ChatGPT for Data Scientists on Mac in 2026: notebooks, SQL and spoken help with Lazur

Where ChatGPT fits in a data scientist's week on Mac

Most data scientists do not use ChatGPT for the core model. They use it for the small jobs around the model. On a Mac, that means a regex, a window function, a plot setting, or a stack trace you have never seen before. Those jobs eat more time than most people admit.

A 2025 survey of scientists who program points the same way. Of 609 respondents who named a primary generative-AI tool, 391 named ChatGPT. That is 64.2%. Only 88 picked a developer-specific tool such as GitHub Copilot, Claude Code, or Cursor, which is 14.4%. The sample is mostly academic and US-based, so read it as a signal, not a verdict on industry teams.

What studies show about scientists who use ChatGPT for code

A CHI 2025 study by Gabrielle O'Brien interviewed scientists and reviewed their logs. The paper is titled How Scientists Use Large Language Models to Program. Its main finding is simple. Scientists mostly used code models as a retrieval tool. They used them to get oriented in unfamiliar languages and libraries.

The same paper raises a risk. Code generation can influence the parameters of a scientific analysis without the person noticing. The scientists in the study built their own ways to judge whether generated code was usable. You need the same habit, even when the stakes are a dashboard and not a paper.

The survey adds a caution about feelings. Self-reported productivity averaged 3.9 on a scale of 1 to 5, from 623 answers. The authors call these perceptions, not measured code quality. The model explained very little of the variation, so do not read the score as proof that ChatGPT makes your code better.

Which ChatGPT app and plan to use on Mac in 2026

On July 9, 2026, OpenAI combined Chat, Work, and Codex in one desktop app for macOS and Windows. The original app is now called ChatGPT Classic. If the window looks different from the one you remember, this is why. Check which mode you are in before you paste anything.

Your plan decides which model you get. Free and Go users get GPT-5.6 Terra in Work and Codex. Plus, Pro, Business, and Enterprise can choose Sol, Terra, or Luna and set an effort level. Ultra mode in Work is limited to Pro and Enterprise.

This page does not list file size or upload limits. Those limits could not be confirmed from OpenAI's help center, so check them in the app before you upload a large file.

How to write ChatGPT prompts for SQL and pandas

Give ChatGPT the shape of your data, not the data itself. A schema, the column types, and the question are enough for most SQL drafts. Paste a few rows only when the values are messy and you need to see them. Name the dialect too, because PostgreSQL and BigQuery write dates and joins differently.

• Table names, column names, and types, copied from your schema or from df.dtypes.
• The grain of each table, such as one row per order or one row per user per day.
• The exact question, with the time window and any filter you need.
• The database or dialect, and your pandas version, which you can read with print(pandas.__version__).
• The output shape, such as one row per customer with a count column.

Here is a prompt that works. My orders table has order_id, user_id, created_at as a timestamp, and amount as numeric. Write PostgreSQL that returns monthly revenue per user for the last year. Explain each join and each filter.

Ask for the explanation with the code. A query with a short note on each join is easier to check. Compare the logic against the grain you wrote down. If the grain and the join disagree, the number is wrong even when the query runs.

Debugging a traceback in ChatGPT without pasting the whole notebook

Paste the last few lines of the traceback, plus the cell that raised it. The bottom line holds the exception type and message. The lines above show where the error started. Keep the failing cell and leave the rest of the notebook out.

Here is a common case. You merge orders with users on user_id. The merge runs, and it returns zero rows. No error appears. The usual cause is a type mismatch. One table holds integers, and the other holds strings that look like integers. Nothing matches.

Run df.dtypes on both tables and compare the user_id line. Then cast one side with astype before the merge, and check the row count again. Ask ChatGPT with the dtypes output and both row counts. The answer will point at the real cause, not a guess.

This pattern shows up beyond merges. Shape mismatches, missing columns, and silent empty results all leave clues in the same places. A small print call often tells you more than a full stack trace.

Checking ChatGPT code before it changes an analysis

Treat every generated line as a claim that needs a test. O'Brien's study names the quiet risk. Generated code can change an analysis parameter without you noticing. Common examples are a default join type, a dropna call, and a time zone on a timestamp column.

• Run the code on a small subset where you already know the right answer.
• Compare row counts before and after every join, filter, and groupby.
• Search the code for dropna, fillna, how=, and drop_duplicates, then confirm each one.
• Check null handling, since a missing value can drop out of a mean with no warning.
• Check time zones on any timestamp before you group by day.
• Save the prompt and the final code in the notebook, next to the result.

Ask ChatGPT to list every assumption in its answer. The list will not be complete. It still shows you where to look first. Then test those spots yourself.

Privacy rules for ChatGPT on a work Mac

Do not paste customer records, personal data, or unreleased results into a chat unless your company allows it. A chat is not a private notebook. Read your team's data policy before the first paste, and ask if the policy is unclear.

Anonymize before you ask. Replace names with codes, and swap emails for placeholders. Keep the real key in your own notebook. Bring only the shape of the problem into the chat.

Business and Enterprise admins must grant access first. Members still have to opt in. Leave it off unless you have read the rules and your admin agrees. Check the data controls in ChatGPT's settings as well.

Using Lazur next to ChatGPT on your Mac

Lazur is a voice assistant for the Mac that lives in the notch. Hold ⌥⌘, which is Option and Command together, and ask about anything on screen. Lazur answers out loud. It sees the screen only while you hold the shortcut.

This helps when your hands stay on the keyboard. Ask Lazur what a chart axis means, or why a button is grayed out. Lazur points at the thing and explains it. Ask where the model picker is, and Lazur draws a circle and a label around it.

Ask how to do something, and Lazur guides you one click at a time. It waits for each click. If Lazur cannot see a window, check Apple menu > System Settings > Privacy & Security > Screen & System Audio Recording.

Lazur does not click for you, edit files, or operate ChatGPT or any other app. It is not an add-in for ChatGPT, and it has no partnership with any app maker.

Dictating prompts and notes while you code

Hold fn to dictate clean, formatted text at the cursor in any app. That works in the ChatGPT box, a notebook cell, or a commit message. Dictation is unlimited on both paid plans.

Use the personal dictionary for product names. Speech tools often mangle pandas, scikit-learn, or groupby. Add each term once. Then the next prompt comes out right.

• A schema header for SQL prompts, with table grain and dialect.
• A standard data check: row counts, nulls, duplicate keys, and the date range.
• A source line for each notebook: the file name, the pull date, and the owner.
• A review request that asks a teammate to list the assumptions in a result.

Clipboard slots help too. Keep the schema in one slot and the traceback in another. You can switch between the notebook and the chat without retyping either one.

Common mistakes data scientists make with ChatGPT

• Accepting the first answer. It often looks right and fails on edge cases.
• Pasting full datasets. You share more than you meant, and the answer rarely gets better.
• Forgetting the library version. Method names change between releases, and old code can fail quietly.
• Asking which model to use without your data. The answer names what sounds common, not what fits your case.
• Skipping the explanation. Code you cannot explain is code you cannot defend in review.
• Losing good prompts. Save the ones that worked, because you will need them again.

The biggest mistake is the quiet one. A number looks plausible but came from a bad join. Check it once against a known value, and the habit will carry over to the next analysis.

Option+Space and quick chat for short questions on Mac

Not every question needs the full window. OpenAI's release notes list Pets quick chat via Option+Space for September 7 to 11, 2026. Use it for a one-line regex or a quick unit check. Keep longer analysis in the main window, where you can paste schemas and review code. For setup, see how to use the Option+Space shortcut in ChatGPT.

Appshots works differently. In the May 2026 release notes, pressing both Command keys sends the frontmost window to Codex. That helps when a chart or traceback is on screen. Check the window for client names and keys before you send it.

When ChatGPT is the wrong tool for a data science job

ChatGPT struggles with work that depends on state. A notebook with cells run out of order is hard to reason about in a chat. A codebase with tests needs a tool that reads the files. A chat window does not see your pipeline unless you paste it.

For that work, try an editor-based assistant. See the guides for Cursor and VS Code. In the 2025 survey, 14.4% of respondents who named a primary tool chose a developer-specific tool. Most people still start in a chat window.

A simple split works well. Use ChatGPT to think, draft, and read errors. Use your notebook or IDE to run, test, and keep the record.

Trying Lazur on a data science Mac: trial and plans

Lazur has a Pro trial that needs no credit card. The trial starts the first time you open the app. Paid plans come as Pro and Power. Dictation is unlimited on both.

Check current prices on the pricing page before you pay. Prices are kept there, so this guide does not repeat them. You can also download Lazur to try the trial on your Mac.

Lazur runs on macOS, on Apple silicon and Intel Macs. Windows is planned, not available. If your analysis work lives on a Windows laptop, Lazur will not run there yet.

Where Lazur helps and where it does not

• Helps: pointing at a button, reading an error aloud, drafting a reply, and dictating a prompt.
• Helps: voice timers, Apple Reminders, Calendar events, notes, music, and opening apps or links.
• Helps: research jobs run in Lazur's own Chrome tabs and come back with a summary and a report.
• Does not: click buttons, edit files, or operate ChatGPT or other apps for you.
• Does not: run on Windows yet, since Windows is planned, not available.

Privacy works like this. 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.

Next steps for data scientists using ChatGPT on Mac

1. Check your ChatGPT mode and plan in the desktop app.
2. Write one schema header and save it as a snippet.
3. Pick one recent query, and check its row counts on a small subset.
4. Read the Computer History rules before you change any setting.
5. Try Lazur's trial when you want voice help on your screen.