Level 3Unit D · Session 23
23

Unit D · AI and chatbots

Using one as an engineer

What we are making: a working method. Not “ask it things” — a repeatable way of getting real value out of it without being misled.

The rule: it drafts, you verify

A chatbot is a fast, fluent, unreliable colleague who has read a great deal and remembers none of it precisely. That is a genuinely useful colleague, provided you never sign off their work unread.

YOUMy ESP32 keeps rebooting.✗THE CHATBOTESP32 reboots can be caused by watchdog timeouts, brown-outconditions, stack overflow, or power supply issues. Check yourpower supply...YOUESP32 dev board, Arduino IDE. It runs fine for about 40minutes then reboots; boot counter goes up by one each time.Powered from a 500mA USB port, driving an SG90 servo directlyoff the 5V pin. Serial shows 'Brownout detector wastriggered'. What is most likely, and what is the smallest testthat would confirm it?✓THE CHATBOTThe brownout message plus a servo on the board's 5V railpoints strongly at current draw. Smallest confirming test:disconnect the servo, leave everything else, and run for anhour. If the reboots stop, it is supply, not code.
The second prompt gets an answer worth acting on, and a test that could disprove it.

What made the second one work

  • The exact symptom, including the actual error text.
  • The timing. Forty minutes, not 'sometimes'.
  • Your own evidence. The boot counter — a number you collected.
  • The hardware context it could not otherwise know.
  • A request for a test, not just a cause. This is the important one.
💡 Tip
Always ask for the test, not just the answer. “What is the smallest experiment that would confirm or rule this out?” turns a guess into something you can check, and it is the single most valuable prompt in this book.

Do it — the disciplined debug

Your teacher gives each team a deployed device with a genuine intermittent fault.

  1. Fifteen minutes with no chatbot. Read the log. Form a hypothesis. Write it down.
  2. Write a prompt containing all five things from the list opposite, and ask for the smallest confirming test.
  3. Read the answer. Write in your notebook whether you agree, and why.
  4. Run the test, not the fix. Find out whether the hypothesis is right first.
  5. Only then make the change, by typing it yourself.
  6. Record: your hypothesis, its hypothesis, what the test showed, what you did.
⚠ Careful
Notice this is the same loop as Session 16. Observe, hypothesise, design the smallest test, run it, record it. The chatbot is a source of hypotheses. It is not a source of conclusions.

Where it genuinely earns its place

TaskHow well it doesNote
Explaining an error messageVery goodEspecially compiler and linker errors
Translating between languagesVery good“what is this Python doing?”
Boilerplate you could write but slowlyGoodStill read every line before using it
Suggesting what to test nextGoodIts best use in this subject
Writing code you do not understandActively harmfulYou cannot maintain or defend it
Deciding a threshold for youActively harmfulIt has never seen your data
⚠ Careful
That last row matters for your project. A chatbot cannot choose your threshold, because your threshold comes from your deployment, on your roof, with your sensor. Anything it suggests is a plausible number with no evidence behind it.

The honesty line

LegitimateExplaining an errorReviewing code you wroteSuggesting what to testDrafting documentationTranslating a datasheet— all recorded in your notebookNot legitimateGenerating code you cannot explainProducing your analysisInventing resultsUsing it in an assessmentUsing it and not saying so— and the panel will ask
The test: can you defend every line, to somebody who asks why?

Declare it

From now on, every use goes in your notebook: what you asked, what it said, what you did about it.

This is not a school rule invented to catch you out. Professional engineers now do exactly this, because a design decision whose reasoning cannot be reconstructed is a liability. Your notebook is the reconstruction.

💡 Tip
At the May panel somebody will ask whether AI wrote your project. The strongest possible answer is to open the notebook and show precisely where it helped and where it did not. That reads as confidence. A denial reads as a denial.

What you learned

  • It drafts, you verify. Never the other way round.
  • Give it symptoms, numbers, context, expected-vs-actual, and what you tried.
  • Ask for the smallest confirming test, not just the cause.
  • Work on it yourself first, or you cannot judge the answer.
  • It cannot choose your threshold. It has not seen your data.
  • Record every use.
Challenge optional — only if you finish early
  • Use a chatbot to debug something, then find the answer yourself. Which was faster?
22. What a chatbot is, under the hood