Doing the Analysis

Getting the whole dataset in front of the model, on a tool that actually runs code against it, is not the finish line — it is where the real risk starts. A vague prompt gets a vague summary; a well-shaped question gets a real answer you could defend to someone who pushes back on it. And even a well-shaped question, answered by a tool running real code against your whole file, can still be wrong: AI can invent a data point that was never in the source, get the arithmetic wrong without flagging any uncertainty, or ride a clean, confident narrative past what the data actually supports. This section covers both halves together, on purpose — asking well and checking the answer are not two separate skills, they are one habit.

Asking Good Analytical Questions · From Raw Numbers to a Defensible Conclusion · Where AI Fabricates Data and Numbers · Silent Arithmetic and Logic Errors

In this section

Page	Last updated
Asking Good Analytical Questions How the shape of the question you ask determines whether AI does real analysis or produces a vague summary.	Updated 2026-07-06
From Raw Numbers to a Defensible Conclusion Going beyond a surface-level summary to a conclusion you could actually explain and stand behind, including the checklist question that catches an overstated trend or a mistaken correlation.	Updated 2026-07-06
Silent Arithmetic and Logic Errors Why AI can get the math wrong on your own data even when it ran real code against the whole file, without flagging any uncertainty, and how to catch it.	Updated 2026-07-06
Where AI Fabricates Data and Numbers Documented cases of AI inventing data points and figures during data analysis specifically, not general AI hallucination, with evidence matched to the mode -- chat or code-executed -- it was actually measured in.	Updated 2026-07-06