When ChatGPT was publicly released in the fall of 2022, I was the chief academic officer and interim provost of a small college in New York City. The faculty and I immediately recognized the dangers this posed to serious learning and academic integrity—but that’s a topic for another post. What wasn’t as obvious at the time but probably keeps at least as many college and university administrators up at night today is the regulatory and legal danger associated with casual AI use by faculty and staff. When everyone has a free account with Claude or ChatGPT and “Googling” has been replaced by Gemini’s “AI Overview,” the probability that your school’s FERPA- or HIPAA-regulated data has been used to train someone’s AI model is very high.
Consider how little it takes. Suppose a financial aid counselor, short on time before a deadline, pastes a student’s file into a free chatbot and asks it to draft a reply to a family’s appeal. This doesn’t look like a data breach: the counselor is trying to help, and the tool is a browser tab away. But if the account’s terms allow the provider to train on what users submit, the student’s record may have left the institution’s control, and no one at the school will know it happened. IT departments have a name for this: shadow AI, the tools people adopt on their own, outside any review. How do you prevent institutional damage from casual misuse of AI tools? This post suggests three steps: finding out what is already happening, adopting policies people can live with, and supplying a tool that makes doing the right thing the easy thing.
Find out what AI tools your faculty and staff already use
While maintaining plausible deniability might be a tempting way out, it won’t impress an accreditor or, in the worst case, a jury. The first step in getting control of unregulated AI usage across your college or university is creating a reporting mechanism that allows departments, offices, and individuals to identify the AI tools and subscriptions they are currently using and the problems they are solving with them without fear of discipline. Of course, asking people to admit to what may be a violation of policy, or of law, depends on trust. The request will work only if the amnesty is real and credible. At this stage, getting a complete picture of usage on campus is of more value than sanctioning your team under what may be a patchwork of hastily composed AI rules or legacy policies shoehorned to answer AI-created questions.
Some of the tools your self-reporting effort uncovers may be reasonably secure. By default, a Claude Team subscription, for example, does not use data shared by team members to train its model. Others may be as bad as you feared. Some of the uses may be wise and appropriate; others may be entirely inappropriate or even illegal. Whatever you find, the more complete your understanding of the who and the how of AI usage on your campus, the better you are able to craft the policies and supply the tools that will mitigate the risk and help you achieve your ultimate goal: securing all the efficiencies AI promises without violating the trust or the rights of your community.
Write an AI acceptable-use policy people can live with
Laws that run against ingrained habits and seem calculated to assert control without reason are not always subject to public campaigns of civil disobedience, but they are often undermined by many acts of quiet resistance. An administrator who has just learned how much student data may already have left the building will want to close every door, but maximalist AI policy restrictions are likely to meet the same sort of response.
On the other hand, if you set clear boundaries that prohibit plainly illegal or unethical behavior, there remains plenty of room for using AI to relieve the burden of mundane tasks or benefit from its extraordinary computing power. The exact contours of back-office AI policies will vary with an institution’s culture and posture toward risk, but no one will have a difficult time defending or defining a policy that prohibits exposure of FERPA- and HIPAA-regulated data, protects proprietary and confidential information, and maintains ethical boundaries around issues like AI-human impersonation and AI-aided high-stakes decision-making. If, after adopting and publicizing such policies, your school still struggles with compliance, the issue is probably not the policies or your people, but the AI tools you have (or haven’t) given them.
Give faculty and staff a compliant AI tool
Our company prohibits our developers from using personal AI accounts for Watkyn-related work. That policy would be a dead letter if, on the positive side, we didn’t provide a Claude Team account to all—and weren’t willing to pay for enough tokens for them to finish their work. Developers, like college administrators, are practical people who want to get the job done. If you give them the tools they need to do it, they will gladly use them. If you don’t, they’ll be tempted to find some other way.
What does that mean for your college or university? If you give your faculty and staff an AI tool that meets their needs and your policies, they are very likely to use it. They want to comply with your policy and best AI practices and they want to be productive. The right tool allows them to do both.
AI platforms for colleges and universities: four options
Which tool, then? The market offers several answers, and they differ chiefly in cost, pricing structure, and flexibility.
The first segment of the higher education AI market is the proprietary AI systems built by flagship universities like the University of Texas and the University of Michigan. These generally provide the whole campus with access to the latest AI models with a UI and related features customized for the specific needs of the university. The price tag for the initial build, however, is beyond the reach of most schools.
A second option is to contract with Google, Anthropic, or OpenAI for their higher education AI solutions. All three offer flat-rate, per-seat campus plans, although the details of their packages—and what is bundled with a contract you already have—differ. Per-seat pricing has the virtue of predictability, since the bill is known in advance. The cost of that predictability is paying for seats that sit idle. Anthropic will also sell Claude to an institution on a pay-as-you-go basis through its API, where you are charged only for the tokens you actually use. With that model, of course, comes the challenge and risk of managing usage across departments and offices.
Alternatively, some schools may prefer to allow different divisions of the college or university to buy subscriptions to different AI tools, recognizing that different companies and models have varying strengths and weaknesses, as well as different pricing models. Rather than supplying a single solution, they can allow department and office heads to judge how best to deploy their AI budget within the general policies adopted by the school. The downside of that flexibility is that the school ends up with several sets of terms of service to read, and several places where student data might go.
Watkyn offers colleges and universities a fourth option: an AI solution that combines the best of these three approaches—Amplify GenAI, an open-source AI platform originally created by Vanderbilt University. We run Amplify GenAI ourselves, and our team of developers has built upon that excellent foundation, adding new security and UI features that make it an attractive platform your school can trust. From there, we can further customize your instance of Amplify GenAI without you having to pay for all the hours spent building the baseline solution.
What about cost? Since it runs in your AWS environment, you only pay for the infrastructure and tokens that you use, but to avoid billing surprises, it allows you to set department-by-department usage limits and budgets. This means the school decides how much each office may spend before the spending happens, not after. Moreover, through AWS you have access to AI models across the leading AI companies, so you aren’t locked into a single company’s offerings or pricing. At the same time, Amplify GenAI allows you to set up model restrictions that fit your institution’s policies. If you want to exclude Chinese models, for example, you can do that—and whichever models you choose, they won’t be trained on your data because your data stays within your AWS environment. With this approach, compliance is a matter of defining what the infrastructure you control does with your data rather than something buried in a third party’s terms of service. Price-conscious users can also compare the pricing for different models in the UI and employ more expensive ones for only their most difficult tasks.
Opening the black box, once
Opening up the back-office AI black box at your school might be scary, but understanding current AI usage across your institution, crafting realistic policies, and supplying useful, compliance-ready AI tools means you will not have to open it blind a second time. Whatever tool your school chooses, we know this: the students whose records are stored in your systems did not agree to become data sources for anyone’s model. Following the steps above, so that you lose neither their trust nor an accreditor’s confidence in your stewardship of it, should be a top institutional priority for any school that hasn’t uncovered and remediated unregulated back-office AI usage.
If Amplify GenAI sounds like it might be the right tool for your school, we’d be happy to show you how it provides access to the leading AI models while managing cost and maximizing compliance. Request an Amplify GenAI fit call → There is no fee for the initial scope assessment, and no obligation beyond the conversation.
