Executive synthesis
The candidate and the idea
A model can be copied like an idea and consumed like a factory. That paradox is the best place to begin the economics of artificial intelligence. On a laptop screen, a frontier model looks almost weightless. One set of weights can be copied, adapted, fine-tuned, and embedded in thousands of products. Yet every useful interaction still draws on rival things: chips, electricity, memory, networks, data rights, engineers, evaluations and—more often than the demo admits—human judgment.
Paul Romer gave economics the language for the first half of that sentence. His 1990 model begins from a deceptively simple asymmetry: ideas are not like machines. A machine used in Ludwigshafen cannot simultaneously be used in Singapore. A formula, software architecture, chemical process, or organizational recipe can. Once discovered, the same idea can be applied again and again across rival labor and capital.[1] That is why ideas can create increasing returns even when ordinary production still faces scarcity.
Generative AI makes the distinction visible and then complicates it. A model can search papers, generate molecules, write simulation code, propose mechanisms, and rank experiments. But the economy does not consume hypotheses. It consumes medicines that survive trials, materials that survive stress, code that survives integration, and decisions that survive contact with reality. The output of a model is not yet an addition to the productive knowledge stock. It is a candidate.
The thesis of this essay is therefore demanding but measurable. AI matters for growth when it increases the flow of validated, adopted, recombinable knowledge and when that knowledge diffuses widely enough to compound. Candidate abundance is not irrelevant. It is simply upstream of the economic object we ultimately care about.
Romer 1990 · reconstructed
The loop that moved technology inside economics
Select a stage. Romer separates the reusable design from the rival object built with it, then lets the accumulated stock of designs make tomorrow's research more productive.
Research sector
Purposeful effort
Research human capital combines with the inherited knowledge stock. Existing ideas are both an input into discovery and the record of solved problems.
Section 01
What Romer changed
Solow explained why capital deepening alone cannot sustain growth in output per person. Diminishing returns pull the economy toward a steady state unless an external technology term keeps improving. The model was analytically powerful and intellectually unfinished: the most important source of long-run growth arrived from outside the system.
Romer moved technical change inside. Firms and researchers allocate scarce human capital to invent new designs because those designs can earn rents. The stock of designs raises the productivity of production and future research. Growth becomes endogenous not because scarcity disappears, but because one output of investment—knowledge—scales differently from ordinary inputs.
Research human capital and the inherited stock of knowledge jointly produce new designs.
With a constant research allocation, the benchmark can sustain a constant proportional growth rate.
The multiplication by A encodes “standing on shoulders.” A discovery is not only an output; it becomes an input into future discovery. This is the intellectual source of many AI growth forecasts even when Romer is never mentioned. If machines become research workers, Hᴬ or δ may rise. If they can draw on accumulated codified knowledge, standing on shoulders may become easier. If AI helps improve the process that improves AI, the feedback can become recursive.
But Romer’s equation is a benchmark, not a law of nature. Research teams duplicate effort, scientific fields become harder to master, distant knowledge is costly to absorb, and more researchers do not always produce proportionately more validated discoveries. Jones’s semi-endogenous response allows both research congestion and a less-than-proportional contribution from the existing stock.[2][3]
Lambda captures duplication and congestion; phi separates standing on shoulders from fishing out and the burden of knowledge.
Section 02
The three-sector economy
Romer’s economy connects final output, intermediate producer durables, and research. Competitive final-goods firms combine labor, human capital, and a continuum of specialized capital goods indexed by the number of available designs. More variety makes the rival factors more productive. The mechanism is not “more patents” as an accounting exercise; it is a better toolkit that allows production to be divided into more precisely differentiated tasks.
A is the measure of available designs; x(i) is the rival durable embodying design i.
A successful inventor owns the right to use a design and can sell or rent the corresponding durable. Downward-sloping demand permits a markup over marginal cost. That temporary monopoly rent is not an accidental blemish. It is the financing mechanism for fixed-cost invention. The same markup also restricts use below the social optimum.
The model therefore contains two classic welfare wedges. Inventors cannot capture all the value their design provides to future researchers, so knowledge spillovers push the social return above the private return. At the same time, exclusion and monopoly pricing suppress diffusion. Stronger property rights can finance more upstream invention while slowing downstream recombination. Weaker rights can accelerate reuse while starving expensive frontier work of appropriable returns.
Interactive figure · non-rivalry
Copy the recipe, not the factory
A design carries a fixed discovery cost but can serve many users at almost zero copying cost. A machine must be reproduced for every user. Scale spreads the cost of the idea; it does not abolish the rival capital that embodies it.
Non-rival design
20.2
cost units per use
Rival machine
24.0
cost units per use
At 6 users, the reusable design costs 1.2× less per use than reproducing the rival asset.
Artificial intelligence sharpens rather than resolves the tension. Open weights can be copied, inspected and localized. Training them remains expensive. Inference remains capacity constrained. Proprietary data and expert validation remain scarce. Openness changes one boundary of excludability and often moves rents downstream—from model creation toward compute, applications, data, integration, evaluation and trusted domain institutions.
Section 03
A modern research production function
An AI-era extension must unpack the machine term into tasks, complementarity and validation. Begin with semi-endogenous knowledge production, then let effective research be an aggregate of task outputs. Some tasks are technically feasible for machines, some remain human, and many require combinations of both.
V is the rate at which generated output survives verification, becomes usable knowledge, and diffuses into the productive stock.
Low theta creates an O-ring process: weak stages dominate the aggregate and narrow machine excellence cannot substitute for them.
Benjamin Jones organizes the machine contribution around three parameters: γ, the share of research tasks AI can perform; M, machine productivity on those tasks; and θ, the substitutability among stages.[7] The triad converts a capability score into an economic production system. A machine that is one thousand times better at literature search can generate only a modest end-to-end gain if experimental design, physical testing or interpretation remains the weak link.
The fourth parameter is V: epistemic proof, engineering reliability, institutional acceptability and adoption. It is not an ethics appendix. It is the conversion rate from plausible output to productive knowledge. A false result can have negative social value; a correct result that no organization can absorb has little realized value; a reliable design that remains locked inside one firm creates private value without the full Romerian spillover.
Live model · not a forecast
Run the idea factory
The same machine can produce a modest level shift or a persistent acceleration. The difference lies in task coverage, complementarity, validation, and how strongly new knowledge raises future research productivity.
1.61×
effective research tasks
138
knowledge index · year 40
0.81%
annualized knowledge growth
46%
candidate survival
Candidate flow
2.31
plausible additions per model period
Validation gap
1.25
candidates that do not enter usable knowledge
Machine productivity matters. But when tasks are complements, broad coverage and validation determine whether the curve bends.
Section 04
The validation economy
When generation becomes cheap, the shadow price moves to trusted evidence, test capacity, provenance and accountable judgment. A useful decomposition treats validation as a chain rather than a single score. Epistemic validation asks whether the result survives replication and adversarial review. Engineering validation asks whether it can be embodied reliably at acceptable cost and latency. Institutional validation asks whether it is legal, safe, auditable and insurable. Adoption asks whether organizations possess the skills, incentives and complementary capital to use it.
If an essential stage approaches zero, its contribution to the usable knowledge stock approaches zero.
AI research funnel · illustrative
Generation is not yet knowledge
AI can widen the top of the funnel. The growth effect is governed by throughput at the narrowest complementary stage—not by proposal volume.
Shadow price moves to
Experiment and replication
- Stage 01
Generated candidates
Starting candidate pool
1,000
100% of initial
- Stage 02
Screened proposals
58% pass-through from prior stage
580
58% of initial
- Stage 03Binding drop
Experiments / simulations
53.4% pass-through from prior stage
310
31% of initial
- Stage 04
Validated knowledge
58.1% pass-through from prior stage
180
18% of initial
- Stage 05
Adopted innovation
55.6% pass-through from prior stage
100
10% of initial
10%
Final yield
900
Candidates filtered
53.4%
Binding pass-through
Illustrative conversion from 1,000 machine-generated candidates
Multiplicative notation is intentional. A wet lab does not become one hundred times larger because the hypothesis queue does. A regulator does not approve one hundred times faster because candidate generation accelerated. A senior scientist does not gain one hundred times more attention. Cheap cognition can therefore increase the value of laboratories, simulations, benchmarks, trusted data, expert review and regulatory science.
The product-design implication follows Bryan and Gans: an AI prediction is one component in a composite experiment involving human verification and other information.[10] Maximizing standalone accuracy can be inferior to creating errors that are legible, calibrated, differently correlated from human errors, and cheap to verify. The economically valuable model is the model that improves the human–machine system.
Recent scientific-agent releases illustrate the frontier and the constraint. Workflow-shaped benchmarks now test evidence handling, analysis and operations rather than textbook questions.[17] Agentic coding can modernize scientific software, while systems such as EvoLib explore how inference-time experience becomes reusable skill.[18][20] Yet biological-agent work still finds reliability limits in real dataset construction.[19] Capability expands the feasible set; validation determines how much enters A.
Section 05
Three ways AI can affect growth
One
A level effect
AI reduces the cost of existing research tasks and creates a one-time rise in useful knowledge. Output moves to a higher path; the long-run growth rate returns to its prior value.
Two
A growth-rate effect
AI permanently raises research productivity, expands task coverage, improves validation, or lowers the burden of knowledge. The flow of validated ideas grows faster.
Three
A recursive research-technology effect
AI improves the algorithms, experiments and tools that improve AI-assisted research itself. The process for improving the process begins to accelerate.
Recursive improvement is powerful only when capability translates into broad, reliable research-task coverage.
Charles Jones emphasizes that weak links can delay aggregate benefits even when selected capabilities become superhuman.[8] The right diagnostic question is therefore temporal: does the system improve one research cycle, every future cycle, or the mechanism by which future cycles improve? These are different claims, different models and different evidentiary burdens.
Most copilots currently look like a mixture of level and transition effects. A durable growth-rate effect requires persistence: lower discovery cost after model migration, broader task coverage over time, reusable organizational memory, and a validation system that scales with candidate generation. Recursive acceleration requires more still—the weak links themselves must become automatable without reliability collapsing.
Section 06
Romerian in the weights, Solovian in the world
The phrase identifies a layered production structure. Model weights, code and protocols can be largely non-rival once created. Training compute, inference capacity, electricity, networks and cooling are rival. Enterprise context is partly reusable within the firm but bounded by permissions and semantics. Validation often consumes scarce experts, laboratories and regulatory capacity. Diffusion requires skills, standards and organizational capital.
| Layer | Economic character | Scarce complements | Likely rents |
|---|---|---|---|
| Weights / code | Non-rival; excludability varies | Talent · training data · frontier compute | Developer · platform · ecosystem |
| Serving / inference | Rival and capacity constrained | Accelerators · electricity · memory · networks | Cloud · chip · energy · operations |
| Enterprise context | Reusable locally; legally bounded | Permissions · semantics · provenance | Data owner · integrator · application |
| Validation | Strongly rival in many domains | Experts · labs · test environments · audits | Trusted institutions · domain specialists |
| Diffusion | Shareable knowledge; costly absorption | Skills · management · standards | Complement providers · capable adopters |
Open weights do not convert the whole system into a public good. They move one boundary of exclusion. Copying cost can approach zero while operating, adapting, proving and governing remain expensive. The best policy question is therefore not “open or closed?” It is which components should be open, when, to whom, under what liability, and with which mechanisms to finance creation and independent verification.
This layered view also explains simultaneous concentration and democratization. Fixed-cost frontier creation may centralize. APIs and open artifacts can democratize downstream recombination. The same technology can strengthen hyperscalers at the top of the stack while lowering entry barriers at the edge. The distribution of complementary assets decides where the surplus settles.
Section 07
The expanding idea space
The knowledge stock A is not one library in which every book sits equally close to every reader. New evidence applies validated language models to more than 11 million U.S. patent claims from 1836 to 2023 and finds that inventions have spread apart in idea space.[5] A 98 percent decline in patent interference rates corroborates the text evidence. The resulting spatial mechanism can explain weaker spillovers, higher patent values and roughly forty percent of the long-run decline in measured research productivity.
11 million patents · conceptual map
The library is becoming a landscape
New patent evidence suggests inventions have spread apart over two centuries. A larger knowledge stock creates opportunities, but distance raises the cost of finding, absorbing, and combining them.
The frontier spreads apart
This reframes the “ideas are getting harder to find” result.[4] The frontier may be expanding as researchers occupy valuable new territory. The cost then moves from finding any idea toward navigating, understanding and combining distant ones. Standing on shoulders becomes distance weighted. A larger A creates more opportunity and a greater burden of absorption at the same time.
AI may be especially valuable as a bridge: semantic search, translation, representation learning and agents can lower the cost of crossing disciplines. But generative abundance can also expand the frontier faster, creating more branches than institutions can test. The relevant statistic is not patent or paper volume. It is successful recombination across distance at constant or improving validation quality.
Section 08
Writing code is not shipping code
The cleanest current evidence comes from a production chain we can observe. Demirer, Musolff and Yang study more than 100,000 software developers using autocomplete, interactive agents and autonomous agents.[9] The tools work. The estimated cumulative effect rises across generations. Then it decays down the output hierarchy.
Study estimates · latest tool generation
Writing code is not shipping code
Estimated cumulative effects in Demirer, Musolff & Yang (2026). The hierarchy attenuates the local gain.
Commits
Intermediate activity
Projects
Integrated work
Releases
Shipped output
72%
of the commit effect is gone by project level
83%
is gone by release level
0.25
estimated substitution elasticity
The latest generation raises commits by roughly 180 percent, projects by about 50 percent and releases by about 30 percent. The estimated elasticity of substitution between AI and human effort is 0.25, indicating strong complementarity. Review, integration, security, testing, deployment, product judgment and customer demand absorb the local gain. That is not a disappointing result. It is a production function revealing itself.
The same mechanism should be stronger in physical and regulated domains. A model can draft a credit assessment quickly while the final decision still depends on data rights, documentation, regulation and accountable ownership. It can generate a molecule while wet-lab throughput, toxicology and clinical trials remain fixed. It can generate legal reasoning while due process and appellate review remain scarce. Each successful automation raises the shadow price of the stages that remain.
Large deployments reinforce the organizational point. Evidence from courts suggests that targeted training and workflow integration shape realized effects rather than access alone.[21] Team-based invention adds another layer: more than eighty percent of U.S. patents are now produced by teams, so matching, specialization and organizational boundaries determine the effective value of research human capital.[12] AI may reduce matching frictions or increase returns to proprietary coordination. Which effect dominates is empirical.
Section 09
Is growth additive?
The strongest endogenous-growth story assumes that the next increment to productivity is proportional to the existing level. A bigger stock generates bigger absolute gains, preserving a constant percentage rate. A new empirical challenge finds that U.S. TFP increments appear conditionally additive rather than exponential.[6] Professional forecasts and international series are reported as more consistent with constant increments than constant percentages.
This does not refute non-rival ideas. It challenges the shortcut from non-rivality to exponential aggregate growth. Local idea-production systems can scale proportionally while aggregation, sectoral composition, depreciation and diffusion produce an additive macro process. The same observed short-run task gain can therefore support radically different long-run paths.
Level shift
A higher path
The stock jumps; trend growth returns.
Semi-endogenous
A long transition
Growth depends on expanding research effort and persistent tool improvement.
Proportional
A new regime
Validated ideas scale with A and permanently raise percentage growth.
AI could still create an enormous level shift: better medicine, faster engineering, cheaper services and richer organizational knowledge. That future is consequential even if the economy does not enter an explosive path. The discipline is to match the claim to the mechanism. “AI raises output” is not equivalent to “AI raises the growth rate forever,” and neither establishes recursive acceleration.
Acemoglu’s task-based macro estimates provide a conservative counterweight to extrapolation from benchmarks.[14] Research on AI as an innovation in the method of innovation makes the more ambitious case that machine systems can improve the process of invention itself.[13] Both can be internally coherent because they assign different values to coverage, translation, persistence and equilibrium adjustment.
Section 10
The institutions of diffusion
Ideas spill across organizations and borders. New causal estimates of U.S. federal R&D find that a one-percent shock to the federal research-capital stock raises foreign TFP by about one percent after twelve years, with global returns roughly twice domestic returns.[11] The country paying for research captures only part of the gain. That is Romer in the data: a non-rival idea does not stop at customs.
AI can enlarge the spillover by lowering translation, search and adaptation costs. It can also shrink it when export controls, incompatible standards, closed data and geopolitical fragmentation prevent reuse. National cost-benefit analysis will underinvest when it counts domestic rents and ignores foreign productivity. Private firms will underprovide shared validation infrastructure when its value leaks to competitors.
This is why AI infrastructure cannot mean data centers alone. The infrastructure of endogenous growth includes universities, skilled migration, public laboratories, secure data environments, independent benchmarks, standards, intellectual property, venture finance, competition, grids, laboratories, regulatory science and organizations capable of adopting what the research system discovers. Subsidizing one bottleneck can raise the price of another.
Institutions also govern epistemic diversity. AI aggregation can broaden access to knowledge while making researchers rely on correlated summaries, rankings and synthetic outputs.[16] If a few models mediate the scientific record, search cost may fall while monoculture risk rises. Independent replication, open methods, model-version records and plural evaluation become growth institutions because they protect the quality of A.
Section 11
Firm, worker, state
For the research-intensive firm
Map the whole discovery chain before selecting a model. Instrument literature search, hypothesis generation, simulation, experiment, replication, engineering, certification and commercial use. When one stage accelerates, move capital and people to the newly scarce complement. Preserve rejected designs and negative experiments as organizational memory: a failure enters A when it prevents another team from repeating it.
For workers and research leaders
Human capital does not disappear when machines enter research. Its allocation changes. Routine search and drafting may lose value. Framing a decisive question, designing a test, recognizing an anomaly, integrating distant domains and accepting responsibility may gain value. The critical skill becomes epistemic judgment: knowing what must be true before a plausible answer deserves to enter the knowledge stock.
For policymakers
Finance public validation capacity: shared laboratories, testbeds, benchmarks, secure data enclaves, standards, replication and regulatory science. Protect diffusion without eliminating appropriability through research exceptions, interoperability, prizes, procurement and time-limited rights. Preserve contestability so a concentrated model layer does not convert a growth engine into a rent-extraction system.
For investors
Look beyond the lowest inference price. Durable value may accrue to the assets that verify, embody, distribute and govern ideas: instruments, simulation platforms, lab automation, proprietary data networks, cybersecurity, regulatory tooling, grid infrastructure and firms with superior absorptive capacity. Candidate generation can commoditize while the bottleneck earns the rent.
Section 12
Build a Romer–AI observatory
The central dependent variable should be the flow of validated, adopted, reusable knowledge—not raw model output. No single dataset measures it. A credible empirical program must link research inputs, process measures, epistemic outcomes, business outcomes and spillovers. The unit should be a workflow or technology domain before it is an entire economy.
Q is epistemic quality, U is embodiment and use, and D is diffusion across projects, firms, countries or derivative inventions.
The Romer–AI observatory
Follow the whole chain
- 01
Inputs
R&D spend · researchers · compute · data access
- 02
Process
Cycle time · candidates · experiments · rejection
- 03
Knowledge
Validation · novelty · replication · reuse
- 04
Business
Time to market · quality · unit cost · TFP
- 05
Spillovers
Citations · forks · standards · global diffusion
Do not stop at prompts, tokens, benchmark scores, patent counts, or candidate volume.
The empirical challenge is selection. High-performing labs adopt early. Easier projects attract automation. Tool access is bundled with training and redesign. Improvements in data or management coincide with rollout. Usage logs alone are descriptive; a correlation between tokens and output can simply reveal that productive teams both use more AI and produce more.
The strongest enterprise design randomizes not only access, but the complement package: access alone; access plus training; access plus workflow redesign; access plus expanded validation capacity. It records model version, prompts and tool calls, human review, rejected output and rejection reasons. It follows immediate task metrics and delayed final outcomes at three, six, twelve and twenty-four months.
Patent, science and software data can connect the micro process to diffusion. PatentsView, PATSTAT and semantic patent representations can measure novelty, distance and inventor teams. OpenAlex and linked data/code records can measure scientific reuse and replication. Git repositories can separate commits, reviews and releases. Clinical and materials databases can follow candidates through physical proof. AI-generated covariates require model-version records, validation sets and uncertainty propagation so measurement technology does not manufacture the trend it is supposed to estimate.[22]
- Define the final outcome and complete workflow before deployment.
- Pre-register cycle time, quality, error, novelty, adoption and spillover measures.
- Randomize access and complementary investments where operations permit.
- Record machine cost, human effort, rejection, rework and expected error loss.
- Follow output to release, experiment, protocol, sale, standard or derivative use.
- Repeat after model migration to measure durability and organizational memory.
- Publish null results and failure modes so the knowledge stock is not a selected archive.
Section 13
What would prove us wrong?
A serious growth claim must specify the observations that would change our mind. The optimistic Romerian thesis weakens if publication, patent, code or hypothesis volume rises while replication, novelty, product entry, clinical success and TFP do not. It weakens if task time falls but total cycle time remains fixed because verification and rework expand one-for-one.
Cross-domain recombination remains flat despite better semantic search.
Frontier labs improve while smaller firms and public institutions cannot absorb the tools.
Causal gains appear only where management, data and skills changed simultaneously.
Additive TFP dynamics persist after a decade of machine-assisted research.
Correlated errors and synthetic-data feedback reduce epistemic diversity.
Model migration destroys workflows because little organizational knowledge was retained.
The thesis strengthens with sustained declines in validated discovery cost; shorter end-to-end cycles; higher replication and novelty at constant quality; faster diffusion across disciplines and countries; randomized productivity gains; and evidence that improved research tools raise the productivity of developing the next generation of research tools.
This openness to falsification matters because AI forecasts are unusually vulnerable to category errors. A benchmark becomes a task. A task becomes an occupation. An occupation becomes a sector. A sector becomes GDP. A one-time gain becomes a permanent growth rate. Each step can be defensible. None follows automatically from the previous one.
Conclusion
From possibility into knowledge
AI may be the most Romerian technology we have built: a machine for recombining non-rival ideas, a tool for producing new ideas, and potentially an input into improving the process of invention itself. But it is embedded in a Solovian world of rival capital and an institutional world of evidence, permission, trust and consequence.
The growth miracle will not arrive when machines can propose more ideas. It will arrive when economies learn to verify, diffuse and compound the right ones. The winner will not be the country or company that generates the most intelligence. It will be the one that wastes the least between possibility and knowledge.
Growth still arrives one validated idea at a time.
Generate.
plausible candidates
Validate.
reliable knowledge
Diffuse.
compounding capability
Linked bibliography
References
Research cut-off: 5 August 2026. Status labels distinguish working papers, discussion papers and institutional releases from journal publications.
- 1
Romer, Paul M. (1990).
Endogenous Technological Change. Journal of Political Economy, 98(5, Part 2), S71–S102.
- 2
Jones, Charles I. (1995).
R&D-Based Models of Economic Growth. Journal of Political Economy, 103(4), 759–784.
- 3
Jones, Charles I. (2021).
The Past and Future of Economic Growth: A Semi-Endogenous Perspective. NBER Working Paper 29126.
Working paperSource - 4
Bloom, Nicholas; Jones, Charles I.; Van Reenen, John & Webb, Michael (2020).
Are Ideas Getting Harder to Find?. American Economic Review, 110(4), 1104–1144.
- 5
Ganguli, Ina; Lin, Jeffrey; Meursault, Vitaly & Reynolds, Nicholas F. (2026).
Spreading Out Across Expanding Idea Space. NBER Working Paper 35499.
Working paperSource - 6
Jones, Callum J.; López-Salido, David & Philippon, Thomas (2026).
Is Growth Additive?. NBER Working Paper 35415.
Working paperSource - 7
Jones, Benjamin F. (2025/2026).
Artificial Intelligence in Research and Development. Working paper.
Working paperSource - 8
Jones, Charles I. (2026).
A.I. and Our Economic Future. NBER Working Paper 34779, revised June 2026.
Working paperSource - 9
Demirer, Mert; Musolff, Leon & Yang, Liyuan (2026).
Writing Code vs. Shipping Code: Productivity Effects Across Generations of AI Coding Tools. NBER Working Paper 35275.
Working paperSource - 10
Bryan, Kevin A. & Gans, Joshua S. (2026).
Training AI for When Humans Will Use It. NBER Working Paper 35490.
Working paperSource - 11
De Souza, Gustavo; Fieldhouse, Andrew J.; Mertens, Karel; Nath, Ishan B. & Ramey, Valerie A. (2026).
The Effects of U.S. Public R&D on Global Growth. NBER Working Paper 35517.
Working paperSource - 12
Kim, Seula; Olmstead-Rumsey, Jane & Wang, Honghao (2026).
Ideas and Firm Dynamics When It Takes Two to Tango. CEPR Discussion Paper 21817.
Discussion paperSource - 13
Bontadini, Filippo; Corrado, Carol; Haskel, Jonathan & Jona-Lasinio, Cecilia (2026).
AI as an Innovation in the Method of Innovation: Implications for Productivity Growth. AEA Papers and Proceedings.
- 14
- 15
Agrawal, Ajay; Gans, Joshua; Goldfarb, Avi et al. (2026).
AI in Science. NBER Working Paper 34953.
Working paperSource - 16
Acemoglu, Daron; Lin, Tianyi; Ozdaglar, Asuman & Siderius, James (2026).
How AI Aggregation Affects Knowledge. NBER Working Paper 35036.
Working paperSource - 17
- 18
OpenAI (2026).
Scientific Computing in the Age of Agentic AI. Official research release.
Institutional releaseSource - 19
Anthropic (2026).
Paving the Way for Agents in Biology. Official research release.
Institutional releaseSource - 20
Microsoft Research (2026).
EvoLib: Reusable Skills from Inference-Time Experience. Research project.
Research projectSource - 21
Ash, Elliott; Mehmood, Sultan & Goessmann, Christoph (2026).
Courts of Tomorrow: Evidence from a Nationwide Rollout of Generative AI. CEPR Discussion Paper 21783.
Discussion paperSource - 22
Duan, Junting & Pelger, Markus (2026).
Inference with AI-Generated Covariates. NBER Working Paper 35481.
Working paperSource