AI code: Dangerous at any speed

As an instructor of future game programmers, it is important for me to stay technologically current. I've been using chat gpt as my programming assistant all summer. I have come to a few conclusions:

1) It can be very handy for the well-known, boilerplate things and most documentation

2) It doesn't always generate the best code and sometimes it misses the mark, and you must go in and fix or make changes.

3) When it comes to things that there are very few examples for, or have been changing rapidly over time, it will not only give you bad code but sometimes a bad approach. If you ask it to fix the errors, sometimes that works, but it continues on the path it was traveling and does not consider other approaches. This can be a big waste of time.

In the end, chatgpt coding is just a more convenient form of cut and paste coding that makes semi-educated guesses when it doesn't find the code already written.

But AI coding has had another impact on my teaching. I teach beginning programmers the way I was taught— by making them write “toy” programs that illustrate a particular technique and/or canonical solution to a problem.

The problem is that there are a lot of examples of similar programs on the internet. Cheating has always been a bit of an issue. Because there are a lot of similar examples to teach AI, AI can get a student 90% of the way there.

The result is that I have had to tighten my previous generosity in partial credit. If it’s not fully working or working the way I specified in the lab assignment, you aren’t getting credit for it, or at least not much.

They are also likely to fail, because my final project always has details you WON’T find on the internet and the students who have been cheating hit the wall.

I am genuinely sorry to have to cut back on partial credit. I miss the days when I could be more generous and lead a student where they need to catch up, but right now that simply is not practical.