AI-Generated Economics Papers: An Experiment
Can an AI do economic research largely on its own? 5 papers, and the system that produced them.
In August 2026 I built a system, which I call Paperwright, to find out how far an AI can get in writing an economics paper with almost no human help. Someone types a topic, often a single line. From that point nobody touches the work: Claude, the AI model made by Anthropic, finds a question worth asking, reads the literature, gathers the data, runs the analysis, writes the paper, puts it through simulated peer review with blind referees, checks every number, quotation and citation against its source, and delivers the finished paper with a replication package.
Read these as experiments. The papers below were produced in a single unattended run from a brief prompt, to show the capabilities and limitations of AI systems in doing essentially unassisted economic research. They are not verified, were not edited or reviewed by anyone, have not been submitted or published anywhere, and may contain mistakes. They are not my views, and some were asked for precisely to see how the system handles a hard question.
The experiment
Two pages describe it. The first explains, stage by stage, what the system does and what it refuses to do: how it picks a paper from a one-line topic, how it reads and gathers data, why the analysis plan is written and time-stamped before the results, how the simulated referees work, and the checks that stop a paper with an untraceable number from going out. The second is the request screen itself, reproduced as a non-working copy so you can see exactly what commissioning a paper asks for. The system is private; only its output is public.
The papers
Each opens as a page with its abstract and files: the PDF with its Word or LaTeX source, the process memo where the run wrote one, and the replication package with every script behind the paper, its tables and figures, and a manifest of where the data came from. Where the paper came from a Paperwright request, the prompt that produced it is shown verbatim: that text is the whole of the human input.
-
Data Centers and Retail Electricity Prices: Evidence from County Openings, 2001–2024
Using Clean Air Act permit records to date the first data center in 182 U.S. counties, 2001–2024, and EIA-861 county retail prices, the paper finds no detectable change in residential prices over the six years after a county's first facility (−0.4 log points, s.e. 0.7; the robust set rules out anything outside roughly −5 to +4.5 log points), with industrial prices falling where the facility is large, consistent with composition rather than cost shifting. A cost-allocation model shows why: a new large load contributes to fixed cost, so a price rise requires under-recovery from the facility or utility-level cost growth, and the paper bounds the implied cross-subsidy. Two blind referee rounds were run; the open items are a utility-unit headline and facility-level load, which public records do not carry.
The prompt that produced it, verbatim
Data centers and retail electricity prices in the United States. Does a county's first data center opening raise residential, commercial and industrial retail electricity prices, relative to counties that have not yet had one? Build county-level load-weighted retail prices from the EIA-861 utility files and service territories (macrolib's eia861.county_prices, 1990 to 2024; archive years are allowed as of 2026-10-01). The data-center side is frontier data: assemble a facility list with opening dates through the Stage 1D path (a public facility list, air permits for backup diesel generators as the opening proxy, cross-checked against news and building permits where the rules allow), with terms and the rights flag recorded for every source and the crosswalk tested. Write a model of utility cost recovery and rate design that signs and calibrates the pass-through of a large new load to each customer class, with at least three propositions and a quantitative exercise. Estimate with a staggered-adoption design (Sun-Abraham and Callaway-Sant'Anna against the two-way fixed-effects benchmark) and a second approach (a stacked design or state-level synthetic controls), the full battery drawn as well as tabled, heterogeneity by regulatory structure and utility ownership, and the standard NBER floors. A precise null is a finding; say what it rules out in dollars on a household bill.
-
The Triffin Margin: Who Absorbs Treasury Supply, and What It Costs
A 73-page NBER-style working paper asking who absorbs a dollar of new Treasury coupon supply and what it costs in yield, built from every coupon auction's investor-class allotments since 2000, the monthly and weekly records of official holdings, and a new series on holder composition by maturity bucket. The marginal dollar at auction is divided among investor classes roughly as the average dollar is, while the official sector absorbed the anticipated path of issuance when it was accumulating reserves (2000 to 2008) and not since 2014; the pre-registered announcement-surprise test has no first stage because Treasury's guidance removes the news, and the paper reports that in full. A multi-bucket preferred-habitat model with a quantity-inelastic official sector is calibrated so the fall in the official share lowers the yield cost of a supply surprise by a factor between 0.63 and one, which the daily data cannot pin down. The PDF, the LaTeX source and package, and the process memo are attached; a replication package is in the repo.
The prompt that produced it, verbatim
The Triffin margin: who absorbs Treasury supply, and what it costs. Does the yield cost of absorbing a dollar of new Treasury supply depend on whether the marginal buyer is a reserve manager or a private investor? The paper writes a two-buyer version of the Farhi-Maggiori hegemon model: an official sector whose demand is set by reserve needs and barely responds to price, and a private sector that pays a convenience premium but is price-elastic. The comparative static is that the yield response to a supply shock rises as the official share of the stock falls, and the model is calibrated so the implied elasticities are passport numbers. The test interacts the Phillot (2025) high-frequency Treasury supply shocks, identified from futures moves in the minutes around auction-size announcements, with the foreign-official share from TIC and the Financial Accounts, 2000 to 2025. Because the official share has fallen from roughly two fifths of the market after 2008 to under a fifth, a plain time interaction is an exogenous shock times a trend, so the identification comes from the cross-section of maturity buckets within each announcement window, where the official share differs sharply and the time trend is differenced out, and the interaction is shown to beat a time-varying coefficient before any magnitude is quoted. Bi, Phillot and Zubairy (2025) show the response is larger when debt is growing fast; Jansen, Li and Schmid (2024) and the granular-instrumental-variable work assign yield impact to the official sector and find its demand the least elastic; a Kansas City Fed note (Dilts Stedman 2025) asserts that composition now matters. The increment is the model-implied interaction with an exogenous supply shock, which none of them estimates. A referee would write: the paper turns the slogan that the Treasury market has become more price-sensitive into an estimated, model-implied interaction, and the magnitude is the one a debt manager and a CBO interest-cost projection would plug in.
This is a fresh, full-length NBER build of this pitch (a first build, delivered 2026-10-01 as paper-triffin-margin, came out short: a one-bucket model with one proposition, no related-literature section, four figures, and two referee rounds saying the quantity-based daily-window test could not carry the question, answered by an honest null). Build the full paper to the venue floors: a two-bucket or multi-asset model with at least three propositions, a calibration and a quantitative exercise; a standalone related-literature section; two distinct estimation approaches for the headline interaction, with the full robustness battery drawn as well as tabled. The Phillot high-frequency shock series sits behind an openICPSR login the engine cannot pass, so construct a supply-shock series with a demonstrable first stage from data the house serves (announcement-window yield or futures moves, auction sizes versus Treasury guidance, or the refunding-statement path), compare it with the quantity surprise, and let the loop-back rule govern what the referees find. Reuse nothing from the first build except what the Library already holds.
-
The paper is the first minimum wage study on the redesigned SIPP: 4.0 million person-months of monthly income, employment, and poverty data covering the 2018-2024 state-led $15-era wave. It finds the wage floor binds (10th-percentile elasticity about 0.25), employment of low-education workers does not detectably fall, and poverty point estimates lean favorable without deciding the literature's dispute - plus two mechanical findings showing why annual data cannot settle it: year-grain aggregation manufactures a spurious poverty 'effect' equal to its own pre-treatment drift, and within-person job-loss estimates are 60-100 percent reproduced by placebos where no increase occurred. The 36-page manuscript survived two five-referee blind rounds and a five-agent final check; every one of its 223 numeric claims traces to code output, and the full deferred-work list is in the accompanying process memo.
The prompt that produced it, verbatim
Impact of minimum wage increases on incomes, employment and poverty. Ideally use the SIPP but other data sets also most welcome instead or in addition.
-
How Financial Markets Respond to Economic Data Surprises
A BPEA-style paper measuring how asset prices respond to macro data surprises across the whole transmission chain — fed funds futures, the yield curve and its real/breakeven components, aggregate equities, and eleven sectors — over 2013–2026, with every consensus-based estimate paired against a consensus-free companion built from official release calendars back to 1999. Three findings: inflation news was sharply repriced after COVID and has come only partway back (the broader inflation complex gave back roughly three-fifths of the move, while per-unit core CPI sensitivity shows no decline); the repricing ran through real rates rather than breakevens, so it measures the perceived Fed reaction function, not an un-anchoring of inflation expectations; and equities' CPI trade reverted while the employment report re-emerged as the calendar's biggest event. Since both question rounds went unanswered, all six design calls were made by the engine and documented in the process memo; the paper passed two blind referee rounds, a full citation audit, and a 181-claim reproducibility passport.
The prompt that produced it, verbatim
How financial markets respond to economic data surprises. Should look at how it has changed over time (particularly pre- and post-COVID) as well as how different industries responds differently. Can use how the NASDAQ surprise data that is used in chartbook / macro world as well as WRDS data to capture a small event study window around data surprises. If possible look at changes in the expected FFR, short rates, long rates (both inflation breakevens, real rates and overall nominal), and equities--including in one section of the paper breaking down equities by industry.
-
Implicit Taxes and Measured Progressivity
Municipal bondholders bear a tax no distribution table records: the yield they forgo because the tax exemption is capitalized into prices. The paper measures that implicit tax household by household — Survey of Consumer Finances portfolios, marginal rates computed on the interest margin with TAXSIM, and a market-implied tax rate identified from the short end of the municipal term structure — and finds it raises the 99th-to-99.9th percentile's average effective federal income tax rate by 0.34 percentage points at the anchor, with essentially no change below the top. An experimental draft, not for citation or circulation.
Written by Claude; may contain mistakes. Nothing here has been submitted or published anywhere. The people who asked for a paper are not named. Corrections and comments are welcome at the address on the home page.