SACRAMENTO, CA — Roosevelt Elementary School teacher Margaret Walsh made the unsettling discovery Monday that her entire third-grade class had been secretly feeding their homework assignments into a shared ChatGPT account, effectively creating a collaborative AI trained exclusively on eight-year-old academic responses.

The revelation came when Walsh noticed that all 23 students had submitted identical book reports on "Charlotte's Web," each concluding with the phrase "Charlotte was probably really good at computers." Further investigation revealed the students had been copy-pasting the AI's output since the first week of school, inadvertently teaching it to write like children.

"What started as simple homework help evolved into something far more sophisticated," explained Dr. James Chen from UC Davis's Department of Educational Technology. "These kids accidentally created a feedback loop where AI writes like a third-grader because it learned from third-graders who were already using AI."

Student ringleader Emma Rodriguez, 8, defended the operation: "Mrs. Walsh said we should use technology to help us learn. We're helping the computer learn too. It's called cooperation."

The custom chatbot, dubbed "Homework Helper 3000" by the students, has developed disturbing quirks after months of training on elementary-level work. It consistently misspells "because" as "becuase," believes George Washington invented the telephone, and insists that all math word problems involve buying candy.

"The most concerning part is that it's actually improved their collective GPA," Walsh admitted. "The AI writes better than half my students did before they started cheating. I'm not sure if I should confiscate their laptops or nominate them for a technology innovation award."

The Sacramento Unified School District has yet to determine whether the collaborative AI training constitutes academic misconduct or an advanced group project.