AI Literacy for Scientific Programming

AI chatbots produce plausible code fast, but cannot know whether that code is right for your data, your units, or your question. This page is a living guide — expect it to change faster than the rest of the book as models and norms evolve — built around three habits: being a precise communicator, a skeptical reviewer, and an active learner.

Be a precise communicator¶

A prompt that gets a useful answer on the first try usually states four things:

The goal — the high-level scientific objective, not just the mechanical step.
The environment — language and libraries (python, pandas, numpy, ...).
The data — its actual structure; paste a .head() or print() output rather than describing it from memory.
The task — the precise action wanted.

Compare two prompts for the same task: averaging a temperature record by month.

A bad prompt:

How do I average my data by month in python?

Without a description of the data, the assistant reaches for a generic example — a three-row toy table that happens to already have monthly, not daily, dates — and returns code that runs but has nothing to do with the actual station data.

A good prompt:

I am using python with the pandas library to analyze weather station data. I have a DataFrame df_temp with a DatetimeIndex and a column air_temp_celsius. Here is df_temp.head():
                      air_temp_celsius
2024-07-01 00:00:00              15.2
2024-07-01 01:00:00              14.9
Please give me code to compute the mean monthly air temperature, stored in a new DataFrame monthly_mean_temp.

Pinning the actual column names and structure gets df_temp.resample("ME").mean() — the right tool, applied to the right data — on the first try.

Be a skeptical reviewer¶

Generated code runs and looks reasonable far more often than it is actually correct — the failures that matter in science are silent, not crashes. Three checks catch most of them:

Understand. Can you explain what every line does? If not, ask for a line-by-line explanation before running it.
Test. Run it on a small, known subset first, and check one value by hand.
Question. Is this the best approach, or just an approach? Ask for alternatives and trade-offs.

A first request — “plot a 30-day rolling average of temperature” — gets a working but bare plot. Two follow-ups sharpen it without starting over. First:

This code works, but the plot is not very readable. How would you modify it so it’s more readable?

adds axis labels, a legend, and the raw data as context. Then:

For climatological analysis, a centered mean is usually more appropriate. How would you modify this to use a centered 30-day window?

is the kind of correction only a reviewer who understands the domain, not just the syntax, would think to ask for — rolling(window=30, center=True) in place of the assistant’s default trailing window.

Be an active learner¶

Three prompt patterns turn the assistant into a tutor instead of a code vending machine:

Explain this error — paste the full traceback and ask what it means and why, not just for a fix.
Compare these methods — “what’s the difference between a for loop and a while loop, and when would I use one over the other?” turns a syntax question into an understanding one.
Suggest a structure — before writing a script that loads 50 files and merges them, ask for a clean approach first; you will read and adapt it faster than debugging one written top-down.

A concrete failure: the unstated unit¶

Asked to flag freezing conditions, an assistant might write:

def is_freezing(temperature):     # unit unspecified: the latent bug
    return temperature < 0

Called on a station that reports temperature in kelvin, this is wrong for every real value: is_freezing(268.15) — a genuine −5 °C — returns False, because nothing is ever below zero kelvin. The code is not wrong in isolation; it is wrong because the prompt never fixed the unit, so the assistant guessed celsius. The fix states the unit everywhere it can — the name, the type hint, the threshold:

def is_freezing_kelvin(temp_kelvin: float) -> bool:
    return temp_kelvin < 273.15

This page will change¶

Models, tools, and institutional norms around AI assistance are all moving faster than the rest of this book. Treat the three habits above — precise, skeptical, active — as the stable part; the specific tools and examples will be revisited as they change.