The mechanism is simpler and more total than prompt-assisted cheating
A New York Times investigation, reported in August, found a meaningful shift in how students are using AI to get through coursework. Rather than asking a chatbot to help draft an essay or explain a concept, some students are directing AI agents to log into learning management systems like Canvas and Blackboard directly, where the agent takes quizzes, watches assigned lecture videos, writes and submits papers, and posts in discussion boards, all while the student does something else entirely.
That distinction matters for anyone building or operating LMS infrastructure. Prompt-assisted cheating leaves a human in the workflow at every step, producing a paper trail of drafts, prompts, and edits that at least theoretically could be audited. An agent completing an entire course end to end removes the human from the loop almost completely, which means the usual institutional response, catching a suspiciously fluent paragraph, does not apply to a course where the student never touched most of the material in the first place.
The chatbots themselves are not the line of defense
The Times tested three major chatbots, ChatGPT, Gemini, and Grammarly, by asking each to write an entire paper for a student, and none of the three refused. That finding should settle any assumption that model-level guardrails are quietly preventing this behavior at scale. The capability is available by default, not behind some jailbreak or workaround, which means the only real friction standing between a student and a fully agent-completed course is whether the student bothers to set it up.
Perplexity, which makes an AI browser capable of this kind of agentic task, went further and defended the practice directly, telling the Times that requiring its agents to identify themselves on learning platforms would put the student's privacy and security at risk. That is a genuinely defensible privacy argument taken in isolation, and also a direct refusal to build the one feature, agent self-identification, that would give institutions a technical way to detect what is happening on their own platforms.
Institutions have no consistent response, and professors know it
With no technical signal to rely on and no company cooperation forthcoming, enforcement has fallen to individual instructors, and the Times found that response varying wildly: some professors ignore the problem entirely, others impose harsh penalties on students they suspect, and most fall somewhere in between with no institutional standard to guide them. Karen Costa, a two-decade teaching veteran quoted in the piece, summarized the resulting uncertainty bluntly, asking whether she is just interacting with bots at this point.
Carol Sewell of the California State University system added a detail that should concern any enterprise edtech buyer: she said she receives no institutional guidance on discouraging AI use, only training on how to use it, even as CSU has a multimillion-dollar partnership with OpenAI. That is not a contradiction unique to CSU. It reflects a structural tension every large institution now faces between the AI partnerships their administration is signing and the classroom-level enforcement problem those same partnerships make harder to solve.
Some institutions are giving up on the unsupervised format itself
A handful of institutions have concluded that the format, not the enforcement policy, is the actual problem. The University of Chicago Law School banned phones, tablets, and laptops in its entry-level classes as part of what the school calls its AI strategy, an acknowledgment that device-level restriction is more enforceable than behavioral policy. Princeton went further, ending a century-old honor code tradition of unsupervised exams after an AI cheating scandal, which is a direct admission that the trust model underlying unsupervised academic work no longer holds.
Both responses are notable because they abandon the premise that drove a decade of edtech investment: that coursework and assessment could be delivered asynchronously and online without meaningfully degrading academic integrity. If more institutions follow Chicago and Princeton back toward supervised, device-restricted formats, that is a direct headwind for any platform or vendor whose value proposition depends on unsupervised, flexible delivery remaining the default, and a reason for those vendors to start building supervised-assessment features now rather than treating this as a fringe reaction confined to two elite institutions.
What this means for enterprise learning platforms and corporate training
This problem is not confined to universities. Any enterprise running compliance training, certification programs, or corporate L&D through an LMS built on the same underlying platforms, or the same assumption that a login equals a human doing the work, is exposed to exactly the same gap. A compliance certification that an agent can complete end to end on an employee's behalf is not a theoretical risk; it is the same capability the Times documented, pointed at a different kind of coursework with real regulatory consequences attached, from safety training at an industrial employer to anti-money-laundering certification at a bank.
The practical move for any learning platform owner, in higher education or the enterprise, is to stop waiting for AI vendors to add agent-identification features voluntarily, since Perplexity's response suggests that cooperation is not coming soon, and instead invest directly in behavioral and timing-based detection signals within your own LMS: unusual completion speed, interaction patterns inconsistent with human pacing, and identity verification at assessment checkpoints rather than only at login. The platforms that build this detection layer themselves will be the ones still able to claim their completion data means anything in a year, and the ones best positioned to sell that assurance back to institutions and corporate buyers as a feature rather than scramble to add it once a compliance scandal forces the issue.
