Start With the Question, Not the Tool
The fastest way to wreck an analysis is to load your data and start poking. AI makes poking frictionless, which is exactly the danger. Before you run anything, get the plan straight.
Hand the AI your research question, your variables, and what you actually measured. Ask it to propose the analysis — not to run it.
Here's my study: I surveyed 120 students on hours slept (continuous)
and exam score (0-100). My question: does sleep predict performance?
What analysis fits this data, what are its assumptions, and what
would violate them? Don't run anything yet — just walk me through
the options and tradeoffs.
You'll usually get a sensible answer (here, a correlation or simple regression) plus the assumptions you'd otherwise forget — linearity, outliers, independence. That list is the point. It turns "I ran a thing" into "I ran the right thing and I know why."
Push back when the suggestion feels too fancy. AI loves to recommend a mixed-effects model when a t-test would do. If you can't explain the method to a classmate in two sentences, it's probably wrong for you. Ask: "What's the simplest analysis that still answers my question honestly?"
Run It — and Make AI Show Its Work
You don't need to be a coder to run a basic analysis. You do need to read what runs. Ask the AI for code with comments on every step, in a tool you can actually open — usually Python in a free Google Colab notebook or R.
Write Python for Google Colab that loads sleep_data.csv, checks for
missing values, plots sleep vs. exam score, and runs a linear
regression. Comment each line so I understand what it does. Tell me
which numbers in the output matter and what they mean.
Run it. When it breaks — and it will, usually a column name typo or a missing package — paste the full error back. AI is genuinely excellent at debugging because the error message tells it exactly what went wrong.
The non-negotiable habit: after every result, ask the AI to interpret it and tell you what could be misleading.
My regression gives a coefficient of 4.2 (p = 0.03), R-squared 0.08.
Explain each number in plain English. What does R-squared of 0.08
mean for how seriously I should take this? What am I NOT allowed
to conclude?
That last question is your guardrail. A p-value under 0.05 doesn't mean sleep causes better scores — it's a survey, not an experiment. AI will say so if you ask. It rarely volunteers it. If you want to build real fluency here rather than leaning on the AI forever, the Python for AI & Data Science course covers this groundwork, and Intro to Machine Learning (No Code) is the gentler on-ramp.
Get Unstuck Without Faking Competence
Stats anxiety is real, and AI is the best tutor you've never had to book office hours with. Use it that way — as a tutor, not a substitute.
When a term ambushes you in a paper, ask for the version you can use:
Explain what a "confidence interval" is to someone who's never taken
statistics. Use my actual result: 95% CI [1.1, 7.3] for the sleep
coefficient. What does this interval tell me, and what would a
narrower or wider one mean?
Anchoring the explanation to your own numbers is what makes it stick. Generic definitions evaporate; "here's what your 95% CI means" does not.
Two failure modes to watch:
- Confident wrongness. AI states statistical claims with the same certainty whether it's right or hallucinating. For anything you'll put in a paper — which test, which assumption, how to report it — verify against your stats textbook, your course notes, or a source like the UCLA statistics guides. One cross-check costs five minutes.
- The understanding shortcut. If your professor asked "why this test?" and your honest answer is "the AI told me to," you don't understand your own analysis. Keep asking "why" until you could defend it with the screen off.
Coding Qualitative Data Without Outsourcing Your Judgment
Interviews, open-ended survey answers, field notes — AI can accelerate qualitative coding, but this is where faking results gets seductive and obvious. The work is interpretation, and interpretation is yours.
Start by developing your codebook on a small sample yourself. Read ten responses, name the themes you see, define them. Then bring in the AI to apply your scheme at scale.
Here's my codebook: [paste your themes + definitions]. Apply these
codes to each of the following 40 survey responses. For every
response, give the code AND the exact quote that justifies it. If a
response fits no code, say "uncoded" — don't force a fit.
Demanding the justifying quote does two things: it lets you spot-check fast, and it stops the AI from inventing themes that flatter your hypothesis. Always audit a random slice by hand. If your codes and the AI's disagree on 30% of responses, your codebook is too vague — fix it, don't paper over it.
Never let AI invent quotes, merge participants, or "summarize" raw data into findings you didn't verify. A fabricated interview quote is the qualitative version of faking a number, and it ends careers. The AI for Academic Research course goes deeper on keeping this defensible.
The One Rule That Keeps You Honest
Run everything through a single test: could you reproduce and explain this result without the AI in the room? Not write the code from scratch — explain what it did, why that method, what the numbers mean, and what they don't prove.
If yes, AI was a tool and you stay in control. If no, you've outsourced the part of research that's actually yours, and a single follow-up question in a viva or a peer review will expose it. Use AI to move faster through the mechanical parts and to tutor you through the confusing ones. Keep the understanding, and the judgment, on your side of the screen.