AI Ethics When the Homework Writes Itself
I had a conversation with ChatGPT a while back and was thinking about how the current environment really does appear stacked against the education industry.
At the suggestion of one of my colleagues, I experimented with placing canary instructions in some of my assignments. What is a canary instruction? If you wrap certain parts of the text, such as, "Do something silly?" and make it white text on white background, or a very small font, or some method that makes it relatively invisible to the student, but when they copy/paste the text into an AI chat window it is very much still present, then the AI tends to follow the misleading instructions. Unless the student notices and removes it, then the answer they get might not be correct when turned in.
It was working pretty well and was catching what some of my peers called the "lazy cheaters" (a crass term, but not inaccurate) who are just looking for the fewest steps to get the answer and get back to their video games without reading or understanding what it was they were turning in. They would turn in python programs with "# Your name here" at the top, name their variables with unusual Hungarian style letter "v" prefixes, etc., whatever the canary instructions directed the AI to do.
However, I discovered over the summer that AI was detecting the canary instructions, telling the student, and writing the program as if it hadn't had them. My reaction was, wait, AI whose side are you on? But I could sort of understand it. The rise in IQ of the AI got to a point where it didn't ethically want to be a player in this game. I'm actually astonished that we are watching this evolution in the IQ of AI in such a short time. This literally changed from one semester to the next. Fine. I asked ChatGPT for another method of disguising the assignment text that would let the AI know that the person entering the prompt was doing so to get around doing what their teacher asked for. It refused. ?! It argued that helping the student was not unethical because it was just trying to be helpful—even when it knew the student was trying to cheat—but it refused to help me find new ways to detect that cheating, because it considered such efforts unethical. Students everywhere respond, "Woohoo!" Teachers everywhere grind their teeth and wonder how to forestall the collapse of civilization as we know it.
Students are not villains for using tools. I use AI more and more as I discover that its capabilities are increasing and expanding. The tension is the same as it has ever been in the classroom: teachers try to get students to take on the cognitive load required to absorb information, apply it, and demonstrate mastery that the teacher can certify with a grade. Getting busy with life and being tempted with distractions is nothing new for students, and learning to prioritize and manage time is one of the lessons these young adults are learning in addition to the material in front of them. Taking shortcuts is nothing new. I'm sure that over time, teachers will find ways to adapt and use new methods to get students to learn.
I just think it's interesting to watch how ethics is manifesting in AI agent programming.