Nearly 90% of Harvard Faculty Say They've Received AI-Written Work. Your Certificate Has the Same Problem.

The most credential-obsessed institution in America just admitted it can't reliably tell who wrote the assignment.

Nearly 90 percent of respondents to The Harvard Crimson's annual Faculty of Arts and Sciences survey said they had received student coursework that they knew or believed was produced using AI. Sixty-four percent said AI had a "somewhat negative" or "very negative" effect on their courses this year — up from 42 percent the year before. The survey was published September 14.

That is a nine-figure-endowment institution with graders, proctors, honor codes, office hours and a teaching center, and it is still losing the argument about who did the work.

Now look at your course, with its 14 videos, a workbook, a multiple-choice quiz and a PDF certificate at the end.

Detection was never the plan

The faculty aren't confident either. Sixty-four percent said they felt somewhat or very confident in their ability to tell student work from AI-produced assignments — which, read the other way, means a third of professors at Harvard are openly unsure. A year earlier, just 14 percent of FAS survey respondents said they were "very confident" in distinguishing AI from non-AI submissions, and a Penn State study found humans identify AI-generated text roughly 53 percent of the time, barely better than a coin flip.

The policy response is telling. Only 3.5 percent of respondents said they have no explicit AI policy, down from 10 percent last year, and the share who "entirely permit" AI tools was roughly halved, from 8 percent to 4 percent. Everyone is writing rules. Nobody has an enforcement mechanism.

So Harvard's teaching center stopped trying to catch people. The Bok Center now advises four strategies: move some assessments in person, pair take-home work with in-person follow-ups, scaffold large projects with checkpoints, and anchor assignments in material that is difficult for AI to reproduce. Its executive director, Rebecca Nesson, framed them plainly: "Each one is already good teaching practice that AI has made newly necessary."

Read that list again as a course creator. Live calls. Submissions reviewed by a human. Projects with stages. Assignments built on the student's own business, numbers or clients. Every one of those is something you can ship this quarter, and every one of them costs you the thing you were promised: full automation.

The platforms are already repricing completion

Here's the timing that should make anyone with a certificate nervous.

Six days before the Harvard survey ran, Coursera previewed an AI-native skills platform code-named Project Helix at its FWD customer event, pitched at helping organizations close talent gaps and translate learning spend into verified workforce capability. It's the company's first big reimagined product since combining with Udemy in May 2026. Broad enterprise availability is expected in the first half of 2027.

The sales logic is stated without euphemism on Coursera's own blog: completion and learning hours are easy to report, validated skills are not, so Helix combines hands-on graded assessments, practical observation and recognized credentials, with earned proof flowing into a portable skills record. The company also argues that enterprise learning built on catalog volume and course completions is no longer sufficient.

Strip out the enterprise vocabulary and that's a $2 billion-class education company telling its buyers that completion is no longer evidence. Coursera's CEO, Greg Hart, put the shift as moving "from standalone content to connected journeys that support skill development, application, and proof."

What a certificate certifies now

If a quiz, a workbook and a capstone can be produced in four minutes by a chatbot, then "87% completion rate" is a claim about video playback, not competence. And the certificate at the end certifies that someone's browser reached the final lesson.

That's survivable for a $47 product bought for inspiration. It is not survivable for anything sold as career leverage — the $2,000 certification, the B2B training contract, the cohort priced against a real salary bump.

The creators who come out of this ahead will be the ones whose assessment can't be faked from a phone: cameras on, work reviewed live, deliverables tied to a real account the student actually runs, checkpoints instead of one submission at the end. It is more expensive to deliver and harder to scale.

It's also the only version of a credential anyone will pay a premium for by 2027.