What Earlier Anti-Cheating Arms Races Teach Us

Typing pattern analysis feels like a new problem, but the underlying shape of it is not new at all. Academic institutions have already run this exact cycle twice before, once with text-matching software, and once with online exam proctoring, and both cycles left behind lessons worth paying attention to before the same mistakes repeat a third time.

Looking closely at how those two earlier waves actually played out, not just how they started, gives a much clearer picture of where the current wave is likely headed.

Three Waves, One Pattern

Researcher Sarah Elaine Eaton, who has written extensively on academic integrity, frames the last two decades as three successive waves of technology aimed at the same underlying problem: text-matching software first, then online exam proctoring, and now artificial intelligence and large language models. Her central argument is that none of these waves produced a silver bullet, and that treating any of them as a final solution has consistently backfired.

Typing pattern analysis and the auto typer tools it responds to fit cleanly as a fourth wave in that same pattern, not a departure from it. The specific technology changes. The underlying dynamic, a detection method rewarded until someone builds something that defeats it, does not.

What makes Eaton’s framing useful is that it treats this as a structural feature of the problem rather than a failure of any particular tool. Each wave arrived promising to finally solve academic dishonesty. Each one instead became one input among several, absorbed into a messier, more human process rather than replacing it.

What CAPTCHA Already Showed About This Cycle

The CAPTCHA system, introduced in 2000 to distinguish humans from bots, offers one of the clearest early examples of this exact arms race outside academia. The original version simply asked users to type a word displayed as an image. As automated tools got better at reading distorted text, CAPTCHA mutated, warping backgrounds, distorting letters further, and eventually moving toward entirely different verification approaches.

The relevant lesson is not that CAPTCHA failed. It is that no single version of it ever stayed effective indefinitely, and the system’s designers accepted from early on that staying useful meant continuously adapting rather than declaring victory with any one version. Typing pattern analysis is currently in something like CAPTCHA’s early phase, a first serious attempt at a new signal, not a permanent solution.

What Online Proctoring’s Backlash Adds to the Picture

Online exam proctoring software offers a more cautionary chapter in the same story. As detection measures grew more invasive, tracking eye movement, requiring room scans, and in some cases monitoring biometric data, the backlash grew alongside it. Cleveland State University settled a lawsuit in 2022 over a proctoring tool’s room scans, and a separate proctoring provider faced multiple class-action suits over its handling of biometric data.

Regulation followed. The European Union’s AI Act specifically classifies student-monitoring AI as high-risk, a designation that took effect in August 2026 and imposes meaningfully higher compliance requirements on any tool that falls into that category. The lesson from this wave is not that monitoring technology is inherently wrong, but that invasiveness carries real legal and reputational cost once it outpaces what students and regulators are willing to accept.

AI Detection Already Showed Signs of the Same Collapse

The most recent wave, AI text detection, has already shown the early signs of the same pattern playing out again. Stanford researchers found one widely used detector scored a 26 percent false-positive rate specifically against non-native English writing, and Turnitin publicly acknowledged in 2023 that its own detector should not serve as the sole basis for an academic integrity finding. By 2025, OpenAI had discontinued its own AI text classifier entirely, citing its low accuracy.

Typing pattern analysis is, in a real sense, a response to AI detection’s known limitations, an attempt to add a second signal once the first one proved unreliable on its own. That is a reasonable engineering response. It does not mean typing pattern analysis itself is exempt from becoming, eventually, exactly as contested as the tool it was built to supplement.

It is worth noting how quickly this particular collapse happened relative to the previous two waves. Text-matching software took roughly a decade to accumulate serious criticism. Online proctoring drew lawsuits within a few years of widespread adoption. AI text detection reached the point of a major provider shutting down its own tool within roughly two years of its introduction, a pace that suggests each successive wave is being scrutinized faster than the one before it.

Applying the Pattern Instead of Repeating It

None of this argues against using detection technology. It argues against treating any single generation of it as a permanent, final answer, a mistake each previous wave made and each previous wave eventually had to walk back, often after real harm had already been done to students wrongly caught in an overconfident system.

The institutions handling the current wave most sensibly are the ones building in the lesson from the start: use the signal, expect it to have real error rates, pair it with human judgment, and stay ready to adjust the moment it starts producing the same kind of backlash the last two waves eventually did.

That readiness to adjust is itself the hardest part to institutionalize. Adopting a new detection tool is usually a single decision made once. Building in an ongoing habit of questioning that same tool a year or two later, once its weaknesses have had time to surface, requires a kind of institutional humility that none of the last three waves demonstrated particularly well until public pressure forced the issue.

Part of applying that lesson well is distinguishing legitimate writing assistance from an actual workaround. A style tool such as the AI Humanizer adjusts phrasing without touching a document’s underlying process record, which keeps it outside the pattern this history describes, unlike software built specifically to defeat the detection layer itself.

For more on how AI detection technology has evolved and where its current limits sit, further reading on the Phrasly blog covers the underlying research for institutions and writers thinking through this history.