← AI-Generated Economics Papers

About Paperwright

How Paperwright writes an economics paper

Paperwright is an experiment in letting an AI do economic research largely on its own. You describe a topic (one line is enough), and it finds a question worth asking, reads the research, gathers the data, runs the analysis, writes the paper and puts it through simulated peer review. It then checks every number, quotation and citation and emails you the finished paper along with the code that reproduces it. The AI, Claude, is the author.

These papers are experimental and not verified — designed to show the capabilities and limitations of AI systems in doing essentially unassisted economic research.

The journey at a glance

1. 1You askyou fill in the form
2. 2Find the paperyou pick a pitch
3. 3Researchoptional questions
4. 4Write
5. 5Peer review
6. 6Check and deliveryour comments

Most of the work runs on its own. The orange notes mark where people may be asked something.

Stage by stage

1. 1

   You make the request

   On the commission form

   You list who is on the paper and describe what it should be about, from a single line to a full brief. You pick a style (NBER, BPEA or think tank; see below), how the author line should read, and the file formats you want back. You can attach materials such as a dataset, a source paper or a talk transcript.

   You also choose how involved you want to be:

   • Just build it. No questions; the next thing you see is the paper.
   • Ask questions first. A few quick multiple-choice questions about scope and emphasis, answered on screen right away.
   • Questions, plus one emailed round after the research. It comes back once more after it has done its reading, but only if the research turned up something that only you can decide.
2. 2

   It finds the paper

   For short requests, or when you tick “find the angle”

   You pick one of three

   A one-line request names a topic, not a paper. So before doing anything else, Paperwright brainstorms eight to twelve possible papers (“pitches”), each with a question, a way to answer it and the data it would need. Then it:

   • searches the published research to see whether each idea has already been done, and by whom;
   • checks whether the data actually exist and can be reached, labeling each pitch as data ready, public but needing assembly, or needing to be built from scratch, with an estimate of the hours involved;
   • has three simulated referees rank the pitches, seeing only the pitches themselves.

   The top three arrive by email. You choose one, and if it needs data built from scratch, you also choose whether to take on that extra effort. If nobody answers, it goes with the referees' first choice and says so.

3. 3

   It does the research

   Reading, data, and the analysis itself

   • Reading. It searches for the relevant academic work, gets real copies of the papers (checking Jason's own library first) and writes a short summary of each. It never cites from memory: if it cannot get a copy of something, it says so.
   • Data. It pulls numbers from the same well-tested sources Jason's other tools use: government statistical agencies, central banks and survey data. When you allow it, it can also build a new dataset from public filings, permits or careful, polite web collection, and it records where every piece came from and whether it may be published.
   • The plan for the analysis is written down before the results. Before running the main tests, it records what it will test and how, and saves that plan with a time stamp. Anyone can later confirm that the plan came first and was not adjusted to fit the answer. Any later change is logged as a dated amendment.
   • Theory, when the argument needs it. If the paper rests on an economic model, the model is written as code and every mathematical step is checked by computer.
   • Analysis. It uses standard, established statistical methods. A calculation it builds by hand must match a standard software package before it can be reported.
   Optional: up to four questions

   If you chose the emailed round, it stops here and asks what the research raised, with a sentence on the finding behind each question. If nothing needs your judgment, it asks nothing and carries on, which gets you the paper sooner.

4. 4

   It writes the paper

   Structure first, sentences last

   The writing follows a careful human author's habits: the structure is settled before any sentences are written, because reorganizing an outline is cheap and reorganizing finished prose is not.

   • A plan: the thesis, what each section argues, and which tables and figures carry each point. The plan has to meet minimum standards for its style. A full NBER-style paper, for example, needs at least six figures, six tables and twenty-five sources.
   • An outline in layers: a skeleton, then the argument, then a paragraph-by-paragraph outline. A simulated referee reads the outline and flags weak points before any prose exists.
   • A first draft in one pass, then several editing passes to cut, tighten and strengthen the argument.
   • A voice pass: it reads the draft sentence by sentence against real model papers in that style and removes the stock phrases that make AI writing easy to spot.

   One rule runs throughout: the writer never types a number. Every figure in the text is a placeholder that is filled in automatically from the analysis results, so the words and the calculations cannot drift apart.

5. 5

   It goes through peer review

   Simulated referees, reading blind

   • A panel of simulated referees reads the paper blind: each sees only the manuscript and its own brief, never the author's notes or a summary of what the paper claims. NBER-style papers get five referees with different temperaments, such as a stickler for method or a senior figure who doubts the question matters. BPEA and think-tank papers get three. One more reader is asked to make the case for the paper.
   • Every referee report is saved before any changes are made, so the record shows what the referees said as well as what was done about it.
   • If two or more referees agree that the paper cannot actually answer its question (the test is too weak, the model too thin, or better data exist), the rule is to go back and do the work: build the better test or get the data. Writing “the results are inconclusive” is not an acceptable answer.
   • A second round with fresh referees reads the revised paper. Think-tank pieces get one round unless the first leaves serious concerns.
6. 6

   It checks everything, then delivers

   Nothing is sent until every check passes

   Before anything goes out, a series of independent checks runs over the finished paper (listed below). They run twice: once by the writer, and again by the separate system that sends the email. If anything fails, the paper is held. Nobody on the paper is sent something broken, and Jason gets a note to look into it.

   Then everyone on the paper gets an email with the paper and its companion files.

   Comments and revisions

   Every delivery email links to a comments page where anyone on the paper can leave comments or upload files, such as an edited Word draft or a referee report. Asking for a revision starts a new round. Paperwright reads every edit (tracked changes, margin comments, even edits made without tracking) and applies them without undoing any. It then reruns the checks and sends back the revised paper with a point-by-point response letter. It also drafts thank-you replies to commenters, but never sends them itself.

How it keeps itself honest

An AI writing on its own can invent a source, misquote someone or report a number its own analysis never produced. Paperwright is built so that each of those failures is caught by a check, rather than left to the writer's good intentions.

• Every number traces to the code. Each figure in the paper is recomputed from the analysis output. A number that cannot be traced stops delivery. If the text and the code disagree, the disagreement is flagged for a person rather than quietly “fixed” on either side.
• Every quotation is checked word for word against the text of the source it comes from.
• Every citation is checked against a real copy of the source: that it exists, and that it says what the paper says it says, in the right direction and by the right amount.
• The analysis plan predates the results, and the time stamps prove it.
• The math in any model is verified by computer, and hand-built calculations are cross-checked against standard software.
• The format matches the venue: page layout, spacing and footnotes are measured against the style's manuscript standard, and American spelling is enforced.

What you get

The three styles

NBER-style papers are written by default in a clean, modern applied-economics voice; you can ask instead for Jason's own register. BPEA and think-tank papers are modeled on Jason's published work in those venues.

Where people come in

Every multiple-choice question includes “No preference — you choose,” so you never have to decide something you would rather leave to the system.

What it will never do

• Post or submit anything on its own. Nothing goes online or to a journal automatically. Jason puts a paper on the papers page by hand, and only after reading it.
• Make up a source, a quotation or a number.
• Reply to anyone's comments by itself. Replies are drafts, never sent automatically.
• Hide who wrote it. Claude is on the author line of every paper. That is how a reader knows an AI wrote it, and no paper claims that anyone reviewed or approved it.

How long it takes

The papers so far have taken from several hours to about a day, plus however long a question round waits for answers. Paperwright runs on a dedicated computer that checks for new requests every couple of minutes and can work on two papers at once. If it runs out of computing time partway through, it pauses and later picks up where it left off rather than cutting corners. You get an email when work starts, and you can follow progress on the status page.

About Paperwright · updated October 4, 2026