ChatGPT for Teachers Adds Kamehameha Schools, Keeps Students Out

A Hawaiʻi first, with limits built in

Kamehameha Schools is now part of OpenAI's latest ChatGPT for Teachers cohort. It is the only Hawaiʻi school in the group, and the deal gives its administrators and faculty access to the platform while giving students none.1 The school system serves more than 5,400 K-12 students on campuses on Oʻahu, Maui and Hawaiʻi Island.1 OpenAI announced the partnership on August 26. It covers institution-managed access to ChatGPT for educators and staff, plus training and support meant to help schools use AI responsibly.2

The decision to leave students out is the most important detail, and every outlet reported it. Star-Advertiser and Pacific Business News coverage, Texas Public Radio as cited by a policy tracker, and OpenAI's own statement all say no student accounts come with the program.125 That limit was a choice, and it says as much about where AI in schools stands as the partnership itself.

The national rollout behind the local deal

Kamehameha is one of 55 school systems in 20 states added in this expansion, which OpenAI says brings more than 100,000 more educators and staff onto the platform.2 Other new members include Baltimore City Public Schools, the Boise School District, Northside Independent School District in San Antonio, Sarasota County Schools and Virginia Beach City Public Schools.2 Business Insider described the expansion as roughly doubling the product's reach and said Houston's district is also now a participant.14

The overall numbers differ slightly depending on the outlet. The Star-Advertiser said OpenAI now works with more than 100 K-12 organizations in 30 states, covering about 340,000 educators and staff who serve more than 2 million students.2 OpenAI's own wording, quoted by 9to5Mac, said "over 300,000 educators and staff."11 One independent tracker pointed out that the list of districts on OpenAI's site names only 31 of the 55 systems and includes some from the first cohort, so it should not be read as a list of new members.5 The gap is minor, but it shows that program-size figures in this market come almost entirely from the vendor.

The free period is now longer. ChatGPT for Teachers launched in 2025 with nearly 150,000 educators and staff, and it is now free for verified U.S. K-12 educators through June 2028.3 When it launched, CNBC reported that free access would end in June 2027. The extension to 2028 got little coverage, even though OpenAI's own launch page was apparently edited to show it.5

Privacy terms matter more than the chatbot

The sharpest analysis argues that the most consequential part of the announcement is a legal document, not the AI. OpenAI also announced a multistate National Data Privacy Agreement built on the Student Data Privacy Consortium framework. It is designed to spare districts from negotiating separate privacy contracts from scratch.3 TheStreet argued that the agreement "is doing more work than the chatbot," and noted that OpenAI education chief Leah Belsky pitched the deal around role-based permissions and administrative control, a message aimed at district lawyers more than at teachers.12

Even the list of states is inconsistent. Hawaiʻi and Garden Island coverage named 16 states, including New York, with California under a separate agreement.23 A tracker comparing OpenAI's own post found it gives the 16-state headline figure but lists 15 signed agreements, without New York.5 Hawaiʻi is not on either list, so Kamehameha, as a private school, joins under OpenAI's general workspace terms rather than a state framework. That is an inference from the published lists, not something the coverage says directly.

The workspace has administrative controls and role-based access, and OpenAI says data entered there is not used to train its models by default.3 Guides for schools add a caveat: the no-training default does not on its own decide what student information staff should be allowed to upload.1618

Saving teachers time is the pitch

OpenAI is selling teacher time, not student tutoring. Its usage analysis for January 1 through July 16 counted more than 1.9 million messages about time-saving tasks. About 900,000 involved report cards and progress reports, about 800,000 involved lesson planning, and more than 100,000 each involved substitute plans and teacher-evaluation materials.2 Lynwood Unified superintendent Patrick Gittisribongul, from an earlier cohort, described the goal as making room for teaching rather than replacing it.3

This is consistent with earlier outside research. A Gallup poll found that 60% of teachers used at least one AI tool in the 2024-25 school year, and weekly users reported saving nearly six hours a week.9 Still, OpenAI's message counts measure activity, not results. The largest single category, report cards, is also where the risk is highest. Some school AI policy templates bar teachers from writing student evaluations "without substantial personal input."13 How Kamehameha and other districts handle that tension will matter more than how many messages get sent.

The Register was harsher about the direction OpenAI is taking. It noted that new agentic education plugins arrived amid concerns about AI-enabled cheating and research suggesting the products can hamper learning.19 Belsky's reply, reported by Business Insider, is that more capable AI makes "teacher judgment, teacher guidance, and district oversight even more important than before."14

The research case for AI tutoring is mixed

The teacher-only design looks more sensible against the evidence on student-facing AI tutoring, which points in different directions. The most cited positive result is a 2025 randomized trial at Harvard. Undergraduates using a custom-built physics tutor learned significantly more in less time than students in an active-learning class, and they reported feeling more engaged.21 Two K-12 tutoring trials reviewed by Stanford's National Student Support Accelerator also found gains when AI was built into live, human-supervised math tutoring. Students using Tutor CoPilot were four percentage points more likely to master a topic, with bigger gains among students paired with less experienced tutors.26

A new working paper complicates that picture. In a randomized trial with 2,379 undergraduates at a large U.S. public university, access to a course-integrated AI tutor lowered final grades by up to 0.37 standard deviations. It also sharply reduced participation in course platform activities, and grade losses were more than twice as large for first-generation students.29 The authors cite other studies showing that general-purpose AI access hurt later independent performance in high school math, while tutoring guardrails reduced that harm.29

The most defensible reading is that design and structure decide the result. A purpose-built, pedagogically constrained tutor can work. A tool simply made available alongside a course can push out the effort that produces learning. Meta-analytic summaries make a similar point: intelligent tutoring systems tend to beat large-group instruction but are roughly equal to individual human tutoring.24 For a school system serving Native Hawaiian students, with an explicit mission tied to educational disadvantage, the evidence that first-generation students lose more is a strong reason to move slowly on student access.2940

Academic integrity and the gap in student policy

The Star-Advertiser's editorial board backed the partnership's efficiency goals but said it, and others like it, "fail to address student access head-on." The board argued that banning AI is no longer workable when any student with a connected device can use a chatbot.6 That criticism holds up. Students' AI use is not on hold just because their teachers have an approved workspace.

Research on cheating suggests panic is the wrong response. A Stanford-affiliated follow-up survey of 4,354 students at six high schools found that self-reported cheating stayed close to pre-ChatGPT levels. About 24% reported using AI or devices as unauthorized aids, and nearly 83% said AI should never write entire papers.44 Private schools in that sample reported lower dishonesty rates than public or charter schools.44 Detection is a weak backstop. One widely cited study found that AI detectors flagged non-native English speakers' essays as AI-written at an average rate above 61%.43

What this points to is redesigning assignments and setting explicit permission tiers, not policing. One policy template, for example, defines which levels of AI help are allowed for which kinds of tasks and warns against relying only on detectors.13 Teacher training is also a precondition for that redesign: teachers who use AI themselves are better placed to write assignments that hold up when students have it.

Hawaiʻi's private schools are splitting

Kamehameha's careful approach puts it in the middle of the state's independent schools. Punahou gives high school teachers and students access to Google's Gemini for Education. Mid-Pacific Institute uses the monitored SchoolAI platform for younger grades and Gemini in high school. Hawaiʻi Baptist Academy is training teachers before rolling out student tools.3231 Punahou President Michael Latham framed the question as how, for what purposes and under what conditions to use AI, rather than whether to.32

That context explains why partnering with OpenAI's teacher program is a measured choice. For now, Kamehameha gets privacy-managed adult access, vendor-run training and a tool it does not pay for. Whether that adds up to a strategy depends on two things the announcement doesn't cover: how the school sets rules for students who already use AI, and what it does when free access ends in mid-2028. TheStreet put the commercial logic plainly, arguing that habits formed while a product is free rarely stay free.12 Districts that build routines around the workspace now should plan for possible pricing later.