What a Professor's Fake-Essay Experiment Revealed
A college writing professor and her institution's first AI faculty fellow set out to test the tools being marketed to students on TikTok and Instagram. She asked ChatGPT to complete a final research-paper assignment as a typical freshman, ran it through Undetectable AI to strip out AI-detector flags, then used an auto-typer called Dripwriter to paste the fraudulent essay into Google Docs. The software spent three hours and 31 minutes 'writing' the paper, pausing between paragraphs, making errors, circling back to correct them, and taking stretches that looked like natural breaks. When she opened the version history, it showed entry after entry left by someone who was never there.
The experiment did not prove that students can always evade detection, but it exposed a serious vulnerability. ChatGPT's initial draft was flagged by both AI detectors her institution subscribes to; after the humanizer, neither detector flagged the prose. The auto-typer then counterfeited the timestamped editing record many faculty adopted as a defense against AI plagiarism. The professor concluded that version history had always been evidence of typing, not authorship—and that faking typing is now a product line.
The broader market shows why this is not a fringe problem. An investigation by CalMatters and The Markup documented more than $15 million in Turnitin purchases across 57 California institutions alone. On the other side, Undetectable AI reports more than 18 million users, and its payment processor Stripe publicizes a 48 percent jump in gross profit. Some companies operate on both sides: ZeroGPT offers a free detector and a humanizer on the same website, while Quillbot sells a detector but warns users not to rely on AI detection alone for decisions affecting academic standing.
A 2023 study in the International Journal for Educational Integrity tested 14 detection tools and found that none correctly classified every AI-generated text after manual editing or machine paraphrasing. For the author, the lesson is not better surveillance but a return to observing student learning over time.
The Two-Sided Market Behind AI Detection and Evasion
At the center of the AI-integrity crisis is a two-sided market: colleges buy detection from Turnitin and similar vendors, while students buy evasion from humanizers and auto-typers—sometimes from the same commercial ecosystem. This is not a secret workaround; it is a growing product category with recurring demand on both sides.
Turnitin, Undetectable AI and the Conflict Inside AI-Integrity Products
ZeroGPT sells both a free detector and a humanizer promising text that scores 'as human' across detectors. Quillbot sells a detector while warning institutions not to rely on its score for high-stakes academic decisions. The implication is uncomfortable: detection vendors have incentives to keep selling certainty even when their own caveats and published research undercut it. The 2023 study provides the technical basis—no detector among 14 reliably identified AI text after paraphrasing or manual editing.
The Version-History Fallacy
The professor's experiment shows that auto-typers can fabricate the exact documentation trail faculty learned to trust: a Google Docs version history that looks like sustained editing over hours. Her result is anecdotal, not a controlled study, but it demonstrates how cheaply process documentation can be counterfeited. The mistake was treating evidence of typing as evidence of authorship.
What This Means for Academic Integrity
If detector scores and version histories are both unreliable on their own, the remaining evidence of learning is the trajectory a teacher observes across multiple low-stakes interactions: drafts, conferences, reflections, and revisions made in response to feedback. That shifts the burden from policing to course design—though it raises hard workload questions for large courses and contingent faculty.
What Faculty and Institutions Can Do Before Fall
For faculty and academic leaders, the most concrete response is to stop treating any single forensic signal as proof of student work.
- Do not base an integrity charge on Google Docs version history or detector scores alone. The author's experiment shows an auto-typer can fake a 3-hour, 31-minute editing trail, and Quillbot itself warns against using its detector alone for decisions that affect academic standing.
- Redesign courses around process rather than product. Build a chain of drafts, conferences, reflections, and revisions observed over the semester—the author's surviving model—since an auto-typer can imitate typing but not the uneven trajectory of becoming a writer.
- Talk to students about the tools being marketed to them. Name Dripwriter, Undetectable AI, and similar products explicitly, and address the pressures that make outsourcing assignments feel normal rather than assuming detection will catch them.
- Audit institutional spending and vendor claims. Review detection contracts using data like the CalMatters/Markup finding of more than $15 million in Turnitin purchases across 57 California institutions, and ask vendors directly whether they or affiliates also sell evasion tools.
- Press edtech providers to separate detection from evasion. ZeroGPT's free detector and humanizer on one site, and Quillbot's detector that warns against over-reliance, create a conflict institutions should raise before renewing contracts.
Risk & Opportunity Assessment
| Commercial Risk | Medium | Faculty and institutions are reconsidering detection spending after the professor's experiment and the 2023 study; if process-based assessment replaces forensic surveillance, detection contracts like the $15 million California Turnitin spending could face renewal pressure. |
| Competitive Risk | High | Evasion tools erode the value proposition of detection products. Undetectable AI's reported 18 million users and Stripe-publicized 48 percent gross-profit jump show demand shifting toward evasion, while ZeroGPT and Quillbot blur the line between detection and evasion. |
| Regulatory Risk | Medium | No specific legislation is cited, but institutional academic-integrity policies will need revision. Penalizing students based on detector scores or version histories that vendors themselves warn against creates due-process and appeal risk. |
| Reputation Risk | High | Vendors caught selling both detection and evasion lose credibility, and institutions risk reputational damage if integrity charges rely on tools whose own makers say the scores cannot carry that weight. |
| Technology Disruption | Transformational | Auto-typers and humanizers represent a new product category that counterfeits the process evidence faculty adopted after detectors failed, undercutting the version-history defense across the sector. |
| Commercial Opportunity | Medium | Institutions may shift budgets from pure detection toward assessment design and longitudinal learning tools. Vendors that support process-based evaluation could capture that demand, though the article does not identify a clear current leader. |
Comments 0