Lecture XI - Project Work I
Programming with Python
EP 11: The Build
The shop has a new owner
MunchCorp runs the shop now. Tobi kept his job, the investor kept her spreadsheets, and her next due-diligence visit already has a date: Session XIII, when your projects run in this room.
Until then, the shop rule stands: pull before you start, push when you stop, and let the commit history do the talking.
The Mark
Could anyone tell?
- Your README carries an AI-disclosure section. Suppose you skipped it: could anyone prove that an AI wrote part of your work?
- Since August 2026, EU rules require providers to mark AI content machine-readably, and the major ones do
- The trick behind it is simple enough for three slides
How a watermark hides in plain text
- Before each word, the model splits its vocabulary into a green and a red half, reshuffled at every step
- It then softly prefers green words. The text still reads completely normally
- By chance, a human writes about half green words. A model writes far more
You already know the mechanism
- Session VI: the model picks a likely next token, never the next token
- A watermark is that choice, steered on purpose: green wins a little more often than chance
- Same next-token choice, one hidden thumb on the scale
Who can actually detect it
- Detection is counting: one paragraph of marked text can push chance below 1 in 10 trillion
- But you need the secret key that decided which half was green
- In practice, only the provider can run the test
What a mark proves (and does not)
- Proves: a model processed this text
- Does not prove: who did the work, or whether a human verified it
- Heavy editing can wash a mark out, and short snippets may not carry one, so an absence proves nothing
Presence without disclosure, though, looks exactly like what it is.
Disclosure is part of the work
- The rules converged everywhere: AI cannot be an author, and its use gets disclosed
- Your README section says which tool, for what, and what you verified yourself
- That is standard practice, not a confession
Wrap-up
Three things to remember
- AI text carries a watermark now: fluent, normal-looking, detectable by the provider
- Disclosure is part of the work: which tool, what for, what you verified
- Small commits, pushed often, are the cheapest insurance any of you can buy
Next week: your agent keeps making the same mistakes. Time to teach it once instead of correcting it every session.