Resources for Critical AI Pedagogy

An image of noise generated as the first step of creating an AI image in a diffusion model.

Eryk Salvaggio (eryk.salvaggio @ gmail.com) LinkedIn

  • Bountiful resources are available at the AI Pedagogy Project website from the metaLAB (at) Harvard; including lesson plans and frameworks for developing critical AI pedagogy.

Why Critical AI Pedagogy?

  • One of the core problems of AI in classrooms comes from an abundance of trust in its false authority. When students encounter any source without informed and calibrated skepticism, it becomes a problem: see Wikipedia.

  • The more informed students and faculty are of AI, the more critical they become. The more critical of AI they become, the more useful AI is.

  • This is because AI is labored with myths that center AI itself as an agent useful to learning. The best use of AI is not as a source of information, but as a source of friction that challenges our understanding in order to confirm and verify what we are learning — to arrive at confident cognition.

Impacts on Creativity

  • A 2024 study found that “using an AI image generator during ideation leads to higher fixation on an initial example. Participants who used AI produced fewer ideas, with less variety and lower originality compared to a baseline,” “narrow[ing] designers’ ability to explore the creative space between abstract ideas and potential solutions.”

  • Fixation on the output reflected a fixation on the assignment / brief: the task was framed as a problem to solve, rather than a space to explore. In design, “more finished (polished)” images tended to restrict exploration and brainstorming, while sketchy, unfinished images tended to encourage it.

  • Counteraction: Framing assignments as exploratory, setting time aside in class for structured brainstorming, and encouraging device-free ideation before turning to AI can be beneficial.

Illusions of Understanding

  • AI tools perpetuate an illusion of understanding: “The proliferation of AI tools in science risks introducing a phase of scientific enquiry in which we produce more but understand less.”

  • In a 2024 study of 572 students who used ChatGPT to study for an exam scored 6.7 points less than those who did not — with high-achieving students showing the greatest drop. (Note that this was assessed with ZeroGPT used to determine if the essays were AI written).

  • Counteraction: Again, we come up with the challenge of AI being the myth of its authority. A 2025 study showed that people with less understanding of how AI works tended to trust it more. Therefore, a key way to diffuse the harmful effects of AI in education is to teach it critically.

Undermining Confident Self-Assessment

  • Deferral to AI as an authoritative source creates conflict between the LLM and the understanding of the student. That is, if theyuse ChatGPT to verify and confirm knowledge, they may confirm inaccuracies about that knowledge — especially for complex, nuanced topics.

  • A 2022 study showed that “human self-confidence, not their confidence in AI, directs the decision to accept or reject AI suggestions. Furthermore, this work finds that humans often misattribute blame to themselves and enter a vicious cycle of relying on a poorly performing AI.”

  • Counteraction: Embrace an understanding of metacognitive processes (evaluation, reassessment, and problem frames) as a means of challenging ChatGPT's feedback. Intrinsic motivation is key; how do you model a process for students to follow that is not focused solely on outcome (ie, “completing the assignment correctly”).

AI Plagiarism Detection

  • Tools used to determine whether a text has been generated by AI are generally not recommended

  • They have been shown to be inaccurate and biased against non-native English speakers and marginalized students.

  • A 2023 systemic review of 17 studies showed each arrived at various, contradictory conclusions about the same tools regarding their reliability.

My Research into Generative AI