This page collects some notes on AI in the classroom.
My Thoughts on AI in the Classroom
I believe that a decision (i.e. a policy about AI in the classroom) is best evaluated with respect to a set of values (i.e. what a classroom is supposed to accomplish).
Going to university serves several purposes. Relative importance varies by student. I believe these purposes include, but are not necessarily limited to:
- Obtaining general skills and knowledge.
- Obtaining specialized skills & knowledge for “practical” reasons. Quick-to-materialize, relatively tangible, relatively predictable benefits of being more educated. Think “trade school for knowledg workers”.
- Obtaining specialized skills & knowledge for “intellectual” or other less-tangible reasons. Slow-to-materialize, less-tangible, less-predictable benefits of being more educated, likely in a particular area.
- Obtaining credentials. Advertising to society that you were educated at a particular institution.
This list is in no particular order, and is likely incomplete.
Assumptions:
- The potential uses of AI are varied. This is true in the classroom, and elsewhere.
- AI tools can, in general, lead to improved productivity. I am most confident about this claim for programming. Most of the students I teach are programmers. I am less confident about this claim for other disciplines.
- Employers, in general, will expect new graduates to be able to use AI tools. Again, I am very confident about this for programming, less so for other disciplines.
- The use of AI tools, in some cases, allows students to shortcut traditional assignments. The latest ChatGPT or Claude models are quite capable of producing solutions to introductory programming problems. Shortcutting assignments could lead to a degraded learning experience.
- Students at the undergraduate level feel pressure to achieve and to obtain credentials.
- Feedback loops on the credential value of education are often shorter than feedback loops on the intangible value of education.
- The use of AI tools, in some cases, can enhance student learning in unique ways. For example, ChatGPT provides good enough subject matter knowledge to answer introductory questions on many topics. While this experience is different and in some ways worse than the experience working with a human tutor, it’s relatively low-cost, and available 24/7.
- The detection of AI-generated content can be viewed as an adversarial game, played between attackers who attempt to substitute AI-generated content for human content without being noticed, and defenders who attempt to accurately discriminate between AI-generated and human-generated content.
- Many adversarial games display a cat-and-mouse dynamic. Vulnerabilities in defense are found and exploited by attackers. Defense responds by patching these vulnerabilities. Attackers respond by finding and exploiting new vulnerabilities. Complexity and effectiveness of attacks and defenses increase over time. Neither attackers nor defenders maintain 100% success rates in the long term.
- From 9 and 10, I expect that detecting AI-generated content in student submissions will display a cat-and-mouse dynamic.
- Maintaining standards of academic integrity serves all the values of education listed above. By “academic integrity”, I mean that there is an appropriate level of mutual understanding between the institution and the student about where the content of their work originated.
- Because AI can increase productivity and students are feel pressure to achieve, students will feel pressure to use AI to complete their work. Using AI for schoolwork can be, but is not always, in conflict with the values of education listed above.
Goals
- Provide opportunities for students to use AI in ways that are compatible with all of the values of education.
- Design teaching material that prevents or discourages the use of AI in ways that are detrimental across all of the values of education.
- In some cases, the impact of AI use on the values of education will not be unidirectional. Strike a balance across these cases.