Dexter schools pilot locally hosted AI tool for classroom use, citing privacy and control

Dexter Community Schools staff on Tuesday outlined a district-built artificial intelligence tool they say will let students and teachers use a chat-style LLM on local servers while keeping student data in-house.

The presentation to the Dexter Community School District Board of Education described a locally hosted model accessed from the district domain, running on a Mac Studio housed at Creekside. "We can host our own AI tool," one presenter said, adding that the district paid about $6,000 for the initial device and has since purchased additional machines to expand capacity. The team said the tool provides roughly "75 to 80% of the capability of chatGPT," but runs entirely on local hardware so inputs do not leave district servers.

District technology staff said the pilot began with roughly a dozen advanced-placement coding students over the summer, expanded to about 20 users and staff, and now has "over 300 to 400 active users" on a regular basis. The staff described three core design priorities: affordability, the ability to customize guardrails, and data security so student prompts do not get sent to external vendors.

"We wanted a solution that can not only be affordable and also protect student rights and data and also allow us to customize it," said Dalton Dieter, one of the district technology staff members who presented the work. Dieter described the team’s approach as moving from generative AI toward "instructional AI" — a system that prompts students through thought processes and references teacher-provided rubrics or syllabi loaded into a district knowledge base.

The presenters described several privacy and safety measures. Memory and long-term context were disabled so the model will not retain per-student histories, staff can review chat logs but students cannot alter them, and the model will not access the district’s student information servers. The team also said they stress-test different models and system prompts to block inappropriate content and to reduce the risk of “injection attacks” that try to trick an AI into producing disallowed outputs.

"Out of the box, it knew not to do those things," Dieter said, referring to attempts to prompt the pilot to produce inappropriate or dangerous instructions. The district said staff tested the system by asking it to produce problematic content and that the model refused those prompts during testing. The presenters also said the model’s training dataset cut-off is about 2023 for the Llama-family models they are using.

Staff described instructional guardrails for classroom use: a four-level scale ranging from level 0 (no AI allowed, for summative assessments) to level 4 (open AI use with disclosure). The district plans a disclosure process so students must report when and how they used AI on assignments, and a separate permission/disclosure workflow for cases where teachers approve specific uses.

The district said uses seen so far include idea generation, text leveling for younger readers, practice tools for music (audio/video features planned), and teacher-facing efficiency tasks. Presenters emphasized that the intent is to have AI initiate or support work but to have a "human finish" — teachers remain responsible for final evaluation and grading.

Board members and attendees asked about updates, model training, energy use, and rollout. The presenters said they will continue stress testing models before broader deployment and will collect regular feedback via Google Forms and chat-log review. They invited the board to return questions about long-term maintenance and training for teachers.

District staff said they have sought outside feedback from AI industry contacts including conversations with Google engineers and other AI leaders. They also emphasized environmental considerations, saying local models reduce the need for large datacenter processing for routine classroom tasks.

As presented to the board, the project remains a pilot with controlled expansion and teacher-driven permissions. The district did not present a final policy or a universal timetable for full rollout; presenters said they will move from the current "initiate and build" phase into "evaluate and engage" as they gather classroom feedback.

The presentation and board discussion occurred during the meeting’s presentation portion and drew questions from multiple board members. The district indicated the pilot will continue to broaden in coming months and that the team will return with further details as they refine guardrails and model updates.