Stop catching AI cheaters. Start testing what students know.

The trap worked—and proved too little

A history professor hid an absurd instruction inside a midterm prompt: place the word ‘Madagascar’ somewhere in the answer. Thirty-two of 35 submissions reportedly followed it. As a demonstration of students pasting an entire assignment into a chatbot, the trap was brutally effective. As a model for education, it is a dead end.

One student successfully challenged the result after explaining that dark mode made the hidden text visible. That correction matters. Secret tripwires can catch misconduct, but they can also turn grading into an adversarial contest in which instructors design traps and students learn to evade them. Neither side necessarily gets better at history, writing, judgment, or responsible AI use.

The labor market is already rewriting the curriculum

The same 24-hour news cycle delivered the other half of the problem. Enrollment in computer and information sciences at four-year institutions fell more than eight percent from spring 2025, according to data cited by the Associated Press, while students in psychology, music, and other fields are crowding into AI courses. Entry-level coding work is under pressure, but AI fluency is spreading across almost every discipline.

Universities therefore cannot solve AI cheating by pretending the technology is outside the classroom. Employers are asking graduates to use it. The educational task is to separate assisted performance from genuine competence—to teach students when a model adds leverage, when it hides uncertainty, and when delegating the work also delegates the learning.

A credential is a promise about a person

A degree is not supposed to certify that a polished document appeared in a learning-management system before midnight. It is supposed to certify that a person can reason, retrieve knowledge, test evidence, explain decisions, and transfer those abilities to a new situation. Generative AI did not destroy that promise. It exposed how often institutions were verifying the document instead of the person.

If a chatbot can complete an assessment invisibly, the answer cannot be an endless escalation of detectors and hidden commands. The answer is to collect better evidence of learning. That is harder than buying surveillance software, but it is also the work students are paying colleges to do.

Make the thinking visible

Every high-stakes assignment should include at least one moment when the student must make the reasoning their own: a short oral defense, an in-class transfer problem, a critique of the model’s output, a version history, a source audit, or a live explanation of why one approach beat another. AI use can be allowed, limited, or prohibited according to the learning objective—but it should always be disclosed.

This does not require turning every course into an interrogation. Faculty can randomly sample oral defenses, require concise process notes, or ask students to revise an unfamiliar passage under observation. The point is not to catch more people. It is to make outsourcing the entire cognitive task less useful than learning how to do it.

• Grade the decision trail, not only the final artifact.
• Require students to identify and correct at least one model failure.
• Use oral or in-person verification for high-stakes claims of competence.
• Publish a clear AI-use rule for every assignment instead of relying on guesswork.

The real scandal is unverifiable learning

The viral hidden-word trap is satisfying because it produces a clean reveal: gotcha. But higher education should resist designing itself around the emotional reward of catching people. The more important question is whether an honest student can leave a course having submitted every assignment and still be unable to explain the underlying ideas without a machine.

That is the standard universities should fear. Stop treating AI detection as the finish line. Build assessments in which students must show their judgment, defend their evidence, and own the consequences of the answer. The future belongs neither to students who avoid AI nor to students who surrender to it. It belongs to people who can use the tool without disappearing inside it.