What the Scale Was Hiding

On May 25, NPR reported that California State University, the largest public university system in the country, has renewed its contract with OpenAI to provide ChatGPT Edu to all 460,000 students and 63,000 faculty and staff across its 23 campuses.¹ The original deal last year was a $17 million no-bid contract.¹ The renewal is roughly $13 million a year for three years, again signed without a competitive bid.¹ When the chancellor, Mildred García, announced the original initiative in early 2025, she described it as a project of unprecedented size. No other university system in the U.S. or internationally is doing anything like this, not at this scale, she said.²

The scale is real. So is the picture the system's own survey draws. In fall 2025, Cal State surveyed 94,060 students, faculty, and staff about how they were using the tool.³ ChatGPT was named the most used AI tool by 84 percent of students, 87 percent of faculty, and 89 percent of staff.³ Sixty-four percent of respondents said AI had affected their learning experience positively.³

The other numbers are the harder ones. Fifty-two percent of faculty said AI was having a negative effect on their teaching.³ Sixty-seven percent of students said their professors were not teaching them how to use AI effectively.³ Roughly two-thirds of students and three in five faculty said they remained skeptical that AI was benefiting education at all.¹ Eighty-two percent of both groups worried about the tool's effect on creativity.¹

What the Petition Was Naming

A faculty-led petition opposing the contract said the quiet thing plainly. ChatGPT Edu, the signers wrote, is "a general-purpose chatbot that is not designed, trained, or optimized for education."⁴ Martha Kenney, a science studies scholar at San Francisco State and one of the petition's organizers, put it more sharply in her interview with NPR. Offering students a chatbot that lets them take shortcuts on assignments, she said, is "cheating our students out of an education."¹

Both lines name the same thing. The system has measured scale, contract value, and adoption. What has not been measured, because the tool was not built to surface it, is what students are actually learning while they use it. Whether they are wrestling with the material or routing around it. Whether they are building the muscle of reading slowly and writing badly first and then better, or being handed a finished sentence and rephrasing it once. The survey can count the logins. The institution cannot see the practice underneath them.

Scale Was the Wrong Unit

What the next five years of higher education will need is not a larger contract. It is a more honest answer to a smaller question. What does it look like, on a given Tuesday, in a given history seminar, when a student uses AI well? What does it look like when she uses it poorly? What does the difference feel like to the professor reading the work three days later, and what would it have taken for the professor to see the difference earlier, while the student was still drafting?

None of those questions can be answered by a tool whose primary feature is fluency on demand. They can only be answered by tools that hold on to the work. The half-written paragraph at 11 p.m. The prompt the student wrote, the answer the AI returned, the parts she kept, the parts she rewrote, the place where she paused for forty minutes and came back with a sharper sentence. None of that survives in a chat window. All of it is the substance of learning.

This is part of why Koan has been built the way it has. Not as a smarter chatbot, but as a quieter one, designed for the classroom and trained to behave like a tutor rather than a finisher. The work, the revisions, the small breakthroughs, all of it gets preserved. A professor can see how a student arrived at her argument. A department can see, across a semester, whether students are getting better at thinking. A chancellor signing a contract can be looking at evidence of learning, not adoption rates.

The Cal State debate is not really a debate about OpenAI. It is a debate about what counts as evidence that learning is happening. The survey numbers are the first honest version of that question. A majority of the faculty closest to the work say the tool is hurting their teaching. Two-thirds of students say no one has shown them how to use it well.³ Those two findings, sitting next to each other, are the whole problem in miniature. The scale was decided. The pedagogy was not.

If your institution renewed its largest software contract tomorrow, would the people closest to your students say it was making their teaching better?