Universities Don't Need a $17 Million Chatbot
Cal State is deciding right now whether to renew its $17 million contract with OpenAI for system-wide ChatGPT access. Faculty are split. Students are split. A petition to end the deal is circulating. Other large systems are watching to see what happens.
The debate is being framed as AI: for or against. That's the wrong frame. The real question is narrower: is a general-purpose chatbot the right tool to put in front of half a million students who are trying to learn?
It isn't.
What a general-purpose chatbot is good at
ChatGPT is an extraordinary research assistant. It can rewrite an email, brainstorm a thesis, debug a script, summarize a paper, translate a sentence. For a knowledge worker who already knows what they're doing, it's a productivity multiplier. That's the use case it was designed for and the use case it's optimized for.
That is not the use case in a chemistry course.
A student working through a stoichiometry problem doesn't need a faster way to get the answer. They need to slow down, work the steps, hit the productive confusion, and come out with a mental model they didn't have an hour earlier. The thing that makes ChatGPT useful at work is exactly what makes it counter-productive in a classroom: it short-circuits the process learning depends on.
Pedagogy is structural, not stylistic
You can't fix this with a better system prompt. You can ask ChatGPT to "act as a tutor" and it'll do a reasonable impression for a few turns, and then a student will type "just give me the answer," and it will.
Tutoring is a different kind of product. It has to refuse the right things, scaffold milestones, surface where a student is stuck, give the teacher signal back about what the class is and isn't getting, and connect to the specific content of a course. Those aren't toggles on top of a general model. They're the architecture.
We've built one of these for STEM and we'll be the first to tell you it was harder than it looks. It's also dramatically cheaper than $17 million.
The equity argument cuts the other way
The standard worry about AI procurement is that flagship institutions can afford the enterprise license and smaller ones can't, leaving regional schools and community colleges behind. That worry is real if you accept the premise that what students need is access to ChatGPT.
If you don't accept that premise, the picture flips. A subject-specific tutor doesn't need a half-million-seat enterprise contract to work. A community college chemistry department does not need to outbid Cal State to give its students a better AI experience than Cal State's students will get. The school that wins is the one that picks the right tool, not the one that signs the biggest check.
What the procurement question should actually be
If you're a provost or a CIO looking at an AI contract this summer, the question to ask isn't ChatGPT or no ChatGPT. It's a sequence.
What are we trying to do? If the answer is "help faculty and staff get more done," license a productivity tool. If the answer is "improve how students learn," that's a different product category entirely, and you should not buy the first one and hope it does the second.
Who controls what the AI will and won't do for a student? If the answer is "the vendor, via a system prompt we can't see," you've outsourced a pedagogical decision to a company with no stake in your courses.
What does the instructor get back? A consumer chatbot gives instructors nothing. A teaching tool should give them a real picture of where their class is stuck — without putting individual transcripts on a dashboard.
What does the integrity story look like? "We trust students not to misuse it" is not a story. It's a wish.
The shorter version
Cal State's $17 million is buying the wrong thing for the use case that's getting the most attention. Faculty are right to push back, but the push-back is too narrow. The right ask isn't "no AI in our classrooms." It's "the right AI for what classrooms actually do."
There's a real version of this product. It costs a small fraction of $17 million. It's built for teaching instead of retrofitted from a productivity tool. And it leaves the people who actually run a course — the instructors — in control of what their students see and don't see.
If you're inside an institution working through this question right now, we'd love to talk. We're running fall pilots with universities and community colleges making exactly this call.