ChatGPT for economists on Mac: spoken help, data checks and citation traps in 2026

The ChatGPT Mac app changed on July 9, 2026. OpenAI said the Codex app was merging into the new ChatGPT desktop app for Mac and Windows. The older app was renamed ChatGPT Classic. If you already use Codex, a normal update gives you the new app.

Which ChatGPT app to open for economics work

The new app has three modes. Chat is for conversation, drafts and quick analysis. Work is for longer jobs that produce documents, sheets, slides or sites across several apps. Codex is a coding agent. For most economists, Chat is the mode you will open every day.

Chat, Work and Codex are available on every plan, including Free, according to an OpenAI FAQ quoted in a third-party write-up. Read the plan details on the ChatGPT pricing page before you pick a tier. Match the tier to the features you will use, not the headline.

Menu names and layout can change with each release. Treat any path on this page as a starting point, and read what your screen shows. Many Mac apps open Settings with ⌘, but check the menu bar first.

What economists actually use ChatGPT for

The most useful map of this work is Korinek's 2023 NBER paper. It describes 25 uses of large language models in economic research. The uses fall into six domains: ideation, writing, background research, data analysis, coding and mathematical derivations. The paper predates the current app, so read its examples as a starting list.

• Ideation: stress-test a question, or ask for the strongest objection to your identification strategy.
• Writing: tighten a paragraph, cut an abstract to a word limit, or draft a reply to referee comments.
• Background research: summarize a paper you already have, or map the main contributors to a question.
• Data analysis: describe a dataset, suggest cleaning steps, or explain what a regression table says.
• Coding: draft Stata, R or Python for a task whose output you can check against known numbers.
• Mathematical derivations: check a step, then work the step yourself before you accept it.

We found no survey that measures how economists use ChatGPT specifically. The closest study reads job-market forum posts rather than asking economists directly. The best nearby number covers all workers. In late 2024, 23% of employed US respondents had used generative AI for work in the previous week. Only 9% used it every workday. Treat those figures as context, not a benchmark for your department.

Non-work use is growing too. The NBER team tracked ChatGPT use through July 2025 and found non-work messages rose from 53% to more than 70%. Keep your research chats in a Project of their own, so they stay easy to find when a paper needs them again.

Data analysis in ChatGPT: what to upload and when it helps

Data analysis is part of the paid tiers. The pricing page lists it for Go, Plus and Pro, and as limited on Free. Upload limits decide what you can send. The table below gives the approximate limits by plan.

Start with a small sample. Export a few hundred rows with the real column names, and ask ChatGPT for cleaning steps and a checklist of likely problems. Then run those steps yourself in Stata, R or Python on the full file. Your code is the record you can defend later. For spreadsheet cleaning, see Excel for economists on Mac.

Coding and derivations with ChatGPT: how to check the output

Coding is where ChatGPT saves the most typing and where it makes the quietest errors. Ask for a Stata do-file that runs a difference-in-differences model with clustered standard errors on your sample. Then run the same regression with a second command, or by hand on a small case, and compare the coefficient. If the two disagree, trust your check, not the draft.

For derivations, ask ChatGPT to show each step with a reason. Then pick a special case with a known answer, such as a parameter value where the result is obvious, and test the result there. A step you cannot justify is a step you rewrite yourself, not a step you paste into the paper.

The 2024 arXiv experiment is the warning here. Economics students used ChatGPT on writing and Stata tasks. Average scores rose and time fell. But 42% saw no gain on math and data-analysis tasks in Stata. Higher-ability students also did worse on writing analysis with ChatGPT. Those were students, so do not read the result as a verdict on your own skill.

Pick a ChatGPT plan by page limits

Pick your plan based on the largest file you expect to send. The numbers below come from a summary of the pricing page, so confirm them on the live page before you buy.

Deep Research and file uploads scale by plan too. Deep Research is limited on Free and Go, expanded on Plus and at its maximum on Pro. File uploads are limited on Free, larger on Go, expanded on Plus and available on Pro.

ChatGPT citations for economics papers: how to check them

Treat every reference from ChatGPT as unverified until you open it. A 2026 Southern Economic Journal paper by Buchanan, Hill and Shapoval tested models including GPT-4o, DeepSeek V3 and web-enabled GPT-5.1. It found non-existent citations in economics. Error rates stay high for scholarly work, according to a co-author's blog summary. That summary gives no exact rates.

Web access does not fix this. Some of the tested models were web-enabled, and they still produced citations that do not exist. Use the steps below for every reference, including ones that look perfect.

1. Search the exact title in Google Scholar, the publisher's site or the journal's site, not inside ChatGPT.
2. Open the paper and check the authors, year and journal. Match each one to the reference.
3. Find the sentence that supports the claim you wanted to cite. If you cannot find it, drop the claim.
4. Delete any reference you cannot open. Do not patch it by hand.

Keep a short log of the references you verified, inside the Project for that paper. The log helps when a co-author asks where a claim came from. It also shows you which sources you already checked.

Writing and editing with ChatGPT without losing your voice

Writing help sits in the ideation and writing domains of Korinek's list. Keep each request narrow. Paste one paragraph, say what it must claim, and ask for two shorter versions. Reject any sentence that asserts something your results do not show.

The 2024 student study is a useful caution here. It found 34% saw no improvement on writing-analysis tasks. Keep your own draft open in another window, and compare every rewrite against it before you paste anything in.

For equations and tables in Word, see Word for economists on Mac. Those tasks need a different check, because the layout can change even when the words stay the same.

Keep one Project per paper

Projects are listed for Free, Go, Plus and Pro on the pricing page. Use one Project for each paper. Put the abstract, the variable definitions and your style rules in it. Start each chat from that Project, so you stop retyping the same setup.

Check who can see a Project before you add unpublished estimates to it. Keep those out unless every co-author on the paper has agreed to the sharing. A Project is a shared workspace in practice, so treat it like one.

Background research: when Deep Research helps and when it does not

Deep Research runs a longer search and writes a report. Use it for a narrow question with a clear end point, such as the main findings on one policy change. It is a poor fit for a broad question with no boundaries. Read the sources it lists, not just its summary.

Lazur can run research too. Research runs in Lazur's own Chrome tabs through the Lazur for Chrome extension, then comes back with a summary and a report. It is a separate job from your ChatGPT chat, so keep the two outputs apart in your notes.

Where Lazur fits next to ChatGPT

ChatGPT does the thinking and the typing inside its own window. Lazur sits beside it. Hold ⌥ ⌘ and ask about anything on screen, such as a table in a ChatGPT answer. Lazur answers out loud. It does not click buttons, edit files or operate ChatGPT for you.

Lazur listens and looks at the screen only while you hold the shortcut. Audio and the screenshot go to AI providers to answer, and they are not stored. Content is never used to train AI models. Read that policy again before you use Lazur on restricted work.

If macOS asks for screen or accessibility permission for any app, the setting lives in Apple menu > System Settings > Privacy & Security. Grant only what the app needs, and check the list again after a major macOS update.

Ask Lazur where things are and how to do them

Ask where something is, and Lazur draws a circle and a label around it on your screen. This helps when a ChatGPT update moves a button. Ask how to do something, and Lazur guides you one click at a time, waiting for each click before the next step.

If you open ChatGPT with a keyboard shortcut, read How to use the Option+Space shortcut in ChatGPT on Mac before you change any settings.

Privacy before you paste anything into ChatGPT

Start with your data agreement. Many datasets forbid sending identifiable records to outside services. Remove names, IDs, addresses and free-text fields before any sample leaves your machine. Then ask whether your university or employer allows consumer AI tools on that data.

Look at every sample before you paste it. Export files sometimes carry hidden columns or notes in an extra tab. Check the first and last columns, because identifiers often end up there.

Common mistakes economists make with ChatGPT

• Trusting a reference because it looks right. Open it first.
• Uploading a full file when a small sample shows the same problem.
• Accepting a derivation step you did not work through yourself.
• Keeping one long chat for three papers. Start a fresh chat or Project for each paper.
• Using Work mode for a one-line question. Chat is faster for that.
• Skipping your plan's page limit. Check it before a large upload.

A working session from start to finish

1. Open the ChatGPT desktop app. Choose Chat for a question, or Work for a job that spans several files.
2. Open or create the Project for your paper. Add the abstract, the variable list and your style rules.
3. Ask for a literature map. Request titles and authors only, then verify each one.
4. Upload a small sample with real column names. Ask for cleaning steps and a list of checks.
5. Run those steps yourself on the full file, and keep the code.
6. Draft one paragraph at a time. Mark every claim your results do not show.
7. Hold ⌥ ⌘ when a table or message is unclear, and ask Lazur what it says.

Pick the step that costs you the most time this week, and test the workflow on one paper first. Lazur vs ChatGPT for Mac explains where each tool fits your daily work.