When an economics professor at Brown University sees his take-home midterm average soar to 96 out of 100 and then crash to 48 on the in-person final, he no longer has any doubt: the suspicion of AI cheating becomes statistical certainty. Roberto Serrano went public with his experience, forcing academia to answer a question that goes beyond academic integrity: what does it mean to trust an assessment process when everyone carries in their pocket an LLM capable of solving exercises without leaving a trace?

The data is sparse but devastating: same course, same topics, yet two incompatible grade distributions. No detective work is needed to guess that students, left free to use any resource at home, tapped ChatGPT or similar tools extensively. Serrano didn’t run a forensic investigation: he simply looked at the numbers, and those numbers are an unmistakable signature of undisclosed automation.

Who wins and who loses in the era of remote assessment

The first casualty is trust. Universities that invested in online exam platforms and automated proctoring now face a dilemma: every unsupervised test is potentially compromised. But the real loser is the cloud-first AI model: the ability to access powerful models via API has made cheating too cheap, too easy, too widespread to be stopped by institutional policies. Students don’t need to install anything, don’t need to configure hardware: a browser and a connection suffice. The barrier to entry is zero.

The winners, paradoxically, will be those institutions that manage to bring control back where it can actually be exercised: not on the student’s device, but on an execution environment that belongs to the examiner. This shifts attention toward assessment solutions running on dedicated infrastructure, possibly on-premise, where test integrity can be guaranteed by a physical perimeter and software that the institution itself manages. It’s no coincidence that some public competitions and professional certifications are already returning to in-person exams or air-gapped labs, precisely to remove the candidate’s control over the software stack.

The structural signal for trust infrastructure builders

The Brown case illuminates a tender nerve in AI deployment for high-stakes contexts. When assessment determines careers, professional access, or funding, offloading inference to third-party cloud services means accepting an information asymmetry: the respondent knows they can cheat, the evaluator cannot know whether the text was generated or not. The only way out is to shorten the chain of trust, bringing the evaluation infrastructure under the certifying body’s direct control.

This doesn’t necessarily imply a return to paper and pencil, but rather a redefinition of what a “controlled environment” means. Imagine an exam station where the only interface is a locked-down terminal running a local evaluation model, capable of proposing questions, collecting answers, and, why not, exploiting an LLM for consistent grading—but without the student being able to interact with it. That would be a self-hosted scenario by definition, with all the VRAM, latency, and TCO constraints we’re familiar with.

Those designing such systems face real trade-offs: an LLM that evaluates open-ended answers requires non-trivial computational power, and centralizing hundreds of simultaneous sessions locally demands GPU investments that many universities aren’t used to considering. Yet the cost of a semester of cheating is not zero: it undermines reputation, devalues degrees, and over time erodes the academic brand’s value. The TCO calculation expands: no longer just hardware cost, but the cost of lost credibility.

Beyond witch hunts: a project for trustworthy AI

Serrano chose not to accuse individual students, but to publicly show the statistical evidence. It’s a stance that teaches us something: a perfect detector isn’t needed when the result distribution reveals the anomaly. But to design future exams that are LLM-proof, it won’t be enough to go back. What’s needed is an architecture that treats the assessment process as a data integrity problem, exactly as banks treat financial transactions. And in those environments, local inference and self-hosting aren’t paranoia—they are audit prerequisites.

The Brown lesson is that “we cannot choose to become idiots,” as the professor warns. But we can choose where to run the intelligence that evaluates us. And that choice, more and more often, will be dictated not by the cloud’s convenience, but by the need for real, measurable, verifiable control. For those evaluating on-premise deployment, AI-RADAR offers analytical frameworks to weigh these trade-offs without shortcuts.