|
Getting your Trinity Audio player ready...
|
The Policy Everyone Wrote and No One Uses
Sometime in the last two years, nearly every school leadership team sat down to write an AI policy. The process was usually the same. A working group was formed, a few examples from other institutions were collected, and a document emerged with familiar sections: a statement of principles, a list of prohibited uses, a paragraph on academic integrity, and a line promising to “review regularly.”
Then the semester started, and the policy did not matter.
Teachers kept making their own calls. Students found the boundary and worked just inside it. Vendors shipped AI features into tools the school already paid for, and no one checked whether the policy covered them. By the end of the term, the document sat in a shared drive while actual AI practice in the school was shaped by individual teachers, student ingenuity and software update notes.
This is not a failure of effort. It is a failure of design. Most school AI policies are written as if the problem were a lack of rules. The real problem is a lack of decision-making structure.
Diagnosis: Three Kinds of Policy That Fail
Most AI guidelines that collapse fall into one of three types.
The prohibition policy. It bans certain uses, usually generative text in student submissions, and treats compliance as a matter of enforcement. It fails because it depends on detection, and detection is unreliable. It also produces a culture where teachers become investigators and students learn that the goal is not getting caught rather than learning anything.
The principles policy. It affirms that AI should be used “responsibly, ethically and transparently” and that human judgment remains central. Nobody disagrees with any of it, and nobody can act on it. A teacher facing a specific Tuesday-morning decision, such as whether students may use an AI tool for a draft, gets nothing from a sentence about responsibility.
The tool-list policy. It names approved and unapproved products. It is at least concrete, but it is out of date before it is published. The AI landscape moves faster than a committee cycle, and the policy has no way to handle tools that did not exist when it was written or features added to products already in use.
All three share the same flaw. They describe what the institution believes rather than who decides what, when, and on what basis.
The Mechanism: Policy Without Decision Rights
A working policy answers operational questions. Who can approve a new AI tool? What information must be reviewed before approval? Who is accountable if a tool mishandles student data? What happens when a teacher wants to use something that isn’t yet approved?
Most policies leave these unanswered, so decisions default to whoever is closest to the problem. That is usually a teacher, an IT administrator or a department head, each working with partial information and no mandate.
The result is a pattern school leaders will recognize:
- Teachers adopt free tools individually because the approval path is unclear or slow. Student data ends up in systems the school never assessed.
- IT teams find out about tools after they are already in use, and are left choosing between blocking them and quietly tolerating them.
- Assessment practice diverges between classrooms. One teacher permits AI assistance, another treats it as misconduct, and students experience the inconsistency as arbitrary.
- Vendor updates change what a product does, and no one is responsible for noticing.
The policy did not fail because people ignored it. It failed because it gave them nothing to follow.
The Assessment Gap Nobody Wants to Own
The deepest weakness in most AI policies sits in assessment. A policy can declare that students must not submit AI-generated work as their own, but that statement only holds if assessment tasks are designed so that the distinction can be observed.
If an assignment can be completed entirely by a text generator, a prohibition does not protect its integrity. It just moves the problem into a dispute about suspicion. Policies that address conduct without addressing task design push the cost onto teachers, who must adjudicate cases with no reliable evidence.
Serious policy work here means asking which assessments still measure what they claim to measure, and which need redesign: more process evidence, more in-class components, more oral defense of written work. That is a curriculum and assessment decision, not a discipline decision, and most AI policies never reach it.
The Workload Problem
A second gap is teacher time. A policy that requires teachers to evaluate tools, disclose AI use, redesign assessments and monitor student behavior, with no reduction in other duties, is an unfunded mandate. Teachers will comply where compliance is cheap and drift where it isn’t.
Any policy that does not state what it costs teachers in hours, and what gets removed to make room, is asking for goodwill it has not earned.
The Equity Dimension
AI policy failures are not evenly distributed.
Well-resourced schools can absorb the gaps. They have IT staff who vet tools, instructional leaders who redesign assessments, and families who supplement at home. When their policies are vague, the harm is limited.
Under-resourced schools face the opposite pattern. With no capacity to vet tools, teachers rely on free products whose business model is data collection. With no assessment redesign, prohibition policies fall hardest on students who lack the language fluency or support to defend their own work when accused. And where detection software is used to enforce the rules, its error rates fall unevenly, with students writing in a second language among the most exposed to false suspicion.
Students also differ in what they can do outside school. A ban on AI in class doesn’t remove AI from the lives of students who have unrestricted access at home. It only removes the guided, supervised use that would teach them to do it well. Students without that access at home get neither the tool nor the instruction.
A vague policy is therefore not neutral. It shifts risk toward the students and schools with the least ability to manage it.
What a Working AI Policy Contains
The alternative is not a longer document. It is a smaller set of operational commitments that people can act on. Five components matter most.
- Decision rights. State who approves AI tools, who reviews data handling, and who is accountable. Name roles, not committees. Include a fast path for low-risk requests and a slower one for tools that touch student data. The aim is that a teacher with a request knows exactly where to send it and how long to expect to wait.
- An approved-tools register. Keep a living list of tools in use, what they are used for, what data they receive, and who approved them. The register is not a ban list. Its purpose is visibility. Include tools that arrive as features inside existing platforms, which is where most surprise adoption happens.
- A review cadence. Set a fixed rhythm, such as each term, for revisiting the register and the policy itself, and assign a named owner. Also define triggers for an off-cycle review: a significant vendor change, a data incident, or a new category of tool.
- An assessment position. Require each department to identify which assessments are vulnerable to AI substitution and to redesign or reclassify them. Distinguish tasks where AI assistance is permitted, tasks where it is permitted with disclosure, and tasks where it is excluded, and make the category visible to students on every assignment.
- A data and procurement baseline. Set minimum conditions for any tool that handles student information: what data is collected, where it is stored, whether it is used for model training, and how it is deleted. If a vendor cannot answer these questions, the tool doesn’t get approved. This connects AI policy to the wider question of how schools govern the platforms they already run.
Making It Survive the Semester
Even a well-designed policy needs conditions to last.
- Involve teachers in writing it. A policy drafted by administrators and handed down tends to be read once. One tested against real classroom scenarios gets used.
- Write for scenarios, not categories. Include worked examples: a student uses AI to outline an essay, a teacher wants to use an AI feedback tool, a vendor adds a chatbot to the LMS. Show how the policy resolves each.
- Give the owner authority and time. A policy owner without either is a title rather than a function.
- Measure whether it is working. Track how many tool requests were made, how long approval took, and how many tools were found in use that were not on the register. Rising discovery of unregistered tools is a sign the process is too slow or too opaque.
- Expect to be wrong. Publish the policy as a version, not a verdict. Institutions that treat their first AI policy as provisional revise it sooner and more honestly.
The Bottom Line
The AI policy vacuum is not caused by a shortage of documents. It is caused by treating AI governance as a statement of values rather than a system of decisions. Principles tell people what an institution hopes for. Decision rights, registers, review cycles and assessment redesign tell them what to do on Tuesday morning.
Schools that build the second kind will still make mistakes, but they will find them early and correct them. Schools that publish the first kind will discover, one semester later, that their policy governed nothing.