BMCC/CUNY GenAI Education Paper Cited in MDPI Smart Cities Article

Vasiliy,
I did an informal poll of my students asking what they wanted their program of study to provide them regarding AI, and the ability to critically evaluate AI output was one of the top things that they reported wanting. This fining could probably be replicated more formally.

My main focus in my courses is on the idea that if students want to be able to critically evaluate what AI produces, they need to learn to complete the task without AI first. This way, they know what is involved in correctly completing the task and what to look for in terms of possible problems with what AI produces.

However, although I have since lost the link, my thinking was significantly impacted by a YouTuber who argued that we should not be using AI as an autonomous assistant who does things for us on its own. Instead, it is more effective to use AI as an automated interlocutor who can help us think things through in dialog with us. On this view, the main strength of AI is that, unlike a human interlocutor, it never gets tired, looses interest, or runs out of time to help. This shifts the object of assessment from an end product produced by AI to the quality of feedback and suggestions offered by AI.

Keith

Vasiliy,
I don’t think that we should try to reach consensus in this forum but I have a rather different view. Perhaps it would he helpful to explain a little further. We were speaking loosely, so to be clear when I referred to AI, I meant generative AI. Perhaps you were speaking more broadly.

As philosopher David Lewis once wrote, calling someone “Armstrong” does not mean that they have strong arms, and likewise calling generative AI “intelligence” does not mean that it is intelligent. Because generative AI simulates intelligence by brute force prediction, I do not consider it intelligent. There is room for disagreement in part because we can have different criteria for intelligence.

I should think that anyone on this list could complete basic arithmetic done by a calculator, just not as quickly. However, trust in a calculator is not just a purely empirical matter, it is also a matter of design. In a sense, a calculator has a model of how numbers behave built into it. However, generative AI is only a model of language, it lacks of model of how the world works. Even in cases where machine learning is applied to things other than language, prediction is not the same as explanation or understanding.

When I suggest that knowing how to complete a task without generative AI is important to being able to critically evaluate the products of generative AI, this in no way implies that doing the task oneself is the same thing a critically evaluating the product of generative AI. However, if one does not know how to complete the task, it creates a double black box: first the usual generative AI black box that we do not know how the model produced the result and second the black box of not knowing how a person would complete the task without AI.

This issue is only amplified when we move from simple right-or-wrong tasks to complex tasks with no uniquely correct answer such that products need to be judged by multiple criteria and alternative solutions can be equally satisfactory. The skills that I am teaching students tend to be of this latter type where there is no definitive test of a correct answer that can be applied without knowing how to complete the task. This is quite different, for example, from vibe coding by writing comprehensive unit tests for each function that can be evaluated without knowing how the code works.

I am not looking to rain on anyone’s parade. I would say, for the time being, let many flowers bloom and with time we will sort out the best approaches to using generative AI. I hope that I have been able to offer you some insight into an alternative perspective.

Regards,
Keith