I’ve been saying for years that schools need to adopt AI, and I haven’t changed my mind. But I’ve also been saying that adoption without infrastructure is a gamble. The Center for Democracy & Technology’s latest report, “Hand in Hand” (Laird, Dwyer, & Quay-de la Vallee, 2025), puts hard numbers behind that gamble, and the numbers are uncomfortable.
CDT surveyed over 2,800 teachers, students, and parents during the 2024-25 school year, and the central finding cuts through every section of the 64-page document: the more a school uses AI, the more risks its community reports. These aren’t speculative projections. They’re reported experiences from people living through the fastest educational technology rollout in memory.
The Dose-Response Pattern Behind AI Risks in K-12 Schools
CDT’s data reveals a consistent pattern across every risk category the report tracks. Laird et al. find that teachers who use AI for many school-related reasons are significantly more likely to report that their school experienced a data breach (28%, compared to 18% among teachers with low AI use). Those same teachers report higher rates of AI systems failing to work as described (23% vs. 4%), AI not treating students fairly (19% vs. 4%), and AI use damaging the school’s trust with the community (17% vs. 5%).
The student data follows the same curve. In high-AI-use schools, 61% of students reported hearing about deepfakes depicting someone connected to their school, compared to 16% in low-use schools. For deepfake non-consensual intimate imagery (NCII) specifically, the gap is 21% vs. 6%. And 45% of teachers in high-use schools heard about deepfakes, compared to just 13% among low-use teachers.
I want to be careful about causation here. CDT’s data is correlational, and schools that adopt AI extensively may also be larger, more digitally integrated, or have more surface area for problems to emerge. But the pattern runs through too many categories to wave away as coincidence. At minimum, it tells us that rapid adoption without adequate risk infrastructure produces exactly the outcomes critics have warned about. Schools that treated AI adoption and risk management as separate projects are now finding out they were always the same project.
What Students Are Actually Doing with AI Chatbots
The section of this report that should command the most attention is about student-chatbot interactions. Laird et al. report that 42% of students say they or a friend used AI for mental health support. The same proportion used chatbots as a companion or as a way to escape from real life, and 19% report using AI for a romantic relationship.
Those numbers are striking on their own. They get more troubling when you add the delivery mechanism: 31% of students who use chatbots for personal reasons say they’re doing so on school-provided devices or software. Schools are, in many cases, providing the infrastructure for these interactions without realizing it. That’s a governance failure, not a technology problem.
It also connects to what Luo (2025) found about generative AI eroding trust in teacher-student relationships. CDT’s data confirms that concern with scale: 50% of all students say AI in class makes them feel less connected to their teacher, rising to 56% in high-AI-use schools. And 38% of students say they’d prefer working with AI over a teacher when they don’t understand something.
I wrote about Bastani et al.’s (2025) finding that generative AI without guardrails harms learning. CDT’s chatbot data is the social-emotional version of that same argument: AI without guardrails doesn’t just affect academic outcomes, it reshapes how students relate to the adults around them.
The report also flags something educators working in special education need to see. Students with an IEP or 504 plan are more likely to have back-and-forth conversations with AI (73% vs. 63%). They report feeling disconnected from their teachers at higher rates (57% vs. 46%) and express more concern about unfair AI treatment (39% vs. 25%). Linsenmayer’s (2025) OECD report on AI and special education raised the question of whether AI tools built for general populations serve students with disabilities well. CDT’s polling data suggests the answer isn’t reassuring yet.
AI Literacy Training Isn’t Keeping Up
Laird et al. find that 48% of teachers and 48% of students have received some form of AI training from their school. Satisfaction is high among those who got it: 86% of teachers and 87% of students found the training helpful. The content of that training, though, is where the real gap lives.
The most common training topics for teachers are basic tool usage (29%) and explaining what AI is (25%). For students, the top topic is school policy about when they can and can’t use AI on assignments (22%). What’s barely covered: only 11% of teachers received any training on how to respond when a student’s AI use harms their well-being. Only 14% got guidance on what to do when AI tools produce biased or incorrect results.
Schools are training people to use AI. They’re not training them for what happens when it breaks. That gap maps directly onto what Hillman, Holmes, and Duarte (2025) identified in their rapid review of AI literacy frameworks for the Royal Society: most frameworks emphasize functional skills and give too little space to critical evaluation and risk awareness.
CDT also reveals a priority mismatch that itself creates risk. Teachers care most about effective tool use, students want to know what to do when AI breaks, and parents are focused on privacy protections. Only 17-21% of parents report receiving information on the topics they ranked as most important.
Three needs, three gaps, and the training as it exists now covers none of them fully. That disconnect is itself a risk factor: when parents don’t feel informed, backlash builds. CDT’s data shows 72% of parents think they should be able to opt their children out of AI tools in class, while only 43% of teachers agree.
What This Means for Schools Adopting AI
This report doesn’t argue against AI adoption, and I don’t either. It argues that adoption and risk are traveling together, and schools that separate the two conversations will keep getting blindsided.
A few things come through clearly. The deepfake and NCII data is alarming enough to demand immediate policy attention: 69% of teachers say their school hasn’t shared any policies on addressing deepfake NCII, and only 9-10% received guidance on handling incidents. That’s a vacuum where there should be a plan.
AI literacy training needs to pivot from “here’s how to use the tool” to “here’s how to tell when the tool is failing you.” Fan et al. (2025) coined the term metacognitive laziness to describe what happens when students let AI handle the cognitive work. CDT’s data shows schools aren’t preparing anyone to recognize or resist that tendency.
And one detail from the report deserves a hard look: 24% of teachers say AI was automatically added to an educational tool they were already using, with no separate adoption decision. Schools aren’t always choosing to take on these risks. The risks are arriving bundled with products they already purchased. Procurement teams need to start asking different questions.
The AI tools will keep improving. The risks won’t resolve themselves. The schools that get this right will be the ones that treated risk management as part of the adoption plan from day one.
References
- Bastani, H., Bastani, O., Sungu, A., Geb, H., Kabakcı, Ö., & Marimane, R. (2025). Generative AI without guardrails can harm learning: Evidence from high school mathematics. Proceedings of the National Academy of Sciences, 122(26), e2422633122. https://doi.org/10.1073/pnas.2422633122
- Fan, Y., Tang, L., Le, H., Shen, K., Tan, S., Zhao, Y., Shen, Y., Li, X., & Gašević, D. (2025). Beware of metacognitive laziness: Effects of generative artificial intelligence on learning motivation, processes, and performance. British Journal of Educational Technology, 56(2), 489–530. https://doi.org/10.1111/bjet.13544 //
- Hillman, V., Holmes, W., & Duarte, T. (2025). A rapid review of AI literacy frameworks, with policy recommendations. A report prepared for the Royal Society. London: The Royal Society. https://royalsociety.org/-/media/policy/projects/ai-in-education/hillman-et-al-a-rapid-review-of-ai-literacy-frameworks.pdf
- Laird, E., Dwyer, M., & Quay-de la Vallee, H. (2025). Hand in hand: Schools’ embrace of AI connected to increased risks to students. Center for Democracy & Technology. https://cdt.org/insights/hand-in-hand-schools-embrace-of-ai-connected-to-increased-risks-to-students/
- Linsenmayer, E. (2025). Leveraging artificial intelligence to support students with special education needs (OECD Artificial Intelligence Papers No. 46). OECD Publishing. https://doi.org/10.1787/edu/wkp(2025)12
- Luo, J. (2025). How does GenAI affect trust in teacher-student relationships? Insights from students’ assessment experiences. Teaching in Higher Education, 30(4), 991–1006. https://doi.org/10.1080/13562517.2024.2341005