The honor code is dead, and vibes won't bring it back.
The Summary
- Harvard professors are abandoning formal AI detection because there's no reliable way to prove students used ChatGPT, so some are threatening to fail work that "feels like AI"
- Students have already figured out how to bypass technical countermeasures like hidden watermark text meant to flag LLM output
- The tell-tale signs people think reveal AI (em dashes, "on the one hand" hedging) are trivially easy to prompt away
- What worked as a social norm for academic integrity doesn't scale when the cost of cheating drops to zero and detection is impossible
The Signal
Harvard's AI crisis isn't about cheating. It's about the collapse of verification in knowledge work. When one professor's syllabus threatens to fail work based on vibes rather than proof, you're watching an institution admit it has no tools left. The honor council stopped taking cases because proof is impossible. Students submit Google Doc version histories, which proves only that they opened Google Docs.
This is the watermark problem at institutional scale. Every technical countermeasure gets broken within a semester. Hidden text meant to trap AI? Students learned to strip it. Stylistic fingerprinting? You can prompt ChatGPT to write like a sophomore with ADHD who's obsessed with Succession. The tells people think they see (excessive hedging, unnatural evenhandedness, those damn em dashes) are just artifacts of lazy prompting, not inherent LLM traits.
"If your AI-dar relies on spotting 'on the one hand, on the other hand' construction, you're already two model versions behind."
The deeper problem: this isn't fixable by better detection. It's a category error. Academic integrity was built on high friction: cheating was expensive, risky, and obvious. Buying a paper from a essay mill left a paper trail. Copying Wikipedia was easy to spot. The social cost kept most people honest most of the time.
LLMs collapsed all three pillars at once:
- Cost: Free or $20/month for unlimited writing assistance
- Risk: Undetectable when done with any competence
- Obviousness: The output quality ranges from "pretty good" to "better than the median student"
What Harvard is discovering, every knowledge institution will discover: you can't police this. The detection arms race is over before it started. Not because the technology is too good, but because the prior equilibrium depended on friction that no longer exists.
So what happens next? Some professors will double down on vibes-based grading, which is just institutionalized bias with extra steps. Others will require all work to happen in proctored environments, which doesn't scale and doesn't reflect how actual knowledge work happens. The smart ones will redesign assessment entirely, away from "produce this artifact" and toward "demonstrate this capability live."
The Implication
If you're building tools for education, hiring, or any other domain where proving someone did the work matters, understand this: detection is not your moat. The Harvard problem is coming for every credentialing system that relies on unsupervised artifact production. Degrees, certifications, portfolio work, code samples. All of it.
The opportunity isn't in better AI detection. It's in systems that assume AI assistance as default and measure what actually matters: taste, judgment, the ability to ask good questions, the skill to evaluate outputs. What you can do with the tools matters more than whether you used them. Harvard's crisis is everyone's crisis. Design accordingly.