Friends,

Since the beginning of this whole journey with AI, two thoughts have jockeyed for supremacy in my brain. The first is that questions over AI in general, and classroom use in particular, feels more like an ethical question than a technical one. I want to write more about this later—how we let ourselves be persuaded that education was broken—but for now I’ll just say this:

We should be deeply suspicious of technologies that absorb power, autonomy, and decision making, and technologies that internalize rewards and externalize risks. We have a social, political, ethical obligation to block this absorbtion by diffusing the power, redistributing rewards, minimizing risks. To this end, I have been deeply moved by the over 600 artists that have stood in solidarity with our call for a moratorium: “We’re calling on you to stand up for the children of this city—the natural-born creatives, the innate innovators, tomorrow’s leaders—whose future is being sacrificed only to further enrich the planet’s wealthiest people.” Rejecting AI feels like a profoundly optimistic, and ethical, choice—stand with the poets and the painters.

The second thought is that bringing such a manifestly disruptive technology into classrooms ought to have an overwhelming research base to support it—the proof, not just potential—of an impact on learning so profound that we as a society believe it’s worth assuming the risks.

And that research base manifestly does not exist. The chancellor, in a recent meeting with AIM, admitted as much - that there are no studies supporting the use of AI in classrooms that are not funded by industry. The city, it seems, is basing its policy on vibes. But you don’t have to take my word for it—we have an expert guest author today.

City Council held hearings about AI in NYC public schools in late June, after a majority of council members signed a letter supporting our moratorium. Notably, the council grilled the DOE for over three hours before parents and community members were invited to testify. The DOE delegation chose to leave rather than hear from the community. You can read my testimony here.

Dr. Krystal Cleven, a member of the AIM Coalition, a mother of two, community organizer with District 30 Parents for Ed or Tech, and a physician with a masters degree in clinical research methods, had this to say about the “research” that the DOE cited during the meeting.

AI Products v. the Gold Standard: Human Teaching

Krystal Cleven, MD

When I toured public schools for our young children, I asked all the usual questions. How big are the classes? How long is recess? I did not ask, “will my four-year-old befriend an AI chatbot?” Frankly, that seemed outrageous…until a parent friend shared that her kindergartener was using AI in her classroom. I soon learned there are hundreds of schools using AI products like Amira, Writeable, Google Lens, in K-12 classrooms without parental consent or knowledge—including some products like Google Gemini that are deemed “high risk” by Common Sense Media. Parents reasonably assume our NYC public school system is providing children with safe and effective learning tools, but during a recent city council oversight hearing on the use of AI in our schools, system officials provided no quality evidence of the effectiveness of these AI products and demonstrated they are providing little protection against its harms.

Early in the hearing, Eric Dinowitz, Chair of the Committee on Education, asked NYCPS officials what empirical evidence they have that AI tools are for the betterment of students’ education and cognitive development. Among all of our concerns regarding data privacy, algorithmic bias, cost, and environmental impacts, this seemed like a very basic and fundamental question. Do these AI products improve learning? As a physician with a master’s degree in research methods, I was curious about the evidence that the district cites to defend its position that this technology belongs in classrooms.

Dr. Alan Cheng, the Supervising Superintendent for High Schools, claimed there are “multiple” meta-analyses and randomized control trials in peer-reviewed journals (considered the highest quality studies we have in research), that show AI can offer careful support around instructional design, feedback, and adaptability.

One of the studies he cited in the hearing reviewed 28 studies that examine Intelligent Tutoring Systems (ITSs). ITSs are AI programs like Duolingo that adapt to a student’s progress. This was not a meta-analysis, as Dr. Cheng claimed, but instead a systematic review which does not involve any real analyses, so is not the quality of evidence he suggested. Additionally, not all ITS studies reviewed had positive results, only one included children younger than fifth grade, and none were younger than third grade. Most were in high school, and the studies did not evaluate any long-term exposure or outcomes. Most AI interventions lasted less than one week and some only one day. 

When examining Dr. Cheng’s claim that there are “multiple” randomized control trials, I found no randomized trials of K-12 students in the United States. Indeed, there are extremely limited studies in this age group. Most are not peer-reviewed, are funded by big tech, include only adult learners, examine teacher AI use rather than student, and were evaluated in very resource limited international settings with different needs from ours. 

Even the poor-quality data available are not robustly positive, and some demonstrate harm. In fact, the only other data cited by the NYCPS was a study that highlighted the negative effects of using ChatGPT-like products on learning. The study showed an increase in correct responses at the time these AI products were used compared to non-AI resources, but when re-assessed, the students who used AI showed no improvement or performed worse than those who did not use it. Perhaps most alarmingly, the students who used AI did not perceive their reduction in sustained learning. 

Dr. Cheng included an important qualifier in his remarks: that AI can assist learning “when these tools are used with educators.” However, when a child’s device includes Amira, Writeable, or Google Gemini, they are not always used with the direct guidance of a teacher. I fear that is part of the appeal: to create products where educators are not needed.

A major flaw of many educational technology and AI studies is that they lack a real control group at all, sometimes only comparing their newer product with an older version, making it impossible to conclude their products improve learning. This is akin to Cheerios claiming their product improves heart health. Compared to what? Fruit Loops? Or a more optimal breakfast like steel cut oats? 

AI companies never compare their products to the gold standard: an optimally supported teacher-led classroom. When educational tech companies compare their products to a non-teacher directed intervention, they are clearly looking to create a product that will not require a teacher in the end. When we make budget decisions as the largest school system, our decisions have ripple effects across the country. When we spend nearly $2 billion on technology annually, millions of it on AI products that do not require human teachers, we are sending a dangerous message about our values and mission. 

In the end, NYCPS officials provided no empirical evidence to justify AI use. They are making decisions for 800,000 children based on an evidence base that does not exist. Unless they take swift action to ban AI, schools will be left alone to interact with predatory educational technology companies pushing their ineffective and harmful AI products into our classrooms.