When a School Starts Talking to Itself

Michael Keany

Imagine a school where nearly every message is composed, summarized, interpreted or answered by artificial intelligence. The humans still press “send,” but increasingly, chatbots are doing the talking. The communication loop described in this New York Times Magazine article has entered the schoolhouse—and a typical day now looks subtly, consequentially different.

At 6:15 a.m., the principal opens an AI-generated summary of overnight messages. A parent’s anxious email about bullying is labeled “emotionally charged but low risk,” and the system drafts a calm response promising follow-up. What the principal cannot see is that the parent also used AI to compose the complaint. One chatbot has converted fear into polished institutional language; another has condensed that language into a dashboard item. The exchange is efficient, but the parent’s urgency has been softened twice before any person responds.

By first period, a teacher asks an AI assistant to turn a district curriculum memo into a lesson. Students then use their own chatbots to interpret the directions. The teacher’s intent passes through one model; the students’ understanding passes through another. Instead of wrestling with the question, several students ask what answer the teacher is “probably looking for.” Their polished submissions satisfy the rubric, yet class discussion reveals that few can explain their reasoning without assistance.

During lunch, two students disagree in a group chat. Maya uploads the exchange to a chatbot, which tells her she may be experiencing exclusion and drafts an assertive email to the counselor. Jordan feeds Maya’s email into another chatbot, which characterizes it as accusatory and prepares a detailed defense. The counselor receives two articulate, logically organized accounts that describe different realities. Because neither sounds like the students’ ordinary voice, it is harder—not easier—to hear the uncertainty, embarrassment, and hurt beneath the conflict.

Meanwhile, administrators review an AI analysis of attendance, behavior and assessment data. The program recommends moving twelve students into an intervention block. Teachers receive an automatically generated request for feedback and use AI to draft their replies between classes. By the meeting, the principal sees what appears to be strong agreement. In fact, several teachers had reservations, but their assistants converted tentative objections into neutral, cooperative prose. The system has not fabricated consensus; it has edited away the friction that might have improved the decision.

At 3:30, a parent receives an automated progress alert. She asks a chatbot whether the grading policy is fair, then sends its forceful response to the teacher. The teacher asks AI to make a reply “warm but firm.” The parent asks AI what the reply “really means.” Each turn produces more language, more certainty and less direct understanding. A five-minute phone call becomes a 14-message exchange between systems acting as advocates, interpreters and shields.

None of these uses is automatically harmful. AI can translate, organize, clarify and help people who struggle to express themselves. The danger arises when assistance becomes invisible mediation and when generated confidence replaces human curiosity. In schools, relationships are not an administrative inconvenience; they are part of the learning infrastructure.

Leaders should therefore establish a simple principle: Use AI to prepare for communication, not to replace it. Disclose meaningful AI assistance. Keep high-stakes judgments and emotionally sensitive conversations human. Teach students to verify interpretations, retain their own voice, and ask another person, “Is this what you meant?”

The central leadership question is no longer whether schools will use AI. It is whether, amid all the fluent messages, the people in the building will still genuinely speak—and listen—to one another.

Original Article

------------------------------

Prepared with the assistance of AI software

OpenAI. (2026). ChatGPT (5.2) [Large language model]. https://chat.openai.com