Although the use of artificial intelligence has gone far beyond search engines, lesson templates, and calendar organizers, many teachers still use AI primarily for such surface-level tasks.
As AI models advance, teachers increasingly need training not only in the basics of the program, but also in how to use their own professional judgment and expertise when working with AI technology.
On March 18, about 50 teachers gathered here to learn about developing “Agent” AI tools, autonomous software systems that can perform more complex multi-step tasks, including elements of reasoning to support teachers across subjects, grade levels, and platforms.
For Jing-Liang Guan, a science teacher at the Brooklyn Academy of Science and Technology, this can mean the difference between a tool that allows her to create lesson plans and a tool that helps her stress-test her lessons for content gaps and confusing language, and refine her teaching methods over time.
The New York City training is part of the National Academy for AI Educational, a five-year, $23 million partnership between the American Federation of Teachers and three leading AI developers: Anthropic, Microsoft, and OpenAI to train 400,000 teachers how to use technology in the classroom. With limited support from developers, the academy leverages teachers to train other teachers on how to use AI to improve instruction.
Percentage of teachers using AI-powered tools will nearly double from 2024 to 2025according to a national survey by EdWeek Research Center. Six in 10 teachers say they use AI in their practice, but it is still used most often for basic lesson planning and administrative tasks rather than to improve instruction.
AFT President Randi Weingarten said, “We are in a race for teacher knowledge” to make more meaningful use of AI. “This is going to be the most disruptive technology of our time. … We really need educators to learn to be in the driver’s seat of AI, not companies, regions, or the technology itself.”
AI used for real-time problem solving
These agent AI tools (agents) allow teachers to narrow down the information the AI uses to respond to requests based on their professional judgment and experience, said Seth Reznik, team lead for Microsoft Elevate, the tech giant’s $4 billion training initiative for schools and nonprofits. This makes it less likely that the AI will hallucinate, fabricate information, or give responses that are simplistic or not based on the latest research, as previous studies have shown AI assistants often do..
Lois Torres, a preschool paraeducator in New York City public schools, wants to develop a research-backed AI agent that can help co-teachers and her brainstorm faster alternative approaches when classroom, behavioral, or academic interventions aren’t working.
“A lot of teachers are doing this from home, and they’re just racking their brains trying to figure out what’s going to help them the next day,” Torres said. “Sometimes teachers and paras like me take a long time to think about how to help a child. And when something doesn’t work, it’s like, ‘Okay, we have to brainstorm on the fly, what else should we do?'”
Similarly, Yaseema Cook is developing agents to help create and monitor individualized education programs and differentiate lessons for 12th graders in New York City’s independent special education classrooms with autism, intellectual disabilities, and a variety of other needs. She said she can use students’ daily progress to adjust lessons throughout the week.
Cook and other teachers participating in the training believe that a careful balance needs to be struck between protecting student privacy while providing AI with enough context about a student’s disability or other needs to get meaningful and targeted assistance. New York City public schools have not yet released formal AI usage guidance, expected in March, and teachers said it is not always clear what data may be used to identify students for each tool.
For example, Jennifer Watters, a third-grade teacher at PS 29 in Queens, said she has been experimenting with AI since 2019 and regularly uses it for higher-order activities, such as developing questioning techniques for students from different cultures. But Watters said he switched from using OpenAI’s ChatGPT to Anthropic’s Claude last month due to data privacy concerns. This followed a high-profile showdown in which Anthropic refused permission from the Department of Defense to use its model for domestic surveillance..
“It is critical that teachers know that if this information and this tool falls into the wrong hands, it can be extremely dangerous to our students, our profession, and our jobs,” Watters said. “When we talk about our ethical obligations, privacy, and student data, it’s very important that we abide by them. But we need to know: Does the AI abide by that too?”
Teachers said the agent is best suited for tasks that have clear rules, guidelines, and tone, and where teachers can direct the AI to specific trusted sources. For example, some teachers have developed agents that write letters to parents about everything from upcoming field trips to academic assignments and discipline. It is tailored to each individual student, follows district and union guidelines for parent involvement, and includes next steps based on appropriate research and resources.
However, some experts caution against entrusting these common teacher tasks to AI agents. For example, a letter-writing agent learns to refer to details about a particular family or to imitate a teacher’s writing voice through examples of real letters, which, paradoxically, the teacher is less likely to remember. Important details about the student.
And teachers said the ever-evolving AI technology will require ongoing training and guidance to use in the most effective and ethical way.
“What we’re seeing nationally is that the more someone uses it, the more [AI]”There are still a lot of fears about the lack of federal guardrails around privacy, safety, disinformation, and academic freedom,” Weingarten said.
