge you're measuring and how you try to fix the resulting slowdown.
Warren: The self-undermining feedback loop creates a hump shape in the average creation rate, which leads to multiple steady states, including an unstable threshold KAIU and a locally stable KAIH. That suggests that there might be different long-term outcomes depending on where you start.
Iris: And the structural minimum viable archive level they find is environment-independent, which means no matter what the specific knowledge frontier is, there’s a floor below which the platform just can't survive.
Ruth: So, to wrap up this paper "When AI Improves Answers but Slows Knowledge Creation: Matching and Dynamic Knowledge Creation in Digital Public Goods," the main idea is that AI speeds up individual problem solving but shrinks the public knowledge base through flow and resolution margins.
Warren: The solution they point toward is using tools like public logging to convert those private AI solutions into public artifacts, which can boost creation if done at a rate greater than zero. But we have to be careful about what that conversion actually fixes; it doesn't fix the underlying issues with contributor participation in resolution.
Iris: I think what's most critical is recognizing that these two margins require different instruments, and you need a multi-pronged approach to keep things healthy, not just one magic fix.
Ruth: So we’ve seen how AI acts as a double-edged sword: it makes finding answers easier for individuals but complicates the process of building shared public knowledge. It seems like the path forward involves carefully managing those specific dynamics—the flow margin and resolution margin—with targeted interventions like logging, while also keeping an eye on the underlying contributor health.
Warren: Exactly. The paper lays out a dynamic picture of how these forces interact, showing that it’s not a simple cause-and-effect story but a complex system with potential traps and different stable states depending on the balance of those competing pressures.
Iris: It's an important model because it helps us diagnose whether we're seeing simple volume drops or something more structural, like the collapse of a minimum viable archive.
Ruth: We’ll leave it there for now, but keep an eye on this kind of work as we track how these digital public goods evolve with AI adoption.
The paper's summary: Ruth: So, to put it simply, this paper models how AI helps people find answers quickly but in doing so might actually make our public knowledge base smaller over time.
Warren: Yeah, it's about this dynamic system where agents contribute solutions to a platform and those solutions become part of a public archive that depreciates.
Ruth: They break down that slowdown into two main areas—the flow margin and the resolution margin—which are caused by different things.
Warren: The flow margin is when fewer people post questions because AI solves them privately before they ever get posted on the platform.
Ruth: Right, and then there's the resolution margin, which is about how much we actually get an answer for a question that *is* posted; that drops because AI raises outside options and thins out our pool of contributors.
Warren: They show these two margins interact in a way that creates this self-undermining feedback loop, which can lead to low-archive traps where the system just gets stuck.
Ruth: The paper also looked at how the composition of what's already on the platform matters—if it's already rich with high quality questions, AI helps more with knowledge creation.
Warren: And they found a condition where resolution falls under AI when pool thinning outweighs that composition boost and the flow reduction. It’s a race between three things there.
Ruth: They then showed this leads to multiple steady states, including a structural minimum viable archive level where the platform just can't sustain itself anymore.
Warren: The authors suggest public logging as a tool to target that flow margin directly, converting private AI solutions into public knowledge with some conversion rate.
Ruth: That conversion helps boost creation overall, but they warn you that it doesn't fix the participation issues in conditional resolution on its own.
Warren: And they point out that platforms dealing with routine tasks—those high routine task shares—are more exposed because AI diverts more of those things privately.
Ruth: So it’s not just a simple volume drop; it’s a structural dynamic where you have to manage both the flow and the quality of participation to keep things from collapsing into that minimum viable archive.
Warren: It highlights that you need different tools for different problems; logging fixes flow, but you still need other things to keep contributors engaged for resolution.
Iris: The identification works by looking at how these margins respond to those specific instruments—logging versus reputation systems—which is a good way to isolate the mechanism.
The paper's improvements: Tom: So, we're looking at what this paper suggests we should actually *do* with these findings about AI slowing down knowledge creation.
Ruth: They introduce public logging as a conversion lever specifically targeting that flow margin we talked about earlier, trying to turn private AI solutions into public artifacts.
Warren: Proposition four shows that for any positive conversion rate—any eta greater than zero—the average creation in the economy is larger than if we only had the human-only economy.
Ruth: That sounds encouraging, but they’re careful with the caveats, saying conversion alone can't fix congestion or participation issues in conditional resolution.
Warren: Exactly; you can’t just log everything and expect the contributor pool to stay healthy if that participation is deteriorating.
Iris: The authors argue that the two margins actually respond to different instruments; logging targets flow, but sustaining resolution needs other things, like reputation systems for keeping contributors active.
Ruth: They are trying to isolate those platform-level mechanisms because aggregated frameworks don't separate those parts as clearly as this model does.
Warren: And they use a matching technology setup in their architecture—a bridge between problems and solvers—to study the whole system dynamically.
Iris: The model moves beyond just simple Q andA platforms to look at the stochastic arrival of problems at a knowledge frontier, which is a pretty big step in scope.
Ruth: It also applies this to different incentive structures, so it could be used for academic research or even private R and D efforts.
Warren: The architecture doesn't claim AI reduces overall innovation or economywide knowledge; they say the conversion parameter eta should just be interpreted as conversion into public artifacts.
Iris: The real takeaway is that this model helps diagnose whether we're seeing simple volume drops or something more structural, like the collapse of a minimum viable archive.
Ruth: It gives us a way to design better policies for archive growth by telling us exactly which instrument—flow logging versus engagement systems—will have the biggest effect.
Warren: So it shifts the focus from just "how much knowledge is created" to "what levers do we pull to keep the system stable and growing."
Iris: It sets up a framework for understanding how these dynamics play out, which could help us think about any future digital public good systems we build.
Conclusion: Ruth: So to wrap up, this paper "When AI Improves Answers but Slows Knowledge Creation: Matching and Dynamic Knowledge Creation in Digital Public Goods" shows that AI can speed up individual problem solving, but it creates a slowdown in shared public knowledge through those flow and resolution margins.
Warren: It really frames the problem as a dynamic system where these two forces have to be balanced to avoid getting stuck in those low-archive traps we discussed.
Iris: I think the main thing they identify is that you need different levers for different problems; logging fixes flow, but you still need other things for resolution participation.
Ruth: That’s right, and it highlights how crucial it is to understand those specific margins instead of just looking at a single creation rate number.
Warren: It’s about recognizing the structure of the decline so you can choose the right instrument to target it, which is important when you’re running these kinds of models.
Iris: The model itself is pretty robust because it's built on matching and stochastic arrivals, which gives us a concrete way to test those identification assumptions.
Ruth: It changes how we think about how digital public goods evolve when AI gets integrated into the core contribution process.
Warren: If you’re tracking the economics of knowledge creation, this provides a more nuanced picture than just looking at raw output numbers.
Iris: It gives us a diagnostic tool to see if we're seeing simple volume drops or something more structural, which is really helpful for econometric testing.
Ruth: We’ll keep an eye on how these conversion parameters play out in real-world systems as AI adoption scales up.
Warren: Next time, we’re looking at how AI affects firm incentives to pursue incremental versus radical knowledge in this paper "Bridging Distant Ideas: the Impact of AI on R andD and Recombinant Innovation."
More episodes
- Daily Summary for 2026-11-19
- 2608.21274-Recommendation Quality and the Concentration of Consumption: Experimental Evidence from Netflix
- 2408.02757-A nonparametric test for diurnal variation in spot correlation processes
- 2306.16393-High-Dimensional Canonical Correlation Analysis
- 2605.17299-Geometric Brownian motion with intermittent entries and exits
- 2607.04753-Fooling Yourself: how narratives shape beliefs
- 2408.14671-Inference on Treatment Effects with Repeatedly Mismeasured High-Dimensional Controls
- 2201.06898-Difference-in-Differences Estimators for Treatments Continuously Distributed at Every Period
- 2501.15307-Influence Functions: A Unified Differentiation-Based Framework
- The day in conversation for 2026-11-19