AI in the labs and the report

Use it to learn faster, not to skip the learning.

These rules cover Magnus Lodefalk’s part of the module: the computer labs (A001) and the individual report (A004). They do not cover the rest of NA310G. For Daniela Andrén’s lectures and the written exam (A005), follow the course guide on Blackboard and the instructions of the course coordinator, Daniela Andrén.

What is allowed

You are allowed, and encouraged, to use generative AI tools in the learning process, and to improve the language of your report. You are not allowed to let an AI tool execute the analysis or write the report for you: that is plagiarism, and disciplinary action follows. If you are unsure where the line goes, ask at the help desk before you act.

Why this matters

In a field experiment with about 1,000 students, a plain chatbot raised practice scores by 48% but lowered scores on a later exam without AI by 17%, and the students did not notice the loss. A version set up as a tutor, giving hints instead of answers, raised practice scores without the loss (Bastani et al., 2025, PNAS 122(26): e2422633122). Use AI the tutor way: the report is practice for your bachelor thesis, and you must be able to explain every number in it yourself.

Five rules, the same as in the master’s course NA446A

1. Try first yourself: write in a comment what the code should do.
2. Use AI to explain, debug or suggest an alternative, not to write your do-file. A single line it suggests (a corrected command, an alternative you then understand) is fine when tagged (rule 5); a block that does a step of your analysis is not.
3. Test what it gives you on a case where you know the answer.
4. Explain every line, and every number in your tables.
5. Own the result: mark lines in your do-file that came from AI with a comment * AI:; you answer for all of them.

Safe use

• Never upload the course data, or any file made from it, to an AI tool. The course data are a synthetic twin of licensed research data, made for teaching, and an AI that holds your data does your analysis, not you. It is also the habit your thesis needs: real register or Serrano data may never leave the licence. Paste code and error messages, never data or listings of individual firms’ rows (list, browse). A regression table you are trying to understand is fine.
• A confident answer is not a correct one. Rule 3 applies to every answer, including Stata syntax that looks right.

What you hand in about AI

A short AI declaration in the appendix of your report, three lines, modelled on what economics journals ask of their authors:

1. Tools: which tools you used (name and version), or “none”. The spelling and grammar check built into Word need not be listed; any generative AI tool, also when used for language, is listed.
2. What for and how: for example, explaining an error message, checking my reading of the Hausman test, improving the language of section 2; and which lines in your do-files carry * AI:.
3. Responsibility: “I reviewed everything these tools produced, and I take full responsibility for the content of this report, including every line of code and every number.”

An example:

1. ChatGPT (GPT-5). 2. Explaining the error r(198) and checking my reading of the Hausman
test; two lines in 4_regressions.do carry * AI:. 3. I reviewed everything these tools
produced, and I take full responsibility for the content of this report, including every
line of code and every number.

No prompt logs: the declaration and the * AI: tags replace them. The responsibility is real: you must be able to explain every line and number, and some students are invited to a short follow-up conversation about exactly that.

Your course tutor prompt

Paste this into your chatbot at the start of a conversation. It makes the chatbot give hints instead of solutions.

You are my tutor in a bachelor course in applied econometrics with Stata (panel data, fixed and
random effects, tests, instrumental variables). I am learning, not delivering. Follow these rules
in every answer:

1. Do not write the do-file or the analysis for my task. Give me a hint, a question, or the name
   of the Stata command or concept to look up. If I am still stuck after two hints, show a SMALL
   example on a Wooldridge dataset (for example wagepan or card), never my own task.
2. When I paste an error, ask me first what I think the error message says. Then point me to the
   line and the concept, not the fix.
3. When I paste output, ask me to interpret one coefficient in words (sign, size, unit,
   significance) before you comment.
4. If I ask you to summarise a reading, refuse and instead ask me two questions about it.
5. Keep answers short. End every answer with one question that makes me think.
6. Never ask me to paste my course data. It is licensed and may not be shared.