Week 8: Catalyzing Innovation

[jamesallenevans]

Please post your memo that engages any of the week's readings about a theoretically interesting empirical case regarding artificial intelligence, innovation, and/or growth you anticipate will become the basis of your final project.

“The Rigid Disk Drive Industry: A History of Commercial and Technological Turbulence” Download “The Rigid Disk Drive Industry: A History of Commercial and Technological Turbulence”. 1993. Clayton M. Christensen. Business History Review.
“Dynamic capabilities and strategic management”. 1997. David J Teece, Gary Pisano, Amy Shuen. Strategic Management Journal 18(7):509-533.
“Battle of Ideologies: Firm Dynamics and Productivity in Planned Versus Market Economies,” Ufuk Akcigit, Richard Bräuer, Andrei Markevich, Javier Miranda, and Anna Zherdeva. Working Paper / Presented via slides in class.
“Lack of Selection and Limits to Delegation: Firm Dynamics in Developing Countries,” 2021. Ufuk Akcigit, Harun Alp and Michael Peters. American Economic Review, 111(1): 231–275.
“The Life Cycle Of Plants In India And Mexico,” 2014. Chang-Tai Hsieh and Peter J. Klenow, Quarterly Journal of Economics, 129(3): 1035-1084.
“Innovation and Top Income Inequality,” 2019. Philippe Aghion, Ufuk Akcigit, Antonin Bergeaud Richard Blundell, and David Hemous), Review of Economic Studies, 2019, 86(1): 1-45.
“Creative Destruction and Subjective Well-Being,” 2016. Philippe Aghion, Ufuk Akcigit, Angus Deaton and Alexandra Roulet, American Economic Review, 106(12): 3869-3897.

Post by Thursday @ midnight. By 12pm Friday, each student will up-vote (“thumbs up”) what they think are the five most interesting memos for that session. These memos should: 1) test out ideas and analyses you expect to become part of your final projects; and 2) involve a custom (non-hallucinated) theoretical and/or empirical demonstration that will result in the relevant analytical visualization. Some of the top-voted memos will form the backbone of discussion in Friday's discussion sessions.

[nicolaspark1]

Nicolas Park

Research Question
Does linguistic surprise in public discourse predict cross-industry reallocation before it becomes visible in income shares, firm performance, or employment data?

Motivation and Link to the Readings
This memo builds on the Schumpeterian framework developed in Aghion et al. (2019), which links innovation to rising top income shares through the expansion of high-markup sectors, and Aghion et al. (2016), which shows that creative destruction affects welfare through turnover and expected future earnings. Both papers measure innovation using realized outcomes such as patents, citations, and job turnover, and then examine how those realized processes affect inequality and well-being.
The present memo extends their framework temporally. Rather than measuring creative destruction after it appears in patent data or firm dynamics, I ask whether belief updating about substitution can be observed earlier in public discourse. If innovation reshapes expectations about which industries compete, those shifts in beliefs may precede observable changes in markups, income shares, or employment. This project, therefore, introduces an empirical measure of the anticipatory stage of creative destruction.

Empirical Case and Research Design
I use the emergence of GLP-1 receptor agonists as a cross-industry innovation shock. Beginning in mid-2023, public discussion on X increasingly linked GLP-1 drugs with reduced consumption of snacks, soda, and fast food. This provides a setting in which substitution beliefs may shift before measurable reallocation occurs in affected sectors.

The research design proceeds in three steps. First, I construct a quarterly linguistic surprise index based on semantic deviations in GLP-1 and snack-related co-mentions relative to a pre-2023 baseline. Surprise is defined as an expectation violation in cross-industry associations. Second, I compile observable outcome series including snack-sector sales growth, relative stock returns of snack and pharmaceutical firms, analyst forecast revisions, and executive compensation in relevant sectors. Third, I test whether changes in the linguistic surprise index precede inflection points in these economic outcomes using distributed-lag regressions and structural break tests.

The central hypothesis is that creative destruction begins with shifts in expectations, and that these shifts occur before realized reallocation appears in economic aggregates.

Visualization

The proposed figure plots the normalized linguistic surprise index against snack-sector sales growth over time. The surprise index rises sharply in 2023Q3, while measurable declines in snack sales follow in subsequent quarters. A shaded region indicates the estimated three-to-nine-month lead window. The figure illustrates the possibility that discourse anticipates economic adjustment rather than merely reflecting it.

Interpretation
This memo conceptualizes creative destruction as a two-stage process. The first stage consists of belief updating about substitution across industries. The second stage consists of realized reallocation in income shares, markups, and employment. The readings focus primarily on the second stage. By measuring linguistic surprise, this project seeks to make the first stage empirically observable. If belief shifts systematically precede economic reallocation, linguistic surprise may serve as a leading indicator of structural change within a Schumpeterian growth framework.

[erindner13]

Erik Rindner
Research Question

Would the adoption of AI-based biometric load management in the NBA increase league-wide economic growth by improving star player availability and reducing inefficient rest decisions?

Currently, NBA load management decisions are largely based on medical staff judgment and precautionary reasoning rather than formal predictive modeling. Star players are often rested due to perceived fatigue rather than quantified biomechanical thresholds. While precaution may reduce the risk of catastrophic injury, it may also result in underutilized labor — healthy players sitting out high-demand games. Given that star presence strongly influences ticket sales, resale prices, television ratings, and sponsorship exposure, precautionary rest may generate measurable short-run economic losses.

This creates an opportunity for technological innovation. AI-based biometric systems — incorporating wearable force sensors, motion capture analytics, and predictive injury algorithms — could transform player health management into a measurable optimization problem. Rather than relying on conservative heuristics, teams could allocate rest based on estimated injury probabilities derived from historical workload data. In economic terms, this represents an innovation in labor allocation that increases the efficiency of human capital utilization.

The growth implication is central. If AI-based optimization allows star players to safely participate in more games without increasing long-run injury risk, aggregate league revenue could rise through higher attendance, greater media engagement, and increased consumer surplus. This resembles the innovation framework in Aghion et al., where technological change can raise long-run expected welfare even if it alters short-run behavior. Here, innovation in health analytics could increase total productive output — measured in games played by high-demand players — thereby expanding the league’s economic base. In a league where marginal revenue is highly concentrated around superstar participation, even small increases in availability could generate nonlinear revenue gains. In this sense, AI load management would function as a growth-enhancing innovation rather than merely a medical tool.

Methods

To evaluate this hypothesis, I propose constructing a panel dataset of star-player participation, game-level attendance, resale ticket prices, and injury outcomes across multiple seasons. First, I would estimate the short-run economic impact of star absences by comparing games where players were rested for precautionary reasons to similar games where they played. Second, I would estimate whether precautionary rest significantly reduces subsequent injury probability.

Using these estimates, I would simulate an AI-optimized rest model based on predicted injury risk thresholds (e.g., cumulative workload, minutes played, back-to-back scheduling). Comparing observed participation rates to predicted optimal thresholds would allow estimation of unrealized economic output — the additional revenue that could be generated if healthy players participated in more games without increasing injury incidence.

By quantifying both attendance gains and injury risks, this project evaluates whether AI-based innovation in athlete management could increase sustainable league growth by improving the allocation of elite human capital.

[verachaudhry]

Firm Dynamics in Developing Countries Applied to Developing States
Vera Chaudhry

Research Question
My final project seeks to analyze the dynamics at play that are encouraging buisiness growth in certain places in the southern United States. Specifically, North Carolina is a state that has benefitted from improved business conditions – many large businesses have chosen to move to NC or expand their existing operations in the state. This is a particularly notable phenomenon given the slowdown of business dynamism nationally. The question is the following: What is driving these changes in in North Carolina?

Research Design
My hypothesis for the final project is that the change in the perception of “business friendliness” in the state prompted improved business dynamism. While my project already proposes an empirical DiD design to measure the effect of the policy change in NC compared to neighboring places that exhibit parallel trends, it will be helpful to also develop a theoretical model to explain why businesses may choose to move.

One theoretical framework can apply the model from Akcigit’s et. al.’s 2021 paper. This model captures the inefficiency of growth in certain economies, where delegation is necessary but inefficient for firms. My model could apply the framework of the paper to a new theoretical problem where cities have to make a move/stay decision (not a delegation decision) based on the inefficiencies that come with staying in a city (captured through “business friendliness” metric). Reinterpretation of variables:

• Delegation choice: firm choice to move or stay in a city.
• Alpha: efficiency of staying in a more “business friendly” city. Cutoff n*(alpha) is decreasing in alpha, so as business friendliness increases, small firms are more inclined to stay in a city. We can envision this happens through a similar mechanism of hiring managers, or through another mechanism like R&D spending.
• This reinterpretation fits within the framework of Figure 1 in the paper, because it shows how the business friendliness of a city is likely to change the slope of the profit function itself, so that more more business-friendly cities have bigger growth.

The results of the Akcigit model, reinterpreted to this new application, would show the following comparative statistics:

• Firms that move only happen above threshold n.
• Changing alpha (the business friendliness measure” has a large impact on the amount of firms that move
• If alpha is low, the optimal innovation rate is lower and large firms are more important. Changes in alpha disproportionately affect large firms. Firm expansion happens when the marginal value is higher, ie, when V(n) = V(n+1) -v(n) is greater.
  Note that this is a theoretical reframing, and to be used as an actual model, would require further modifications, especially given that the original model does not account for leaving a city. Nonetheless, the theoretical exercise is useful.

Data Visualization:
My model shows employment expansion in NC post business friendly policy at the end of 2021. The second chart shows the initial comparable trends in tax policy using state tax data.

[ukaishan]

Research question: Does innovation-driven growth improve long-run health outcomes, as measured by life expectancy across Illinois counties?
Urjita Kaishan

Do counties with higher levels of creative destruction (captured through patenting activity or job turnover) exhibit higher life expectancy after controlling for unemployment and other local economic conditions? The goal is to determine whether the growth benefits of innovation translate into measurable improvements in objective well-being, or whether the disruptive effects of creative destruction offset these gains.
This project builds directly on the Schumpeterian model of turnover-driven growth developed in Aghion et al. (2016). The authors show that turnover can increase subjective well-being when controlling for unemployment, because the growth channel outweighs the displacement channel in expectation. I extend this logic to an objective outcome (life expectancy) arguing that the same mechanisms should operate through health. Innovation may raise longevity via higher incomes, improved healthcare access, and stronger public health investment (growth channel), but it may also reduce longevity through stress, income volatility, and temporary insurance loss associated with job displacement (disruption channel). To study this question, I propose a county-level panel analysis across Illinois. The primary outcome variable will be life expectancy at birth by county and year. The key explanatory variables will include (1) patent counts per capita as a proxy for innovation intensity, (2) job turnover rates to capture creative destruction, and (3) an AI exposure index based on occupational composition to measure emerging technology shocks. The baseline regression will estimate the relationship between innovation intensity and life expectancy while controlling for unemployment rates, median income, education levels, healthcare access, demographic composition, and insurance coverage. To mirror the theoretical predictions in Aghion et al. (2016), I will compare specifications with and without unemployment controls to isolate the capitalization effect from the disruption effect. I will also test for heterogeneous effects across counties with different education levels or insurance generosity, and examine whether AI exposure amplifies dispersion in life expectancy. If feasible, I would explore instrumental variable strategies, such as historical patenting patterns or exposure to national technology shocks, to strengthen causal interpretation. Together, this empirical strategy aims to determine whether innovation generates healthier growth or exacerbates spatial health inequality.

As shown in this figure, life expectancy in Illinois counties increased from 1990 to 2010, but substantial dispersion persists across counties and between men and women. This cross-county heterogeneity provides a natural setting to examine whether innovation-driven growth translates into improvements in long-run health outcomes:

[bheim]

Which fields of research have seen the highest degree of LLM penetration? Measured by change in frequency of words commonly used by LLMS.

Guess before looking - which field do you think uses the most AI in its research production?

Context & Justification

We’ve discussed a myriad of potential effects of AI adoption on research output and reception. Our final project aims to get at this question and an important step in answering it is identifying whether or not specific papers are using AI.

By treating the introduction of AI as an exogenous shock within these fields and measuring the impact of that shock (i.e., is it actually being used), we can then better observe how AI is changing the research landscape.

To that end, I used OpenAlex data and [Kobak’s 2025 methodology](https://www.science.org/doi/10.1126/sciadv.adt3813) to determine LLM penetration across research fields. Kobak’s method focuses on looking at the uptick in specific words commonly generated by LLMs. For example, if a field suddenly starts using “delve” more frequently, this is a good proxy for that field significantly using LLMs.

These data can then be used in the next stage of our analysis to see how AI adoption has influenced research between fields.

Results

From a cursory glance, it appears that those research fields that are perhaps more ‘commercial’ have higher AI penetration than those that are more theoretical, but that’s not a clean way of explaining the data. Indeed, it is likely no surprise that computer science has seen an incredibly high usage as LLM’s impact on writing code has been massive. The most theoretical and hardest of sciences, however, (those that are arguably furthest from the description of language) see the least penetration.

With these data, we’ll attempt to estimate the following model in our next memos and final project.

[ShiyangLai]

Research Question.
Does the diffusion of AI for managerial tasks increase delegation efficiency and thereby strengthen firm growth and selection in developing countries, potentially narrowing the productivity gap documented in?

This project extends the model of delegation and firm dynamics in by reinterpreting the delegation efficiency parameter (α) as “AI-augmented delegation efficiency.” In the original framework, firms face decreasing returns due to the entrepreneur’s fixed managerial time, and can overcome span-of-control limits by hiring outside managers. The key mechanism is that higher α increases the slope of the profit function for large firms, raising expansion incentives and strengthening selection. I ask whether AI tools—used for forecasting, monitoring, coordination, or process optimization—function as a new delegation technology that effectively increases α, especially in environments where professional managerial labor is scarce.

Research Design.
I propose a cross-country structural–empirical design combining firm-level AI adoption data with dynamic moments from establishment microdata. First, I will construct a panel of manufacturing firms in a developing country (e.g., India) and a developed country (e.g., the United States) that records AI adoption intensity for managerial tasks (e.g., AI-driven ERP modules, predictive analytics, automated monitoring tools). Second, I will estimate how AI adoption affects firm life-cycle growth, exit rates, and managerial employment shares using a difference-in-differences strategy exploiting staggered AI adoption or policy-induced variation (e.g., subsidized digitalization programs).

Third, I will embed these reduced-form effects into a calibrated version of the delegation model in , allowing α to depend on AI intensity. Following the identification strategy of Nakamura and Steinsson (2018), I will treat causal micro-estimates of AI’s productivity impact as identified moments that discipline the structural parameter governing AI-enhanced delegation efficiency. The key test is whether AI adoption increases firm expansion rates more strongly for high-growth firms and whether it reduces the equilibrium share of subsistence producers—mirroring the selection mechanism emphasized in the AER model.

[TarunSethi13]

Research Question
Do reductions in corporate income tax rates increase innovation incentives by raising the returns to firm expansion, as reflected in differential firm growth across the size distribution?

Background and Motivation
In models of firm dynamics, firm size is best understood as an equilibrium outcome of firms’ decisions to invest in productivity-enhancing activities, such as organizational capital, managerial capacity, and process efficiency, rather than as a direct measure of innovation itself. Because these investments often involve fixed costs, expansion incentives are highly nonlinear: firms may face diminishing returns at small sizes, but once organizational or managerial thresholds are crossed, the marginal returns to growth increase sharply.
Corporate tax cuts can be interpreted through this lens as a policy that raises after-tax profits and therefore the expected payoff to expansion. Importantly, such a policy should not affect all firms symmetrically. Firms near expansion thresholds should respond more strongly than subsistence firms, while very large firms may adjust more slowly due to longer investment horizons and organizational inertia. The key empirical implication is a reallocation within the firm size distribution rather than uniform growth.

Research Design and Empirical Strategy
I study this mechanism using differences in corporate tax policy between North Carolina and South Carolina. The analysis centers on the November 2021 announcement of corporate tax reductions in North Carolina, which provides a clear policy signal affecting firms’ forward-looking expansion decisions.
To capture adjustment before and after this announcement, I compare firm size distributions in 2019 (pre-announcement) and 2023 (post-announcement). This window is chosen to allow sufficient time for firms to respond along margins such as hiring, organizational restructuring, and expansion, while avoiding short-run noise immediately around the announcement.
The empirical approach is descriptive and comparative. Using employment-based firm size bins, I construct a difference-in-differences measure of changes in the share of firms by size in North Carolina relative to South Carolina. Firm size growth is interpreted as an outcome reflecting firms’ responses to improved incentives for efficiency-enhancing investment, not as a direct proxy for innovation. The focus is on heterogeneity across firm sizes rather than average effects.

Figure and Interpretation

The figure plots the change in the share of firms by employment bin in North Carolina relative to South Carolina between 2019 and 2023.

The smallest firms (1–9 employees) experience a relative decline in North Carolina, suggesting that tax cuts do little to relax constraints for subsistence firms with limited growth potential. Second, there is a clear positive divergence for medium-sized firms (10–49 employees). This range plausibly corresponds to firms near organizational or delegation thresholds, where higher after-tax returns can justify investments in expansion, managerial capacity, and efficiency. Third, effects in the upper tail (100+ employees) are positive but modest, consistent with slower adjustment among large firms whose expansion decisions depend on longer-term investment cycles.
Overall, the figure indicates that tax cuts reallocate firm mass toward firms most able to translate higher post-tax profitability into productive expansion.

[csardinas]

Research Question:

Do divergent national copyright frameworks governing AI training data produce measurably different outcomes for creative sector employment and AI investment?

Background:

The United States, European Union, and Japan have adopted distinct legal approaches to the use of copyrighted works in AI training: litigation-driven fair-use ambiguity in the U.S., a mandatory opt-out regime under the EU’s Digital Single Market Directive, and a broad statutory exemption under Article 30-4 of Japan’s Copyright Act. These institutional differences create a natural comparative policy experiment directly aligned with the course’s focus on how institutional design shapes the distributional consequences of technological change.
The analytical backbone of this project draws on the Schumpeterian framework of Aghion, Akcigit, Deaton, and Roulet (2016), which models innovation as a process of creative destruction. In their framework, innovation increases worker turnover by displacing incumbents while creating new vacancies. Worker welfare reflects three competing forces: (i) an unemployment effect from increased job destruction risk, (ii) a job-finding effect from expanded hiring opportunities, and (iii) a capitalization effect, whereby faster growth raises the present value of future earnings once unemployment is controlled for. Generative AI represents a particularly salient form of innovation shock to creative industries, making this framework well-suited to interpret cross-country labor market responses.

Methodology:

For the United States, job opening rates are taken from the BLS Job Openings and Labor Turnover Survey (JOLTS) for NAICS 71, measured as job openings as a percentage of total employment plus vacancies. For the European Union, vacancy rates are drawn from Eurostat for NACE Rev. 2 category R. For Japan, where no direct vacancy-rate equivalent exists, labor demand is proxied using job postings data from Revelio Labs / CEIC, measured as the share of arts-related postings in total postings. While these measures differ in construction, all capture employer-side demand for creative labor and allow for meaningful comparison of trends.
The visual presents both raw levels (left panel) and indexed series normalized to 2019 = 100 (right panel), enabling comparison of both cross-country differences and post-2022 trajectories.

Visualization and Interpretation:

The graph reveals a sharp divergence in creative-sector labor demand after 2022. Japan exhibits a dramatic surge, with indexed demand reaching nearly three times its 2019 level by 2023. In contrast, the European Union shows moderate post-pandemic growth that stabilizes around 120–130 percent of its baseline, while the United States peaks in 2021 and then declines modestly during the AI expansion period. the timing and magnitude of Japan’s post-2021 acceleration align closely with the rapid diffusion of generative AI. Japan’s permissive copyright regime likely reduced legal uncertainty and compliance costs, facilitating faster AI adoption within creative industries and increasing demand for complementary creative roles such as AI-assisted content production and supervision. By contrast, the EU’s opt-out framework may have dampened the pace of adoption by raising regulatory frictions, producing steadier but slower growth. The U.S. pattern is consistent with high underlying innovation potential offset by legal uncertainty, where unresolved copyright litigation may have discouraged sustained hiring despite strong demand for creative content. Overall, the graph suggests that copyright institutions may meaningfully mediate how creative labor markets adjust to AI-driven innovation.

[nbcodes57]

Nick Baumgartner

Research Question:
Does the negative well-being effect of job destruction from creative destruction diminish in countries with younger average employed workforces, owing to greater labor market resilience among younger workers?

Motivation:
Aghion et al. (2016) demonstrates that job destruction depresses subjective well-being (SWB), but that this harm is moderated by unemployment insurance generosity (Prediction 3) and individual patience (Prediction 4). Their US-MSA framework, however, holds demographic composition largely constant. This memo proposes extending the logic to look across different countries: if younger workers face shorter unemployment spells, exhibit greater occupational and geographic mobility, and discount the future less steeply, then countries with younger average employed populations should exhibit a structurally attenuated well-being cost of creative destruction, independent of UI policy.

Research Design:
The proposed design is a cross-country OLS regression (with robustness via 2SLS) using country-year panel data from roughly 40–60 countries over 2010–2022. The dependent variable is country-average SWB from the Gallup World Poll (Cantril ladder). The key independent variables are: (1) job turnover rate (OECD/ILO), (2) median employed-worker age (ILO), and (3) their interaction term. The interaction coefficient is the central test: a positive sign would indicate that the negative effect of job destruction on SWB is attenuated in younger-workforce countries.

The figure above illustrates potential data supporting the hypothesis that the well-being cost of job destruction steepens markedly with workforce age. Countries in Q4 (Japan, Germany, Italy, etc.) show both a higher baseline cost and a sharper slope as destruction rates rise, consistent with older workers facing longer unemployment spells and lower re-employment probability. This information would greatly inform policy, particularly to emphasize the importance of creative destruction in developing countries with lower median age'd workforces.

[Rjlee22]

Research Question
Does the positive relationship between entrant-driven innovation and social mobility invert or collapse in commuting zones experiencing sustained population and economic decline, and if so, through what mechanisms does this inversion occur?

Proposed Research Design
This memo proposes a heterogeneous treatment effects design that interacts local innovation intensity with demographic shrinkage to estimate whether the canonical innovation-mobility relationship identified by Aghion et al. varies systematically across city types. The design is informed by Chetty et al.'s (2014) commuting-zone-level mobility estimates and extends their spatial approach by conditioning on long-run population trajectories from the Census (2000–2020).
The unit of analysis is the U.S. commuting zone (CZ). "Shrinking" CZs are defined as those experiencing population decline of 5% or more between 2000 and 2020 which is a threshold capturing approximately 180 zones concentrated in the Rust Belt, Appalachia, and rural Great Plains. Innovation intensity is measured using USPTO patent applications per 10,000 workers, disaggregated by entrant firms (under 5 years old) following the Aghion et al. framework. Social mobility is operationalized using the rank-rank slope from Chetty et al. (2014), where lower values indicate greater upward mobility.
The core empirical strategy is an OLS regression with an interaction term: Mobilityᵢ = α + β₁(Innovationᵢ) + β₂(Shrinkingᵢ) + β₃(Innovationᵢ × Shrinkingᵢ) + Xᵢγ + εᵢ. The key coefficient is β₃, which tests whether the innovation-mobility slope differs in shrinking versus growing CZs. To address endogeneity, innovation will be instrumented using historical university R&D expenditure (lagged 15 years), following Moretti (2004). Controls include local industry mix, education levels, racial composition, and baseline income.
A second-stage mechanism analysis will use IRS outmigration data (by income decile) to test whether talent flight mediates the null or inverted mobility effect in shrinking zones. This is theoretically motivated by Diamond (2016), who shows high-skill workers increasingly sorting into booming cities, potentially draining shrinking cities of the human capital needed to convert innovation spillovers into broad-based mobility gains.

Figure 1 presents the conceptual framework underlying the research design. Entrant-driven innovation is the treatment, but its effect on social mobility is moderated by city trajectory. In growing CZs, the full β₁ + β₃ pathway operates: local spillovers are absorbed through skill diffusion and supplier linkages, producing the positive mobility effect. In shrinking CZs, that pathway breaks down, talent outmigration and rent exit the region rather than diffusing downward, producing a flat or null mobility effect (β₁ alone, attenuated). The β₃ interaction term is therefore the empirical target: it isolates how much of the innovation-mobility relationship is contingent on a city's demographic vitality rather than innovation. Both pathways converge on a shared policy implication that place-blind innovation subsidies may inadvertently deepen spatial inequality.

[maksimsolomon]

Research Question

In mobile gaming YouTube advertising, does generative AI improve outcomes primarily by increasing creative experimentation (more distinct ad variants per week), and do the returns to experimentation diminish at higher levels of experimentation?

Motivation and Conceptual Background

I’m treating generative AI as a “creativity shock”: it lowers the cost of producing many ad variants (scripts, hooks, voiceovers, edits), so teams can search the creative space faster. This is consistent with evidence that AI tools can push experts toward more exploration, for example, Shin, Kim, van Opheusden, and Griffiths document that after superhuman Go AI, professionals adopted more novel moves and improved decision quality. A second idea is that what feels “easy to generate” depends on what others are already doing; Sourati and Evans emphasize that models can be steered to complement or “avoid the crowd,” which maps naturally to creative strategy (break away from dominant templates).

Research Design

Sixun’s key feedback is right: “a single studio adopts AI” is not a quasi-random shock, it’s endogenous. So I will not use studio adoption timing as identification. Instead, I will use the widespread arrival of major generative AI tools as the time variation and estimate an event study around those tool-release dates, checking pre-trends.

Data (creative-level). I will build a dataset of ad-like mobile gaming YouTube videos where each row is one creative. For each creative I record: publish date, channel/advertiser identifier, game name (when identifiable), title/description (and transcript if available), video length, and public outcomes (views, likes, comments). Then I aggregate to the channel-week (and game-week when feasible) level.

Key measures:

- Experimentation intensity: number of distinct creatives uploaded that week (distinct video IDs).

- Template diversity: share of creatives in a small set of hook families (fail montage, before/after, POV problem-solution, UGC testimonial, etc.).

- Outcomes: (i) mean performance per creative that week (average views, like-rate, comment-rate across that week’s variants), and (ii) total performance (sum of views across that week’s variants). These capture different managerial questions (average quality vs total reach).

Exposure definition (to strengthen quasi-experimental credibility). Because the shock is common across the economy, I will define “high exposure” vs “low exposure” groups ex ante using pre-shock behavior. Concretely, I define high exposure as channels in the top quartile of pre-shock creative churn (variants/week), since those workflows should benefit most when GenAI lowers production cost. Low exposure channels are the bottom quartile. I will also test a format-based exposure split (short, ad-like videos vs longer, organic content) as a robustness check.

Identification strategy (event study + exposure). I will estimate an event-study / difference-in-differences design around the tool arrival dates, comparing high-exposure vs low-exposure channels before and after the shock, and verifying parallel pre-trends.

Diminishing returns + competition. I will test whether the experimentation–performance relationship is concave (bins or flexible curves). If returns diminish, it could reflect a concave response to variety or rising competition. To separate these, I will proxy competition using the number of active creatives/channels in the same game/keyword-week and test whether the experimentation–performance slope flattens more in high-competition weeks.

Panel A (evidence): AI can shift experts toward more exploratory choices (Shin et al., 2023).
Panel B (hypothesis): In marketing, a widespread GenAI tool arrival should raise weekly creative experimentation, with diminishing marginal returns to additional variants.

This figure motivates why “AI arrival → more experimentation” is a plausible mechanism, and why the second-stage relationship should be concave.

[HenryPinner7]

Do inventors from companies acquired by large tech companies (Apple, Microsoft, Google, Meta, and Amazon) become more or less innovative after the acquisition (as measured by patent output and patent impact)?

Motivation: See Hudson's memo below for the motivation behind our project.

Progress this week: To provide some context, and has been discussed with Sixun, this memo updates my work on the research, writing, and framing components of Hudson and my final project. For the sake of writing two distinct memos, I focus here on the motivation/background, literature review, and introduction sections that we have written so far in the draft of our paper; Hudson will document our acquisition and patent data collection/cleaning pipeline in his memo.

On the research side, I completed a structured literature review of 25-30 published papers related to declining business dynamism, market concentration/incumbent advantage, acquisition-based mechanisms (including “killer acquisitions”), and the recent rise of M&A activity in technology-intensive sectors (especially after the AI-era). Some of the most important papers/studies include:

• Akcigit and Ates, 2021
• Ackigit and Attes, 2019
• Balsandze, 2021
• Cunningham, Eder, and Ma, 2018
• Bryan and Hovenkamp, 2019
• Letina, Schmutzler, and Seibel, 2024

For each paper, I extracted: (1) the core claim, (2) the mechanism linking firm behavior to dynamism/innovation outcomes, and (3) the empirical approach and identification assumptions. From there, I organized the papers into thematic clusters and wrote the preliminary draft of both the introduction and literature review sections for our final research paper.

Preliminary takeaways from this research: Rather than summarizing my findings from the literature review, I think a key outcome of my work is a clearer statement/understanding of where our project contributes to the literature. A large strand of the M&A literature, especially work on "killer acquisitions," often evaluates deals through broader, downstream outcomes: product-market effects (prices, concentration, entry) or project-specific continuation versus termination. Our central contribution is to bring the dynamism debate – between diffusion/creative-destruction accounts and incumbent-insulation/market-power accounts – to the level where diffusion and creative destruction ultimately occur: the inventive workforce. If acquisitions reduce dynamism by insulating incumbents or weakening diffusion, we should see it not only in markets or projects, but in the rate and direction of inventive effort itself – whether the engineers and inventors generating ideas at young firms continue producing new, high-impact ideas once integrated into an incumbent. We therefore take an inventor-level perspective (something less common in the studies I have read) and test whether acquisition by a frontier incumbent is associated with changes in (i) innovation quantity and (ii) innovation impact, using patent-based measures.

Moreover, the literature review also helped clarify our main hypothesis for our project, repeatedly highlighting the increased number of killer acquisitions and decreased/stalled firm productivity post-acquisition in the past 20 years, aligning with our prediction that inventor-level innovation will decrease after being acquired.

Figure connection to our data: To complement the narrative more broadly, I include two motivational figures built from the dataset my partner and I assembled using Refinitiv M&A data: (i) acquisitions per year for the five frontier incumbents in our sample and (ii) the same series for Microsoft alone. Both figures show that acquisition activity rises sharply from the late 1990s into the mid-2010s, then remains elevated through the late-2010s (with the visible dips around 2008 and 2020) – evidence that aligns with the literature emphasizing how M&A has become an increasingly central feature of the innovation landscape these firms operate in. Showing Microsoft separately also highlights meaningful within-firm variation in deal intensity, reinforcing why an inventor-level, event-timed perspective is useful rather than relying only on aggregate trends. This pattern is also consistent with our research’s emphasis on an “AI-era” shift in dealmaking: the mid-2010s acceleration in acquisitions coincides with the period many studies identify as a major inflection point in AI-related investment and consolidation (2014-2016), which in turn further motivates testing inventor-level outcomes after acquisition as a mechanism in the dynamism debate.

Next steps (for this sole dimension of the project): Finalize a polished intro/lit review draft, integrate the motivation figures directly into the argument, and tighten the handoff from “why this matters” to “how we test it.” This writing is a work in progress, but it is now grounded in a clearer literature map and a sharper statement of our project’s contribution.

[hcarpenter4]

Hudson Carpenter

Research Question:

Do inventors from companies acquired by large tech companies (Apple, Microsoft, Google, Meta, and Amazon) become more or less innovative after the acquisition (as measured by patent output and patent impact)?

Motivation:

We study whether inventors from startups acquired by Apple, Microsoft, Google, Meta, and Amazon become more or less innovative after acquisition, as measured by the difference in their total patent filings and patent citations at the acquired startup and acquiring company. The question is motivated by concerns that dominant incumbents may use acquisitions not only to scale new technologies but also to neutralize potential competitors (“killler acquisitions"), potentially dampening inventive productivity at the worker level. By examining how individual inventive careers evolve post-acquisition, we provide evidence on whether large tech M&A activity reinforces or undermines innovation-driven growth and creative destruction. We hypothesize that patent count and citations will decrease on average when an inventor moves from a target company to the acquiring incumbent, supporting the idea that big-tech acquisitions can function as “killer-acquisitions” to dampen innovative competition.

Data Collection

We first constructed a dataset of startup acquisitions by Apple, Microsoft, Google, Meta, and Amazon using LSEG Workspace. We restrict the sample to transactions classified as completed mergers, acquisitions, or majority acquisitions, yielding 704 acquired companies between 1987 and 2026.

Using this list of target companies, we compiled patent data for each acquired firm and identified 19,475 inventor-patent observations filed prior to acquisition, along with their associated inventors using Google BigQuery. To ensure accurate inventor-level tracking, we apply the Python package RapidFuzz to reconcile minor variations in naming conventions across patent records and consolidate patents by unique inventors. This cleaning process produces a sample of 4,857 distinct inventors.

Finally, to focus on inventive activity most proximate to the acquisition event, we restrict attention to patents filed within five years prior to the deal. The resulting dataset contains:

• 704 acquired companies
• 4,857 unique inventors
• 8,565 pre-acquisition inventor-patent observations filed within five years of acquisition.

The next steps are to use this list of acquired inventors to filter for patents produced by these inventors at the incumbent firm that acquired them. After that, we will join on patent citations to shift our analysis from changes in raw patent output to patent impact.

Preliminary Pre-Acquisition Results

This graph shows the average number of patents filed per inventor in each year leading up to acquisition. On average, each inventor contributed to 2.89 patents within 5 years of being acquired. The data reveals a general upward trend in patenting activity as companies approach their acquisition date, suggesting that inventors become progressively more productive in the years immediately preceding acquisition. This may indicate that rising inventive output is a key factor in making both companies and their inventors attractive targets for acquisition by large incumbent tech firms.

This chart provides the first half of our analysis, establishing a baseline for pre-acquisition patenting rates. Our next memo and phase of the project will analyze the post-acquisition patenting rates of these same inventors to estimate if there is a statistically significant change in invention rates. Our primary empirical method will be a t-test to evaluate the statistical significance of the change in mean output per inventor pre-acquisition vs. post-acquisition.

[jgoberman]

Research Question: Are hospital-level data silos preventing scale economies and slowing innovation in AI-driven radiology diagnostics?

Motivation:
Innovation led growth models emphasize that technological progress generates long run welfare gains but may be dampened by institutional frictions. In a Schumpeterian framework, Aghion et al. (2016) show that creative destruction increases welfare through growth externalities once short-run labor market disruptions are accounted for. Their analysis highlights that innovation’s aggregate benefits depend critically on whether complementary institutions allow growth effects to materialize.

Medical AI represents a similarly scale dependent innovation. Model performance improved with larger and more diverse training datasets, implying increasing returns to data scale. However, unlike digital platform industries where data is centralized, healthcare data remains fragmented across hospitals and electronic health record systems. If data silos limit pooling and interoperability, they may function as a structural friction that prevents AI systems from achieving efficient scale. In this sense, healthcare data fragmentation may dampen the growth externalities of AI in the same way that labor market frictions can dampen the welfare effects of creative destruction. Understanding whether data silos act as a binding constraint on medical AI innovation is therefore central to evaluating AI’s long run productivity impact on healthcare.

If data fragmentation prevents radiology AI systems from achieving efficient scale, it may dampen the growth externalities associated with medical AI adoption. In this sense, interoperability failures function as a structural friction, analogous to the labor market frictions that can attenuate the welfare effects of creative destruction in Schumpeterian models. Understanding whether data silos act as a binding constraint on AI-driven radiology innovation is therefore central to evaluating AI’s long-run productivity impact in healthcare.

Research Design:
To formally evaluate whether hospital level data fragmentation constrains the diffusion and impact of radiology AI, variation in hospital system integration can be exploited within a differences in differences framework. Figure 1 documents substantial fragmentation across U.S. hospital systems alongside concentration in radiology AI development, suggesting a structural mismatch between distributed data ownership and concentrated innovation supply. The following specification can be estimated:
$$Y_{ht} = \beta \left( Post_t \times LargeSystem_h \right) + \gamma_h + \gamma_t + \varepsilon_{ht}$$

$Y_{ht}$ denotes hospital level outcomes such as radiology adoption, imaging throughput, diagnostic turnaround time, or imaging intensity; $LargeSystem_h$ identifies hospitals belonging to large multihospital systems (e.g., 20+ hospitals, proxying for pooled data capacity); and $Post_t$ captures the post interoperability reform period for example, the implementation of information-blocking provisions under the 21st Century Cures Act). Hospital fixed effects $ \gamma_h$ absorb time-invariant institutional characteristics, while year fixed effects $ \gamma_t$ control for aggregate technological trends and national AI diffusion. The coefficient \beta measures whether hospitals with greater scale exhibit differential adoption or productivity gains relative to independent hospitals following reductions in data-sharing frictions.

Event study specifications can assess pre-trends and dynamic adjustment:
$$Y_{ht} = \sum_{k \neq -1} \beta_k \left( D_{t=k} \times LargeSystem_h \right) + \gamma_h + \gamma_t + \varepsilon_{ht}$$

Where D_{t=k} indexes event time relative to reform implementation. Parallel pre-trends would support identification, while post-reform divergence would indicate heterogenous diffusion linked to data scale.

Consistent with scale-dependent innovation models, a positive post-reform effect on AI adoption or radiology productivity among integrated systems would indicate that pooled data capacity enables realization of increasing returns in medical AI. Conversely, muted responses among independent hospitals would suggest that data silos operate as a binding constraint on diffusion, limiting spillovers and reinforcing structural asymmetries between fragmented data holders and concentrated AI developers.