The Short Answer
An AI detector's score is not reliable enough to accuse a student on, so a university protects academic integrity by assessing how the work was made and by giving students an AI workspace whose agents are limited to permitted tasks. On ibl.ai the university runs that workspace itself: you own all the code and the data, choose any model, and decide what each agent may do.
This is the practical follow-up to MIT Sloan's guidance that AI detectors don't work. That guidance says what to stop doing. This post is about what a university builds in its place.
Why isn't an AI detector's score enough to accuse a student?
Because it is a probability about one finished document. A detector judges the output of models that keep changing, and a student can edit that output or simply ask for it in a different voice.
The early record was poor. OpenAI's own classifier correctly identified 26% of AI-written text and wrongly labeled human writing as AI 9% of the time.
On July 20, 2023 OpenAI withdrew the classifier, citing its "low rate of accuracy".
Independent testing reached the same place. A 2023 study led by Debora Weber-Wulff tested 14 detection tools, including Turnitin, and concluded they were "neither accurate nor reliable".
Detectors have improved since. A September 2025 working paper by Brian Jabarian and Alex Imas, summarized by Chicago Booth Review, tested GPTZero, Originality.ai and Pangram.
All three kept false positives below 1%, and all three lost accuracy on passages under 50 words. That is better than 2023, and still a score: it tells an instructor how likely, never how the work was made.
Who gets hurt when a university relies on AI detection?
The students who are wrongly accused, and they are not a random sample. Stanford researchers ran 91 TOEFL essays written by non-native English speakers through seven detectors.
Averaged across the seven detectors, 61.22% of those essays were wrongly labeled AI-generated, and all seven agreed on 18 of the 91.
The same detectors were near-perfect on essays by U.S. eighth graders. The paper warns against using detectors in educational settings for that reason.
Even a small error rate is large at campus scale. Vanderbilt disabled Turnitin's AI detector in August 2023 after doing the arithmetic on Turnitin's claimed 1% false positive rate.
Vanderbilt had submitted 75,000 papers in 2022, so around 750 could have been wrongly labeled. At that scale, a university that punishes on a detector's score risks accusing hundreds of students a year over work they wrote themselves.
What should a university assess instead of the finished essay?
The process that produced it. A finished document says nothing about how it was made, but a record of the work does, and that record is much harder to fake than a paragraph.
MIT Sloan's teaching guidance makes four recommendations: set clear AI policies, promote transparency, foster intrinsic motivation and teach inclusively. Its practical suggestions include process statements, scaffolded stages and revision.
In practice that means four kinds of evidence, none of which depends on guessing at authorship:
- Staged work. An outline, a draft and a revision, each submitted and each given feedback.
- The working conversation. What the student asked a course agent, what it answered and what they did next.
- A process statement. The student says which tools they used and for what.
- A short oral check. Five minutes explaining the argument shows quickly who wrote it.
On ibl.ai an instructor can read an agent's past conversations in its History panel, and an agent can grade work against a rubric the instructor defines, with a person able to override any score.
How does an AI workspace with limited tools protect academic integrity?
It makes the permitted path the convenient one. Students are going to use AI, so the useful question is which AI, set up by whom, allowed to do what.
On ibl.ai each agent has its own instructions and its own tool list, switched on or off tool by tool. A research agent for a writing course can search, explain and cite sources, and be instructed to question a draft instead of producing one.
Two independent checks sit around every conversation. A moderation prompt screens what the student asks, and a safety prompt screens the model's answer before the student sees it. The university edits both, and can review what was flagged.
None of this makes cheating impossible, and a vendor who says otherwise is selling a detector by another name. A student can still open a different tool. The three approaches differ in what they cost honest students:
| Approach | What it examines | Cost to an honest student | What it takes to cheat |
|---|---|---|---|
| AI detection | One finished document | A false accusation: an average 61% false positive rate on non-native English essays in the Stanford test | Edit the text or ask for a different voice |
| Process assessment | Drafts, feedback, a process statement, an oral check | More stages to submit | Fake every stage and then explain the work in person |
| A university-run workspace with limited agents | The working conversation, on the university's own systems | None: it is the AI they were going to use anyway | Leave the provided tool, and still pass the process checks |
Does limiting AI treat university students like suspects?
Not when the limits are stated and the same for everyone. University students are adults, and the respectful version of an AI policy tells them what is allowed, what is recorded and why, then trusts them inside those lines.
Detection does the opposite. It says nothing in advance, examines the work in secret and reverses the burden of proof onto the student, which is why even a 1% error rate is hard to accept there.
Who holds the record matters as much as what it contains. On ibl.ai, you own all the code and the data, so the conversations students have while learning sit on the university's systems under its own retention rules, not in a vendor's cloud.
Why do students cheat less in courses they care about?
Because cheating is a response to an assignment that feels pointless. Of MIT Sloan's four recommendations, the third is to foster intrinsic motivation: work that is relevant, meaningful to students and done in stages with room to revise.
An agent helps here in a plain way. It can restate why a module matters to the goal the student enrolled for, and turn a large assignment into the next ten minutes of work. A student who can see the next step has less reason to skip all of them.
This is the part no software settles. A course that teaches nothing a student wants will be gamed with or without AI, and the honest response to widespread cheating in one course is to look at the course.
What does a university need to run its own student AI workspace?
Three things that a per-seat chatbot subscription does not give it. The first is ownership: the source code and the conversation data on the university's own infrastructure, so integrity evidence never depends on a vendor's export.
The second is model choice. ibl.ai is model-agnostic, so each agent runs on the model the university picks and can be switched without rebuilding the agent.
The third is a price that lets every student in. Per-seat pricing bills every enrolled student whether or not they open the tool, which is the wrong shape at campus scale. ibl.ai has no per-seat pricing; cost follows use.
The platform already serves 1.6M+ users across 400+ organizations. The agents, settings and deployment options for universities are on the higher education page, and our team builds the first agents alongside yours.