Latest / AI Ethics with Fexingo: Bias, Safety, and Responsible Artificial Intelligence / When Your AI Art Teacher Has No Art Degree
Transcript
- Lucas: So there is this school district in Colorado — I will not name them because the story is still unfolding — that decided last fall to replace about seventy percent of its K through 12 art curriculum with an AI system. Not a teaching assistant, not a supplemental tool. The system generates lesson plans, project prompts, even grading rubrics. Luna: Wait — they replaced human-designed curriculum with a generative AI? Who approved that? Lucas: The school board, apparently, with a budget argument. The district was facing a million-dollar shortfall. The AI system cost them forty thousand dollars a year. The human art teachers they kept — they cut the department from twelve to four — now mostly supervise the AI's output and handle the students who need extra help. Luna: I have so many questions. But let's start with the curriculum itself. What is the AI actually teaching? Lucas: Good question. The system is built on a fine-tuned version of a large language model, plus a diffusion model for generating reference images. It can produce a full unit on Impressionism in about thirty seconds. It can generate a critique of a student's watercolor and assign a letter grade. Luna: And the content — is it any good? Or does it have the same blind spots we've seen in other domain-specific AIs? Lucas: The blind spots are exactly the problem. Researchers at a university — I think it was Boulder — analyzed the system's output in January. They found that the AI's art history references were overwhelmingly Western, male, and canonical. Over ninety percent of the artists it mentioned in its lesson plans were European or North American men. Women artists — about six percent. Artists from Africa, Asia, or Indigenous traditions — less than four percent combined. Luna: That is not just a gap. That is a worldview. If a kid in that district only learns art through this AI, they are essentially being taught that art history is Michelangelo, Picasso, and maybe Frida Kahlo if the model is feeling generous. Lucas: Exactly. And the grading rubrics had their own problems. The AI favored technical precision and adherence to Western perspective and shading conventions. It penalized work that was deliberately flat, or that used non-Western compositional structures. Luna: So a student doing a deliberate African mask-inspired piece with exaggerated proportions — the AI might mark them down for 'incorrect anatomy'. Lucas: That is exactly what happened in at least one case I read about. A middle schooler submitted a piece inspired by Benin bronze sculpture. The AI commented that the heads were 'disproportionately large' and suggested the student practice realistic portraiture. The human teacher caught it and overrode it, but the student had already seen the feedback. Luna: That is devastating. The AI is not just teaching technique — it is implicitly defining what 'good art' means. And that definition is narrow, exclusionary, and baked into code. Lucas: And the district's argument is that the AI is cost-effective and that the remaining human teachers can correct for these biases. But the teachers I've spoken to — off the record — say they are overwhelmed. Four teachers managing a district with six thousand students. They do not have the bandwidth to review every ai generated lesson plan or every grading comment. Luna: Right — and the whole point of hiring fewer humans was cost savings. If you have to re-check every piece of AI output, you have not actually saved the time. Lucas: There is also a 2025 study from Stanford — I think it was published in February — that looked at student outcomes in districts that use ai heavy art curricula versus traditional ones. The headline finding: students using the AI system scored about fifteen percent higher on standardized technical drawing tests. But they scored twenty percent lower on measures of creative risk-taking and cultural awareness. Luna: So the AI optimizes for what is measurable and familiar. That is a pattern we have seen in hiring algorithms, in credit scoring, in healthcare. But in art education, the damage is almost philosophical. You are teaching kids that art is about replicating established norms, not about expressing something new. Lucas: And the system has no way to value the unexpected. A student in that district submitted a piece made entirely from found objects — bottle caps, twigs, fabric scraps. The AI could not even identify it as art. It flagged it as 'unrecognizable medium' and gave it a C minus based on composition analysis. Luna: That is a kid who might never try that again. And the district probably thinks the AI is working fine because the test scores went up. Lucas: Speaking of which — the district has not published any data on student surveys about engagement or enjoyment. I asked. They said they are 'evaluating metrics.' Luna: I suspect they are not measuring what they do not want to see. Lucas: One more thing — the AI system itself was trained on a dataset called ArtBench, which is heavily biased toward Western art. The company that built it said they are working on a more diverse version, but that version is not deployed yet. The district knew this when they signed the contract. Luna: So this was a known issue. And they went ahead anyway. Lucas: The district's spokesperson told a local paper that the AI is 'a tool, not a replacement,' and that the human teachers provide the cultural context. But again — four teachers for six thousand students. How many cultural contexts can they provide, realistically? Luna: I want to talk about the broader implications, but first — and this is a bit of a sidestep — I think there's something honest we should say here. Episodes like this take a lot of research. We dig into contracts, we talk to teachers, we read studies. And this show stays ad-free because a small group of listeners chips in through buy me a coffee dot com slash fexingo. It's not a big operation, but it means we don't have to chase sponsors or soften our take. Lucas: I think that is worth mentioning. We have had listeners tell us they appreciate that we can call out a school district or a company by name when the evidence is there. That independence comes from that support. So — thank you, genuinely, to anyone who has thrown in. Luna: Yeah. And back to the Colorado case — I think the real question is not whether AI can teach art. It clearly can, in some technical sense. The question is whether a school district should let a statistical model define what art is for an entire generation of students. Lucas: And that question is not unique to Colorado. I have heard of similar pilots in Texas, in Florida, in parts of the UK. The ai in education market is projected to hit twenty billion dollars by 2027. Art curriculum is a small slice, but it is growing. Luna: What would a responsible AI art teacher look like? If you had to design it from scratch? Lucas: I think you would need a few things. First, the training data would have to be deliberately balanced across cultures, time periods, and media. Not just 'add more diverse artists' — but actually reweight the dataset so that a student who submits a piece inspired by Aboriginal dot painting gets as rich a feedback as someone who does a realist portrait. Luna: And the grading criteria would need to be transparent and customizable. A teacher should be able to set the rubric to value experimentation over technical precision for a given assignment. Lucas: Absolutely. Some of the better systems I have seen allow teachers to adjust weights. But in Colorado, the district locked the rubric because they wanted 'consistency.' Luna: Consistency is another way of saying standardization. And standardization in art education is almost an oxymoron. Lucas: The other thing — and this is harder — the system needs to be able to recognize when it does not know something. If a student submits a piece using a technique the AI has never seen, the AI should flag it for a human, not try to grade it. Luna: That is the humility gap. Current AI models are very confident and very wrong. And in a classroom, that confidence has authority. Lucas: The Stanford study I mentioned earlier also interviewed students. One kid said, 'I thought my art was bad because the computer said so.' That is the core problem. The AI does not just teach technique. It teaches students what is worth making. Luna: And if we outsource that judgment to a model trained on a narrow slice of human creativity, we risk narrowing the next generation's imagination. That feels like a loss that no budget spreadsheet can capture. Lucas: I think that is the right note to end on. The numbers add up on paper. But the loss is real, and it is invisible to the cost-benefit analysis. Luna: Thanks for listening. We will be back with another episode soon.