David Malan
๐ค SpeakerAppearances Over Time
Podcast Appearances
Students are just going to end up typing into the prompt.
Whereas I think it's a lot cleaner, if not simpler to know, I cannot, through policy,
go to ChatGPT, Gemini, Copilot, any of these things, but I can go to and use as much as I want CS50.AI or the plugin we have in VS Code of the same.
And that to me is a very healthy line because you know if you're crossing that line, if you're pulling up ChatGPT.com or Gemini.Google.com and the like, and at that point it's a conscious choice to be academically dishonest as we would describe it in CS50's syllabus.
Statistically, we are not catching more.
I would like to think that the misbehavior has not markedly increased, if only because we, like a lot of intro courses around the world...
have a tradition of looking for academic dishonesty plagiarism code that was copied and pasted off the internet or YouTube video.
And we within CS50, like a lot of peer institutions are very good at catching that.
And like we have tools that cross compare all of students submissions against each other, against GitHub repositories of past submissions that we have of YouTube videos that have been described.
So we have historically administratively disciplined, so to speak, between 5% and 10% of CS50 student body every semester.
And that's kind of a norm across peer institutions, upper bound of roughly 10%.
There's certainly probably some percentage of students who have been cheating in some form all these years and never have been detected.
But just based on the rigor with which we go through this, the messaging throughout the course, I'd like to think that a good 90% of students are behaving in the way that they should, at least when the boundaries are set through not only policy, but course culture and through the support structure that we provide to students.
That said, we've not seen an uptick in what in detections of academic dishonesty.
What has gotten harder is the so-called prosecution of those cases in the sense that it's harder for us now to hand to the administrative board of Harvard or the Honor Council, so to speak, a smoking gun.
Like here is the URL from which this code was copied because it's not really coming from a URL or YouTube video.
It's coming from the composition of all of the URLs out there and all of the YouTube videos.
about CS50's problem sets because the AI somewhat pseudo-randomly is spitting out an amalgam of these various training inputs.
That said, when you've been doing this long enough, you can tell that this is not the work of this particular student, either in comparison to past work that they submitted or the sophistication with which they're solving a problem or the incorrectness with which they're solving the problem.
And there's often telltale signs like,