College Students Build Custom AI Bots to Evade Homework Detection

Students engineer around detection systems

At New York University, an economics sophomore named Jared learned that a classmate had built a custom bot using AI coding tools to complete calculus homework on WebAssign, a platform designed to prevent cheating. The bot mimics a struggling student's pace to avoid suspicion from time-stamp analysis. That bot has since circulated among multiple students.

This represents a fundamental shift in academic dishonesty. Rather than simply pasting ChatGPT output into assignment boxes—a method students now describe as outdated—undergraduates are deploying agentic AI systems that can plan multi-step tasks, execute code, and complete complex assignments with minimal oversight. These newer models became publicly available in spring 2025, creating capabilities that didn't exist a year earlier.

Why it matters

The arms race between student cheating and institutional detection reveals a deeper crisis in higher education. Universities are investing in surveillance infrastructure and AI detection tools while students spend hours deliberately degrading AI output to appear authentically mediocre. This dynamic suggests that technical solutions alone cannot address what is fundamentally a pedagogical problem about the perceived value of coursework.

Detection tools struggle to keep pace

When Harvard's student newspaper surveyed undergraduates this spring, nearly 40 percent admitted using AI in ways instructors would consider inappropriate. Universities have responded with expanded technology contracts—California State University schools paid additional fees for Turnitin's AI detection feature, while other institutions adopted tools like GPTZero, which claims 99 percent accuracy in identifying AI-generated text.

Yet students report these measures are easily circumvented. Will, a recent Carleton College graduate, described feeding course readings to Claude, having the model draft challenging essay sections with proper citations, then rewriting each AI-generated sentence to evade detection. A UMass Amherst sophomore named Theo used OpenAI's Codex agent enhanced with web-search capabilities to create a presentation, then spent four hours deliberately making it look worse—"restoring the mediocrity of a real college sophomore," as he put it. He received an A.

Surveillance becomes the new enforcement method

Some institutions have turned to network monitoring. At Princeton, according to a junior familiar with the school's Honor Committee, investigators pull browsing histories from campus networks to check for ChatGPT or Claude visits during exams. While the university states it only accesses search histories for students already under investigation, these network logs have become critical evidence in academic dishonesty cases.

Max Spero, CEO of detection firm Pangram Labs, believes improved detection technology could eventually make cheating more difficult than completing assignments honestly. But enforcement remains inconsistent, varying widely by professor and their familiarity with generative AI.

Low-tech alternatives emerge

Some faculty have abandoned the technological arms race entirely. Danielle Carr, a UCLA historian and anthropologist, primarily assigns writing that students complete under supervision in what she calls "reading jail." She views AI cheating as a pedagogical failure rather than a technical problem requiring surveillance solutions.

The students interviewed insisted they still value learning—they simply don't use AI in courses they care about. Their selective deployment of these tools suggests the real challenge for universities is demonstrating that coursework provides irreplaceable value that AI cannot replicate.

These details were first reported by New York Magazine's Intelligencer.