Week 6: Constrained Innovation and its Avoidance

[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.

“Accelerating Science though Human-Aware Artificial Intelligence”. 2023. Jamshid Sourati and James Evans. Nature Human Behaviour.
“The Turing Trap: The Promise & Peril of Human-like Artificial Intelligence”. 2022. Erik Brynjolfsson. Daedalus.
“Superhuman artificial intelligence can improve human decision-making by increasing novelty”. 2023. Minkyu Shin, Jin Kim, Bas van Opheusden, and Thomas L. Griffiths. PNAS.
“Reasoning Models Generate Societies of Thought.” 2026. Junsol Kim, Shiyang Lai, Nino Scherrer, Blaise Aguera y Arcas, James Evans. preprint.
“Ten Facts on Declining Business Dynamism and Lessons from Endogenous Growth Theory,” 2021. Ufuk Akcigit and Sina Ates. American Economic Journal: Macroeconomics, 13(1): 257–298.
“What Happened to U.S. Business Dynamism?” 2023. Ufuk Akcigit and Sina Ates. Journal of Political Economy, 131(8): 2059–2124.

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.

[erindner13]

Erik Rindner

Research Question

Could experienced Major League Baseball organizations use AI-driven lifting and strength-training systems to promote innovative player development strategies, or primarily to refine traditional performance-enhancement practices?

Artificial intelligence is increasingly used in professional baseball to guide physical player development, particularly in strength training, workload management, and injury prevention. Teams now combine AI models with biometric data from wearable technologies—such as heart-rate variability, muscle recovery, and sleep metrics—to design individualized lifting programs and optimize performance. In principle, these systems enable innovation by linking specific training loads to physiological responses and long-run performance outcomes in ways that exceed human intuition. However, the impact of these tools depends on how development staff interpret and act on AI-generated recommendations.
This dynamic closely reflects the argument in Accelerating Science with Human-Aware Artificial Intelligence, which emphasizes that innovation is shaped by human expertise and incentives rather than data alone. Evans and Sourati show that AI systems embedded within dense expert communities often reinforce existing discovery paths, while disruptive innovation requires exploration beyond familiar strategies. In baseball, experienced strength and conditioning staff represent such expert clusters, grounded in traditional lifting philosophies and conservative injury-avoidance norms. As a result, AI-driven training systems may be used to optimize established routines—such as refining recovery thresholds—rather than to pursue novel strength-development strategies suggested by physiological data.
This framework generates clear empirical predictions. If AI promotes innovative player development, teams adopting AI-driven, recovery-based lifting programs should exhibit faster improvements in performance metrics over time, reflecting successful exploration of unconventional training-load combinations. If AI primarily reinforces traditional practices, performance gains should be more gradual, consistent with incremental optimization rather than experimentation. These effects are likely to vary by leadership experience, as less-entrenched staff may be more willing to follow AI recommendations that challenge conventional beliefs about training intensity and recovery.

Methods
I propose a comparative analysis of MLB teams that adopt AI-driven strength-and-conditioning systems, using biometric data. Teams would be stratified by development-staff experience, and player performance outcomes—such as strength gains, availability, and injury-adjusted performance—would be tracked before and after AI adoption. The key test is whether AI adoption accelerates performance improvement trajectories and whether these gains differ across leadership experience levels. Faster improvement among less experienced staff would suggest AI-driven innovation, while slower, steadier gains among experienced staff would indicate refinement rather than disruption.

[Arthur-Langlois]

Antitrust Enforcement, Incumbent Size, and Innovation: A Comparative Assessment

Research Question: Does antitrust regime enforcement increase innovation through incumbent firm size?

This research elaborates previous work by Akcigit & Ates (2021) to further investigate the institutional drivers of market concentration and its relationship with innovation and business dynamism more broadly. By building on an extensive body of literature, the authors document significant increases over the last six decades of market concentration measured by HHI, sales capture by largest firms, and patenting activity. I setup and test the hypothesis that antitrust regulations increase innovation by reducing firms' relative sizes and thus incentivizing innovation-based growth. This is in particular because firms –when unable to growth through anti-competitive practices and "inorganic" growth– have an incentive to innovate to gain market shares (Akcigit et al. 2023).

Literature: A large economics literature links innovation to incentives, competition, and institutions. Recent models and evidence emphasize an inverted-U between competition and innovation (Aghion et al., 2005) and show that stronger patent rights can shift innovation incentives toward incumbents (Acemoglu & Akcigit, 2012). We largely build our hypothesis on Akcigit et al. (2023), which document that larger firms tend to engage in non-competitive behavior. On antitrust and firm size, research connects enforcement to market concentration and dynamic competition: tougher merger policy and scrutiny of dominant firms can reduce “kill zones” and preserve entry (e.g., Cunningham, Ederer & Ma, 2021), while evidence on common ownership and product-market concentration suggests that weaker competitive pressure can raise markups and tilt innovation toward large incumbents (Azar, Schmalz & Tecu, 2018; De Loecker, Eeckhout & Unger, 2020). The literature has extensively studied the relationship between antitrust policies and firm size, as well as specific drivers of innovations, it hasn't focused on the explicit causal link between antitrust regulation, firm size and innovation levels.

Data: We build a 20 year comprehensive country-level dataset from multiple public sources to study antitrust regimes, firm sizes, and innovation levels. In particular, we measure antitrust and competition law using the (1) competition law index, and (2) antitrust enforcement, both collected form the Comparative Competition Project This sources allows us to measure antitrust regimes across 131 countries for years 19902-2010.
In addition, we measure mean and median firm size in given country-years using the World Bank Enterprise Surveys. Lastly, innovation is measured by patent application, share of new firms, both compiled by the World Bank.

Method: We test our specific mechanism using an Instrumental Variables (IV) approach. Our first stage looks at the impact of antitrust on firm size, and our second stage studies the impact of firm size on innovation. This design allows us to exogenize firm size variation by using antitrust regulations as an instrument, and then study its relative impact on innovation. Our argument relies on the two following assumptions: antitrust laws only impact innovation through firm size (exclusion restriction) and antitrust enforcement has a significant effect of firm size (validity condition).

More formally, our design follows the following 2SLS procedure:

$$\text{Firm Size}_{ct} = \gamma_0 + \gamma_1 \text{Antitrust}_{ct} + \gamma X_{ct} + \epsilon_{ct}$$

$$\text{Innovation}_{ct} = \beta_0 + \beta_1 \text{Firm Size}_{ct} + \beta X_{ct} + u_{ct}$$

where $X_{ct}$ denotes a set of country-level characteristics. The causal effect of firm sizes on innovation resulting from shifts in attributed to antitrust laws is our main coefficient of interest $\beta_1$.

[bheim]

Does the introduction of LLM agents with diverse backgrounds improve ideation and surprise?

Motivation and Context:

Following Professor Evans’ lecture on how researchers with different backgrounds can create more disruptive ideas through collaboration, I wanted to test if we could replicate these results with LLM agents with different backgrounds.

Methods:

I gave two ‘economist’ agents the problem of increasing the test scores of the bottom 20% performers in the US. After they reached an idea, I then introduced an education researcher, then a neuroscientist, then an urban planner, and then a microbiologist. The goal was to become increasingly distant from the problem. Each time I introduced someone, I would then record the solution and then introduce the new person. I also had an LLM rank each idea on semantic drift form the original idea, the number of fields included in the idea, feasibility, impact, and uniqueness.

Results & Discussion

The results were unimpressive, but I think this helps push forward the work of Professor Evans. Over time, while the ideas drifted more from the original idea and the corss-domain ratio somewhat improved, the feasibility and impact steadily declined. This relates to Evans’ finding that the best results come from people with similar background but with different view points. In this example, I didn’t vary view point but rather just background.

I think a strong next step for this research would be to test 2 things. The first would involve testing how telling an agent that they’re researcher X and another that they’re researcher Y (where X and Y disagree) would influence results — this would mirror the highest impact combination from Evans’ research.

I think it would also be useful to systematically vary the personalities of the models like was mentioned in the last lecture. Perhaps, we can vary the combativeness, peace-keeping, openness etc. of these different researchers and see how that influences the quality of ideas.

Lastly, I’m not a huge fan of using an LLM-as-a-judge and it may be better to develop a predictive model of paper impact and then run the hypotheses/experiments designed by the agents in that model.

For those interested, the final idea developed by the most diverse team was an experiment with four experimental arms that included tutoring, tutoring _ school means, tutoring + improved food infrastructure, and tutoring + cash transfers.

[ShiyangLai]

Shiyang Lai

Research Question.
Does the diffusion of AI tools increase the speed of knowledge diffusion while reducing its technological radius, thereby reshaping the “best-versus-the-rest” dynamics emphasized by Akcigit and Ates (2019)?

Akcigit and Ates argue that declining knowledge diffusion (δ) is the dominant force behind rising concentration, widening frontier–laggard productivity gaps, and declining business dynamism . In their model, diffusion is scalar: a higher δ enables followers to catch up more easily. My project extends this framework by distinguishing between diffusion intensity (how fast knowledge travels) and diffusion radius (how far it travels across technological space). I hypothesize that AI increases diffusion intensity but compresses diffusion radius: AI lowers search and comprehension costs for nearby technologies, accelerating incremental catch-up, while reinforcing dominant paradigms and reducing long-distance recombination.

Research Design.
Using patent-level and publication-level data from 2005–2024, I will measure two margins of diffusion before and after widespread AI adoption (post-2022). First, I construct a diffusion speed metric: the time from a frontier patent (top decile by citations within cohort) to its first follower citation. Second, I construct a diffusion distance metric: the embedding distance between citing and cited patents, using fixed semantic embeddings (e.g., SciBERT or patent-level transformer embeddings) trained on pre-2022 corpora to avoid contamination.

The empirical strategy is a difference-in-differences design comparing fields with high ex ante exposure to AI tools (e.g., software, biotech text-heavy domains) to fields with low exposure (e.g., mechanical manufacturing). The identifying variation leverages pre-2022 differences in task automation potential and AI-assisted coding or writing intensity. Outcomes include: (1) citation lag (speed), (2) mean and upper-tail citation distance (radius), (3) cross-field recombination entropy, and (4) evolution of frontier–laggard productivity gaps at the sector level.

This design is informed by Akcigit–Ates’ emphasis on the distribution of technology gaps and endogenous markups . Rather than treating diffusion as scalar, I model δ(d)—diffusion as a function of technological distance d—and test whether AI steepens this function (higher local diffusion, lower long-distance diffusion).

This analysis will demonstrate whether AI acts as a diffusion accelerator but also a diffusion concentrator, offering a structural reinterpretation of knowledge diffusion in the AI era.

[verachaudhry]

Predictive Models for Business Dynamism

Vera Chaudhry

Research question: Can AI successfully predict new hot spots for business dynamism?

Motivation: Prof. Evans’s article “Accelerating science with human-aware artificial intelligence,” successfully describes an AI model that is able to predict what the next scientific breakthroughs will be, and which scientists will make them. By testing different “random walks” between scientific research, the model in the paper is able to “yield a ~100% increase in the precision of forecasts regarding future material science discoveries” (pg. 1683). What if a similar AI modeling approach could be applied to understanding the business dynamism landscape that Prof. Akcigit describes in his 2019 paper, “What happened to U.S. Business Dynamism”?

Research Design:
From the charts below, note that the effects of a nationwide slowdown in business dynamism are felt differently across states. For example, not that states like North Carolina are experiencing higher rates of employment growth than some of their historically innovative counterpart states. This is in-line with the body of economic literature that discusses in-migration trends increasing in some southern states due to more positive business policies.

Drawing from Prof. Evans model, image a research project that uses an LLM to turn economic indicators into a prediction of future business dynamism:

• Define Ygt as the output variable: the # of new firms that will exist in geography g at time t. Use years for unit of time t and establish y as the # of firm births.
• Input a feature set of structured predictors such as age mix of population, city density, land use restrictiveness (if available), and text analysis from state hearings on business policies.
• Use a spatiotemporal model, like a hierarchial Bayesian model, to predict where future hot spots for new business growth will be by geography based on the predictive elements.
• Train the model on past data for business formation that can distinguish between historically strong innovative states and states that have more recently reached inflection points for new business growth.
  See Latex explanation of mean model and variable definitions for the predictive model.

Data Visualization:
Using BLS data I mapped the employment growth data for different states over the past 10 years (year 1 on x axis = 2015, year 11 = 2025) to show that dynamism is different across states. The dip in the numbers in the middle of the graphs is due to COVID-19.

[ukaishan]

Are occupations with higher AI impact levels projected to experience lower future job growth?

• Urjita Kaishan

This is inspired by the concept of the Turing Trap, which argues that the economic consequences of AI depend on whether it is deployed as a labor-substituting or labor-augmenting technology (Brynjolfsson).This project would examine whether occupations with higher AI impact levels are projected to experience lower future job growth. If AI primarily substitutes for labor, occupations with greater AI exposure should exhibit systematically lower projected growth. In contrast, task-based and augmentation frameworks suggest that AI reshapes the composition of tasks within occupations without necessarily reducing overall labor demand, implying no strong negative relationship between AI exposure and projected occupational growth.

As a preliminary step (visual attached below) , I construct an occupational-level measure of projected growth using current job openings and projected openings. Regression-adjusted predicted means across AI exposure categories show only modest differences in projected growth, with overlapping confidence intervals. High-AI-impact occupations do not exhibit dramatically lower projected growth relative to low-impact occupations. This descriptive evidence suggests that projected labor demand does not strongly reflect substitution-based automation at the occupational level.
I construct an occupational-level measure of projected growth using current job openings and projected openings:

This measure captures expected structural change in labor demand across occupations.
Then, I estimate a baseline model relating projected growth to AI impact levels:

Here AIImpact is categorical (Low, Moderate, High). A negative and statistically significant coefficient for high-exposure occupations would support substitution-based automation models.
Preliminary findings show that projected growth rates do not decline monotonically with AI impact level. Differences across exposure categories are modest, and confidence intervals overlap substantially. High-exposure occupations are not projected to contract dramatically relative to low-exposure occupations.

However, projected openings represent institutional forecasts rather than realized labor market adjustments. To evaluate whether AI exposure meaningfully affects actual employment dynamics, I propose extending the analysis to a longitudinal occupational panel study.
First, I would merge AI impact levels with historical employment data by occupation over time, using sources such as the Current Population Survey or BLS Occupational Employment Statistics. This will help me find realized employment growth rates before and after periods of rapid AI diffusion.
Then, I’d do a diff-in-diff design comparing employment trajectories of high- and low-AI-impact occupations before and after major AI adoption shocks (like post-2015 deep learning expansion).
The estimating equation:

Here occupation fixed effects control for baseline differences and time fixed effects capture aggregate labor market trends. And then to distinguish substitution from task restructuring, I would incorporate within-occupation wage changes and skill requirement shifts using job posting text data. If AI substitutes labor, high-impact occupations should experience employment contraction and downward wage pressure. This directly tests whether AI exposure causally alters occupational employment trajectories. By distinguishing projected equilibrium expectations from realized structural adjustment, the study contributes to the debate on whether AI drives occupational displacement or task-level reorganization.

Visual:

[genevese]

Geneva Kirk Drayson

Do “societies of thought” induced by reasoning-model agents produce measurable scaling gains in collective epistemic performance (accuracy, self-correction, calibration) as group size increases, relative to instruction-tuned agents, in Moltbook-like social interaction?

Drawing from Kim, J. et al. Reasoning Models Generate Societies of Thought. (2026). Google.

Kim et al. argue that reasoning models’ accuracy gains are partly explained by an implicit multi-perspective conversational organization (“society of thought”)—internal questioning, perspective shifts, reconciliation, and higher diversity of personality/expertise features—rather than trace length alone. They support this with benchmark sampling across BBH/GPQA/MATH/MMLU-Pro/MUSR/IFEval and controlled RL conditions showing that conversational scaffolding accelerates reasoning improvements.

This memo proposes an empirical analysis of field-style causal inference using Moltbook + controlled “agent swap” interventions, grounded in Mercier & Sperber’s argument that reasoning functions socially via adversarial exchange (as discussed in Kim et al.). Use a Moltbook corpus plus an experimental fork: recreate matched discussion threads where agents tackle verifiable tasks, varying only the agent model class (reasoning vs instruction-tuned) and group size. Moltbook provides an ecologically realistic interaction substrate (posts, comments, voting, community structure), and prior work shows it can be analyzed at scale with social-science methods.

The verifiable datasets that would be used:
Moltbook posts/comments dataset for naturalistic structure and sampling of interaction motifs. https://huggingface.co/datasets/ronantakizawa/moltbook

Mechanism:
Inject benchmark micro-tasks with objective keys: GSM8K-style arithmetic, MATH (short-answer), GPQA (multi-choice), BBH subsets, IFEval (format adherence). (Kim et al. explicitly use GPQA/MATH/BBH/MMLU-Pro/MUSR/IFEval to probe reasoning.)
Look for “society-of-thought markers” in text: question–answer sequences, explicit viewpoint shifts, conflict/reconciliation moves (codebook + LLM-as-judge, paralleling Kim et al.’s behavioral quantification).

Observable outcome: final answer correctness (ground truth), plus revision quality (did the thread move from wrong→right?), and calibration (confidence vs accuracy). We would also want to look for whether there is marginal gain per added agent and whether gains saturate differently by model class.

Preliminary experimentation with real dataset:

I analyzed deliberation/contrast markers like “however”, “on the other hand”, “tension”, “tradeoff”, “no conclusions”, inquiry markers (question marks), to create a "Society-of-Thought index" (SoT-index).

The plot shows a comparison of the posts with the lowest 25% and highest 25% SoT scores, and found that high-SoT posts have about 2x higher average AI agent engagement and have about 2x more comments. Posts exhibiting higher “society-of-thought” markers (contrast, questioning, structured argumentation) show significantly higher engagement and more sustained interaction, indicating that these conversational features are behaviorally consequential.

This establishes two key points for the proposed experiment: (1) Moltbook contains sufficient natural variation to operationalize and quantify social reasoning structure, and (2) discourse organization meaningfully affects downstream collective dynamics in a way that is empirically tractable.

It validates Moltbook as an appropriate substrate for testing whether reasoning-model agents generate scaling gains in epistemic performance via structured social exchange.

In further analysis, this would be combined with benchmarking scores similar to those used by Kim et al., and I would test whether SoT posts predict higher rates of error correction, consensus convergence, and truth-tracking.

[zararania]

Zara Rania Suran

Research Question
Does the current industry focus on developing human-level artificial intelligence in medicine, and specifically models designed to replicate the diagnostic patterns of average physicians, create a clinical Turing trap that stagnates medical innovation?

Background and Motivation
In his 2022 critique of the AI development trajectory, Erik Brynjolfsson defines the "Turing Trap" as a subtle but dangerous economic phenomenon where researchers over-index on creating machines that mimic human performance. This "Replacer" mindset leads to mere labor substitution, which may reduce costs but fails to generate the new value required for broad economic growth. When applied to the medical field, the Turing Trap is particularly seductive; we currently celebrate large language models that can pass the medical boards or radiology algorithms that match human consensus, effectively treating the current standard of care as the ground truth of medicine. However, this approach ignores the warning issued by Sourati and Evans (2023): if we train AI solely to replicate human agreement, we are essentially automating the status quo, thereby freezing medical knowledge at its current, imperfect level.

The danger of this trajectory is illuminated by the work of Shin et al. (2023), who demonstrate that the true potential of "superhuman" artificial intelligence lies not in its ability to agree with experts, but in its capacity to introduce novelty into the decision-making process. By generating strategies that appear alien or counter-intuitive to human operators, these models can expose the "invisible" spectrum of pathology that human intuition systematically misses. If we build medical AI to replace doctors, we risk declining business dynamism and a stagnation of ideas similar to the trends identified by Akcigit and Ates (2021/2023); we will get cheaper diagnostics, but we will not discover new cures. Therefore, my hypothesis posits that we must shift from an automation paradigm to an augmentation paradigm. An augmented system-where the AI is penalized for merely copying human biases and rewarded for finding orthogonal predictive features-will create a "society of thought" (Kim et al., 2026) that creates an endogenous growth loop in clinical science, effectively solving the cases that currently baffle the best human physicians.

Methods
To empirically test the divergence between these two development philosophies, I will conduct a simulation analysis using a synthetic dataset of 500 clinical cases generated with a "latent feature" space representing increasing biological complexity. I will model three distinct decision-making agents to test the hypothesis: a "human doctor" who relies on linear heuristics and excels at routine care; a "replacer AI" trained to minimize the difference between its predictions and the doctor’s labels (representing the Turing Trap); and an "augmented AI" trained on objective biological outcomes with a specific incentive to identify non-linear patterns. The simulation will subject these agents to a complexity gradient, ranging from routine diagnoses (low novelty) to high-entropy rare diseases (high novelty).

This visualization demonstrates that "replacer" AI merely automates existing medical errors, offering zero marginal value in complex cases because its failures are perfectly correlated with human failures. True innovation requires "augmenter" models that are penalized for correlating with human judgment, thereby forcing the discovery of orthogonal biological signals. Ultimately, the future of medical AI lies in systematically identifying and solving the specific edge cases that the doctor misses.

[johnab20]

John Butka

Research Question
Does the adoption of AI-driven precision agriculture by frontier firms like John Deere exacerbate the "declining business dynamism" observed in the U.S. economy by increasing the knowledge gap and reducing the "catch-up" speed of incumbent chemical and equipment competitors?

Research Design
Building on Akcigit and Ates (2021, 2023), I propose a "Knowledge Gap and Catch-up Analysis" to evaluate the competitive landscape between John Deere (the AI frontier) and incumbent chemical manufacturers like Bayer and Syngenta. Akcigit and Ates (2023) argue that a primary driver of declining dynamism is the reduction in knowledge diffusion from frontier firms to laggards.

In this design, I will utilize a Patent Citation Lag (PCL) metric. Following the "Fact 1" logic from Akcigit and Ates (2021) regarding market concentration, I will measure the time it takes for Bayer’s R&D department to cite the foundational AI and computer vision patents held by Blue River Technology (Deere’s AI arm). If Akcigit and Ates are correct, frontier firms like Deere now use "strategic patenting" to protect their lead, making it harder for "follower" firms to innovate near the frontier.
This builds upon my previous analysis of "Collaboration Proximity," where I noted that Bayer’s dispersed R&D network faces a "coordination tax". By incorporating the "Aging and Narrowing" framework, I can observe if this coordination tax translates into a "Citation Lag"—where legacy firms are not just physically distant from innovation hubs, but also temporally distant from the frontier’s knowledge.

Furthermore, I will incorporate a "Market-to-Book" (Q-ratio) Comparison. Akcigit and Ates (2021) observe that frontier firms have seen a rise in markups and valuation despite declining overall entry. By comparing the valuation of Deere’s "Precision Ag" segment against the stagnant valuation of legacy herbicide business units, I aim to show that "See & Spray" represents a "winner-takes-most" dynamic. This dynamic discourages new entrants in the AgTech space due to the high "barrier to entry" created by proprietary data moats, effectively slowing the "creative destruction" process I initially set out to explore.

This visualization illustrates a central theme of the Akcigit and Ates (2021, 2023) readings: the structural divergence between a "frontier firm" and industry laggards that drives the decline in U.S. business dynamism.

By mapping the Total Factor Productivity (TFP) of John Deere against legacy chemical incumbents, the graph captures the "productivity shock" following Deere’s 2017 acquisition of Blue River Technology, which shifted the competitive landscape from mechanical engineering to AI-driven precision agriculture. The widening shaded area, or the "Dynamism Gap," represents the slowing of knowledge diffusion; as Deere’s proprietary AI and data moats expand, the probability of "catch-up" by laggards like Bayer decreases, effectively locking the frontier.

This divergence exemplifies Fact 1 and Fact 2 from the course literature—increasing market concentration and reduced imitation—suggesting that while "See & Spray" technology is a feat of innovation, its commercial scaling may inadvertently stifle industry-wide dynamism by forcing competitors into non-productive defensive strategies rather than breakthrough research.

[jgoberman]

Research Question:
Did stronger patent protection under TRIPS reduce knowledge diffusion and young firm activity in India’s pharmaceutical sector?

Motivation:
Recent research documents a broad decline in business dynamism characterized by rising market concentration, increasing markups, widening productivity gaps between frontier and laggard firms, and declining startup activity. In their synthesis on these trends, Akcigit and Ates (2021) argue that many of these patterns can be jointly explained by a decline in knowledge diffusion. In their Schumpeterian framework, weaker diffusion allows leading firms to maintain technological advantages for longer periods, increasing markups and profit shares while discouraging entry and reducing competitive pressure.

The pharmaceutical industry provides a compelling setting in which to examine this mechanism. Over the past three decades, intellectual property (IP) protection has strengthened globally, especially following the implementation of the TRIPS agreement under the World Trade Organization. Stronger patent enforcement increases the private returns to innovation, but it may also reduce knowledge spillovers by limiting imitation, reverse engineering, and follow on innovation, mechanisms that are especially important in pharmaceuticals, where cumulative innovation and generic entry historically played a large role in industry dynamism.

The pharmaceutical sector is uniquely suited to test the diffusion mechanism emphasized by Akcigit and Ates. First, patents are economically central to the industry’s business model, making IP policy a clear and measurable shock. Second, the sector exhibits rising concentration, increasing markups, and substantial productivity gaps between multinational frontier firms and smaller domestic producers, patterns consistent with several of the ten facts on declining dynamism. Third, many developing and middle-income countries experienced discrete changes in patent regimes due to TRIPS compliance, providing variation that can be empirically exploited.

Research Design:
To formally evaluate whether India’s 2005 TRIPS compliant patent reform reduced knowledge diffusion and young firm activity in pharmaceuticals, I propose exploiting the reform as a discrete policy shock within a difference-in-differences framework. Figure 1 documents a structural break in pharmaceutical patenting activity around 2005 alongside observable changes in business entry dynamics, while Figure 2 highlights the institutional sequence from TRIPS adoption (1995) to the 2005 product patent enforcement reform. I will estimate:

$$ Y_{st} = \beta \left( Post_{2005,t} \times Pharma_s \right) + \gamma_s + \gamma_t + \varepsilon_{st} $$

where $Y_{st}$​ denotes sector-level outcomes (entry rates, concentration, or markups), $Pharma_s$ identifies the pharmaceutical sector (high IP exposure), and $Post_{2005,t}$ captures the post-reform period. Sector and year fixed effects control for time invariant sector characteristics and aggregate macro shocks. The coefficient $\beta$ measures whether pharmaceuticals exhibit differential changes relative to less patent-dependent sectors after reform. Event study specifications will test for pre-trends and dynamic adjustment during the 1995–2005 mailbox transition period. Consistent with the diffusion-based mechanism in Akcigit and Ates (2021), a negative effect on entry combined with rising concentration or markups would indicate that stronger patent protection reduced spillovers and shifted industry composition toward incumbent dominance.

Figure 1. Pharmaceutical Patent Data (WIPO), New Business Registrations (World Bank) Figure 2.

[lianduann]

Lian Duan

Research Question
Does the adoption of Generative AI (GenAI) mitigate linguistic frictions in global innovation? Specifically, does GenAI function as a “Linguistic Equalizer” that decouples the quality of scientific ideas from the researcher’s English proficiency? I ask whether the post-2022 GenAI shock has disproportionately reduced the “English penalty” for non-native researchers, thereby enhancing the diffusion of knowledge from the Global South to the scientific frontier.

Motivation
Recent literature suggests GenAI may reshape the diffusion of knowledge through three mechanisms: Reducing Frictions in Knowledge Diffusion. Akcigit and Ates (2021) identify the decline in knowledge diffusion from frontier to laggard firms as a key driver of declining business dynamism. Language barriers act as a significant transaction cost in this diffusion process. GenAI may lower this cost, allowing high-quality ideas from non-native speakers to enter the global "idea market" more efficiently. Augmentation vs. The Turing Trap. Brynjolfsson (2022) warns of a “Turing Trap” where AI merely mimics human intelligence to displace labor. However, in scientific writing, GenAI may instead serve as a complementary tool that augments the communication capabilities of non-native scholars, allowing them to compete based on the merit of their reasoning rather than their syntax. Unlocking “Alien” Hypotheses. Sourati and Evans (2023) argue that human-aware AI can accelerate science by identifying “alien” hypotheses that human experts might overlook. By lowering linguistic barriers, GenAI may prevent "Perspective Collapse" (Kim et al., 2026) and allow distinct, non-Western interpretive angles to surface, enriching the global stock of knowledge.

Proposed Research Design
The empirical strategy utilizes a large-scale bibliometric analysis of arXiv preprints, employing a Difference-in-Differences (DiD) framework around the release of ChatGPT. Data Source: I will construct a panel dataset merging arXiv metadata (for text and timing) with OpenAlex (for author affiliation and country identification). Key Metric:

1. Linguistic Fluency (Mechanism): Textual perplexity scores and readability indices (e.g., Flesch-Kincaid) to measure the "smoothing" of abstract text.
2. Innovation Impact (Outcome): Early-stage citation counts (within 12 months) to proxy for idea diffusion. Event-Study / DiD Framework: Using November 2022 (ChatGPT release) as the cutoff, I will compare outcomes between the Treatment Group (Non-Native English regions, e.g., China, Brazil) and the Control Group (Native English regions, e.g., US, UK). The primary hypothesis is that the gap in textual fluency and citation rates between these two groups significantly narrows post-2023.
   Identification: Geography as Exposure. Following the logic of Akcigit and Ates (2019) regarding leader-laggard dynamics, I identify "Linguistic Laggards" based on the country of the primary affiliation in OpenAlex. I will control for author productivity and field-specific fixed effects to isolate the "translation effect" from general quality improvements.

Feasibility
To ensure data currency, I will utilize the OpenAlex API to retrieve affiliation data for arXiv papers up to early 2025. Identification will focus on fields with high international participation but high linguistic barriers (e.g., Computer Science and Economics).

Visualization
To illustrate the "Linguistic Equalizer" effect, I present a hypothetical event-study plot of Mean Textual Perplexity for Non-Native vs. Native authors (2019–2025). A vertical line marks Nov 2022. I expect to see parallel pre-trends followed by a sharp convergence (a "kink") in the Non-Native series post-2022, while the Native series remains relatively flat.

[elektra773]

Elektra Makris

Research Question: Does proprietary AI-driven drug discovery reduce knowledge diffusion and business dynamism in biotechnology?

AI-driven drug discovery increasingly relies on large clinical datasets, genomic repositories, and model-training infrastructure controlled by major pharmaceutical firms. Unlike earlier periods of biotechnology that depended heavily on open scientific spillovers and publicly accessible genomic data, many contemporary AI systems derive value from proprietary training data that cannot easily be replicated. If access to these data is concentrated, smaller firms may find it harder to catch up to incumbents, potentially changing the competitive structure of the industry.

AI adoption creates two plausible possibilities. On one hand, AI may lower the cost of searching for new compounds and make experimentation cheaper. If that is true, we should see more entry and less concentration. On the other hand, if AI-driven innovation depends heavily on proprietary datasets, it may reduce effective knowledge diffusion. In endogenous growth models of strategic innovation, lower diffusion shifts markets toward “unleveled” states in which leaders remain ahead, concentration rises, and follower innovation declines (Akcigit & Ates, 2021; Akcigit & Ates, 2023). Under this mechanism, AI could entrench incumbents rather than democratize innovation.

Research Design
To distinguish between these possibilities, I exploit the rise of deep learning-based drug discovery around 2015 as a technological break. I classify AI-biotech patents using machine learning-related keywords (e.g., “neural network,” “deep learning,” “training data”) and compare them to non-AI biotech patents in a difference-in-differences framework. The treatment group consists of AI-tagged biotech subclasses, and the control group consists of other biotech subclasses without ML language.

At the subclass-year level, I measure concentration using the Top-4 assignee share and track the share of patents attributed to new entrants. The key test is whether concentration rises and entry declines more sharply in AI-intensive subclasses after 2015 relative to non-AI subclasses. Rising concentration alone would not be sufficient evidence of reduced diffusion: it could also reflect increasing fixed costs or scale economies. However, a joint pattern of higher concentration and declining entry would be consistent with the discouraged-follower mechanism described by Akcigit and Ates (2021).

Top-4 concentration in AI-tagged subclasses after 2015 relative to non-AI subclasses.

Decline in new assignee participation. The figures are illustrative and depict the patterns expected if AI reduces effective knowledge diffusion.

[csardinas]

Research Question

Do countries with high TFP but low human capital show signs of Turing Trap vulnerability in their current economic structure, and does the strength of convergence (2010-2020) vary systematically by labor share and human capital levels?

Background

Brynjolfsson's Turing Trap argues AI either augments or substitutes labor depending on production structure. Akcigit-Ates provide the mechanism: labor share = 1 - μ(1 - 1/λ), where μ is market concentration. When knowledge diffusion (δ) declines, economies shift toward concentrated sectors with low labor shares. I argue current economic structure reveals which countries face augmentation versus substitution. Capital-intensive economies (high TFP, low labor share, low HC) achieved productivity through resource extraction, not human knowledge. They exhibit low effective δ—workers aren't the source of productivity, so there's minimal knowledge to augment. Human capital-intensive economies (mid TFP, high labor share, high HC) derive productivity from skilled workers, exhibiting high effective δ that AI can amplify.

Research Design

Using Penn World Table 2010-2020 data, I test whether convergence strength varies by production structure. First, I split my sample by labor share (<35% vs >50%) and human capital (<2.5 vs >3.0). Then I estimate separate beta convergence regressions:
TFP_Growth[2010-2020] = β₀ + β₁(Initial_TFP[2010]) + β₂(HC[2010]) + β₃(Labor_Share[2010]) + ε
I predict β₁ is more negative (stronger convergence) for high labor share, high HC countries. For low labor share, low HC countries, β₁ approaches zero or goes positive (weak/no convergence), indicating their high TFP is unsustainable. Second, I construct the Turing Trap Risk Index = (TFP₂₀₂₀ × Capital_per_Worker)/(HC₂₀₂₀ × Labor_Share₂₀₂₀) and test whether high-risk countries showed TFP deceleration from 2010-2015 to 2015-2020, while low-risk countries showed acceleration. The prediction is that high-risk countries experience productivity deceleration in the later period as substitution effects dominate, while low-risk countries maintain or accelerate productivity growth through labor-complementary adoption.
Finally, AI readiness measures are incorporated to examine whether countries positioned to adopt AI technologies experience declining labor shares rather than productivity catch-up. Linking AI readiness to labor share outcomes using the Turing Trap framework allows identification of whether technological adoption primarily replaces labor in productivity leaders while remaining complementary in middle-income economies.
Together, these tests evaluate whether convergence dynamics depend on production structure and whether AI adoption risks reinforcing productivity leadership through substitution rather than enabling broader global catch-up.

Visualizations

Turing Trap vulnerability plot shows whether countries combining high productivity with low labor share face substitution risk. These visuals show that AI readiness and productivity leadership do not automatically imply shared prosperity. Instead, they may signal exposure to automation-driven substitution, supporting the central research question: whether AI reinforces existing leaders while leaving middle-income countries on different development paths.

[tepiv1]

Ved Jethani

How would reducing the threshold for required FTC approval in acquisitions improve TFP for biotech firms?

Motivation:
This question is partly motivated by Professor Akcigit’s discussion of talent hoarding in private industry and once again the idea of creative destruction. Looking at the space of acquisitions broadly, it is plausible that pharmaceutical companies might acquire prospective drugs which they can develop with more resources. However, this study aims to expand on the notion raised in Killer Acquisitions () who find that a high proportion of acquisitions occur just below the FTC clearance limit and that an outsized proportion of projects acquired just below this threshold are not seen through to commercialization. When under the threshold, the FTC can still investigate the acquisition but companies are allowed to close the acquisition at any point. This makes behavior that stunts innovation much easier to go unnoticed or unchallenged by regulation which is intended to do so.

Research Design:

I propose a difference-in-differences study to see how TFP differs when the threshold for acquisition size is lowered (making it harder for killer acquisitions to occur). However, since TFP is typically measured using revenue, an alternative method must be used due to many early stage pharmaceutical companies having 0 revenue. Instead, I would use citations on patents per $1 million spent on R&D. This comparably measures the scientific yield and productivity of the firm with what resources they are given and in a more meaningful manner (dubbed Knowledge TFP). By looking at firms who are close to the FTC threshold, it is possible to look at the difference in knowledge TFP before and after the threshold is changed. For example, a firm might even know they are going to be acquired and preemptively become less productive. But if the FTC threshold for approval was moved to below the deal size, the firm would continue along the previous growth trajectory.

The image below uses simulated data where one line shows incumbent players who are not affected by the threshold and therefore continue to become more productive. Firms within the “scrutiny zone” actually could outperform as they might be at an inflection point of growth and given they have already reached past a point where they will be acquired to be shut down, they are only incentivized to grow and disrupt. In reality, to collect data I would aim to compare companies who are below scrutiny one year and above the next year to see if there is a clear difference in knowledge TFP. To see how lowering the threshold impacts TFP, we could look at 2019 where it was lowered (but COVID-19 might muddy the data), however other years with small/no raises might serve as useful points to look at on a firm level.

[clsun2004]

Chloe Sun

Research Question:

Can reasoning-based artificial intelligence function as a quasi-entrepreneurial agent—generating novel recombinations independent of individual human cognition—and does this shift innovation from individual-level entrepreneurship to AI-augmented organizational search?

Motivation:

Schumpeter’s theory of growth centers on the entrepreneur: a rare individual who recombines existing knowledge in novel ways to generate economic transformation. But what happens if the cognitive function of recombination can be partially externalized to machines?

Recent work suggests this possibility is no longer speculative. Sourati and Evans (2023) argue that AI accelerates science by expanding the accessible search space of ideas. Shin et al. (2023) show that superhuman AI increases novelty in decision-making. Kim et al. (2026) describe reasoning models as generating “societies of thought,” structured internal deliberation that mirrors collective cognition. Together, these papers imply that reasoning AI is not merely automating routine tasks (Brynjolfsson’s “Turing Trap”), but potentially altering the mechanism of innovation itself.

At the same time, Akcigit and Ates (2021, 2023) document declining business dynamism in the U.S.—fewer startups, less job reallocation, and greater concentration. A central concern in endogenous growth theory is that innovation has become increasingly dependent on frontier firms and star inventors.

If reasoning AI lowers the cognitive cost of distant recombination, two opposing equilibria are possible:

• Democratization equilibrium: AI reduces reliance on rare “star” entrepreneurs and broadens who can innovate.

• Concentration equilibrium: AI amplifies already dominant firms, reinforcing declining dynamism.

This project tests whether we are witnessing the emergence of “artificial Schumpeterian entrepreneurs” embedded within firms—systems that partially substitute for individual entrepreneurial cognition.

Research Design:

The empirical strategy exploits staggered firm-level adoption of reasoning-based AI tools between 2022 and 2025. AI exposure can be measured using enterprise licensing announcements, AI-related job postings, or earnings-call AI intensity measures.

The central empirical question is whether AI adoption changes who drives innovation within firms.

1. Measuring Novel Recombination

Innovation outcomes will include:

• Patent combinatorial novelty (distance between CPC codes)
• Atypicality of technology pairings
• Citation originality scores
• Text-embedding distance from the firm’s prior knowledge base

These measures capture whether AI increases distant recombination rather than incremental refinement.

2. Measuring Dependence on Individual Entrepreneurs

To test whether AI substitutes for individual-level entrepreneurial ability, the analysis decomposes innovation variance into:

• Firm-level effects
• Year effects
• Inventor-level heterogeneity

The key prediction: If reasoning AI acts as an artificial entrepreneur, the share of innovation variance attributable to inventor-specific ability should decline after AI adoption.

Visualization:

To illustrate why this question matters, consider the distribution of innovation across inventors. If innovation depends heavily on rare entrepreneurial talent, a large share of variance in novelty should be explained by inventor-specific effects. If reasoning AI reduces cognitive constraints, innovation may become less dependent on those fixed individual differences.

A useful visualization is the share of total innovation variance attributable to inventor ability over time. A decline after AI adoption would suggest entrepreneurial cognition is becoming organizational rather than individual.

Below is simulated data and a random effects model to illustrate this variance decomposition. My code computes the share of total variance in novelty explained by differences across inventors and plots how this share evolves over time.

The figure captures the core theoretical tension of the project.

If the share of variance attributable to inventor differences declines after AI adoption, it suggests that innovation is becoming less dependent on rare entrepreneurial talent. That would support the hypothesis that reasoning AI acts as an artificial Schumpeterian entrepreneur embedded within firms.

If instead inventor-level variance remains dominant—or increases—AI may simply amplify existing hierarchies.
The broader implication is that AI may not just increase productivity. It may alter the structure of entrepreneurship itself, with profound consequences for business dynamism and long-run growth.