How to Analyze Data With ChatGPT (and What to Never Trust)

To analyze data with ChatGPT, upload the spreadsheet instead of pasting it, ask one specific question at a time, and have it show the calculation behind every number. An uploaded file lets ChatGPT run real code on your data. A pasted table is text in the chat, and the reply will not always tell you whether a total was calculated or estimated. Even a calculated number needs one more check: whether it used the right rows.

The workflow that holds up

1. Upload the file, don't paste it. OpenAI's help center says ChatGPT can run Python calculations on files you give it, and recommends structured data with clear column names and one record per row. Paste a table instead and you often can't tell from the reply whether it ran code on the numbers or worked them out in the text.
2. Ask one specific question at a time. "Analyze this" invites a generic answer. "What's the month-over-month change in average order value, and how many rows is that based on?" gets you something you can verify.
3. Ask it to show its method. A number with no calculation behind it is unverifiable. A number with the formula or code shown next to it is a five-second check.
4. Cross-check on a sample. Recalculate one or two of the reported figures from the raw file before trusting the rest of the summary.

What to never trust without checking

• Any total, average or percentage that will be used anywhere that matters. Recalculate at least a sample.
• The row count behind a figure. Confirm the number of rows it worked from matches your file, especially after a long paste.
• "Significant" or "strong trend" language. Ask what sample size the claim rests on before you accept the framing.
• Causal words in the answer. Treat any sentence with "causes," "drives" or "leads to" as a correlation claim. A spreadsheet alone almost never proves causation. Our post on correlation vs causation explains why.

Right about the wrong rows

Uploading fixes the arithmetic. It does not fix the question the arithmetic answers. When ChatGPT runs code, a sum is a real sum, but a sum of the wrong rows is still wrong. A filter applied without comment, blank cells read as zero, or a header line counted as data all produce a figure that was calculated correctly and is still not your answer.

Reading the code helps if you read Python. If you don't, the check that works on every route is plain: ask how many rows each figure used and what was excluded, then compare that with your file.

Prompts that produce checkable answers

A vague prompt gets a vague answer delivered with certainty. A specific one makes the model hand over what you need to check it. Try this: "For each finding, give the number, the rows it is based on, how you calculated it, and a HIGH, MEDIUM or LOW confidence level with a one-line reason." That one sentence turns a paragraph you have to take on faith into a list you can audit.

Where ChatGPT earns its keep

None of this argues against using it. Profiling a new dataset, drafting a first summary, picking the three columns worth a closer look out of forty: ChatGPT does all of that faster than you can by hand, and the draft is usually a good place to start. The discipline is keeping "draft" and "verified" as two different labels.

Pairs well with

The Prompt Practice Lab ($69) grades the prompts you write STRONG, WORKABLE or WEAK, and a prompt that never asks the model to check its own work cannot score STRONG. The AI Over-Reliance & Critical-Thinking Check ($69) seeds AI text with a fabricated statistic and other planted errors and scores how many you catch as INDEPENDENT, ASSISTED or RELIANT. If the data you want to analyze started life as a PDF, invoice or receipt, the AI Document Extraction Kit ($99) gets it into structured fields first and flags low-confidence ones for review.