TL;DR: The attempt to police generative AI in coursework by detecting it after the fact has largely failed. Student AI use is now close to universal, detectors produce false positives that fall hardest on non-native English writers, and a growing list of universities has switched them off. What is replacing detection is not a better detector — it is assessment redesign, evidence of the writing process, and more in-person and oral verification. For edtech platforms, that shift changes what institutions will actually pay for.

How the arms race ended

Three facts closed the debate for most institutions.

AI use became the norm. The Higher Education Policy Institute's Student Generative AI Survey 2025 found that 92% of UK undergraduates used AI in some form, up from 66% a year earlier, and 88% had used generative AI for assessments, up from 53%. Around 18% said they had included AI-generated text directly in their work. When nearly every student uses a tool, a policy built on catching its use stops being enforceable.

Detectors proved unreliable in the ways that matter most. A Stanford study led by Weixin Liang and published in Patterns tested seven GPT detectors on TOEFL essays written by non-native English speakers and found an average false-positive rate of 61.3%, against near-zero on essays by native-speaking US students. The same research showed that simple prompt-based rewriting could push AI-generated text under detection thresholds. A tool that both over-flags honest students from one group and is easy to evade is a poor basis for misconduct findings.

Institutions started switching them off. The University of Waterloo discontinued Turnitin's AI detection feature in September 2025, citing bias against students whose first language is not English. Curtin University in Australia announced it would disable the feature across all campuses from January 2026. Trackers compiled in 2026 count more than 50 universities across the US, Canada, the UK, Australia and South Africa that have disabled or formally discouraged AI detection. These lists are unofficial and vary in rigor, but the direction is clear.

What is replacing detection

1. Assessment redesign: the "two-lane" model

The most influential institutional response has been the University of Sydney's two-lane assessment framework, rolled out from 2025. Lane 1 covers secure, supervised assessments — exams, in-person tasks, vivas — that verify what an individual student can do. Lane 2 covers open assessments where AI use is permitted and often scaffolded, and the task is designed around authentic, real-world work. Sydney's own review reportedly found that around 90% of its existing assessments were vulnerable to AI generation — which is why the answer was redesign, not enforcement. Faculties are now completing degree-level assessment plans to guide redesign through 2026.

Versions of this idea are spreading: fewer high-stakes take-home essays, more staged assignments, more in-class writing, and explicit rules about where AI is allowed.

2. Process evidence instead of product analysis

Rather than guessing whether a finished essay was machine-written, tools now capture how it was written. Turnitin launched Clarity in 2025 as a paid add-on — a composition workspace that records version history, pasted text, active writing time and revision patterns. Grammarly Authorship takes a similar approach inside Word and Google Docs. The shift matters: a keystroke-level timeline is evidence an instructor can discuss with a student, whereas a detector's probability score is not.

3. Oral and conversational verification

Short vivas, presentations and "explain your submission" conversations are returning, because they are hard to outsource. The constraint is instructor time — which is exactly where software can help, through scheduling, structured question banks and recorded mini-orals.

4. Teaching AI use rather than banning it

HEPI's data shows students mostly use AI to explain concepts, summarize readings and generate ideas. Institutions are increasingly writing policies that define acceptable use per assignment and require students to disclose and reflect on it.

Where the new approaches are still immature or overhyped

Process tracking is not a lie detector. A student can type out AI-generated text by hand, or draft in another tool and paste in pieces. Process data raises the cost of misconduct and produces better conversations; it does not produce proof. Vendors that market it as "verification" are overselling.

Surveillance creep is a real risk. Recording every keystroke and editing session is a significant data collection exercise. Under FERPA in the US, GDPR in Europe and India's DPDP Act, institutions need a clear purpose, retention limits and transparency with students. Student data privacy concerns — and faculty unease — can stall adoption.

Redesign is labor, not software. The two-lane model depends on academic staff rethinking assessments course by course. No platform does that for them, and many institutions underestimate the workload.

Secure assessment has scale limits. In-person exams and orals do not scale cheaply to large online or distance-learning cohorts — the very market many edtech platforms serve.

"AI-proof" assignments rarely are. Tasks built around personal reflection or local context were once considered safe; current models handle them well. Durable designs focus on observed performance and iteration, not on topics AI supposedly can't handle.

A practical sequence for the next academic year

For institutions still relying on detection, a realistic transition looks like this. First, stop using detector scores as standalone evidence and update misconduct procedures accordingly. Second, audit high-stakes take-home assessments against current AI models and triage them: redesign, move to a secure setting, or convert to an open task with explicit AI rules. Third, pilot process-capture and staged submissions in a handful of large courses, with a published data-retention policy, and measure staff workload before scaling. Fourth, invest in academic development — the redesign work lives with faculty, and they need time and examples. Institutions that skip the first step tend to keep generating disputes; those that skip the last tend to end up with policies nobody applies consistently.

What this means for edtech platforms

The market is shifting from "catch the cheater" to "evidence the learning." That creates concrete product opportunities:

  • Staged submission workflows — outlines, drafts, peer review and final submission as linked artifacts, with feedback at each stage.
  • AI-use disclosure built into assignments — per-task policy settings, student declarations and reflection prompts captured alongside the work.
  • Assessment-plan tooling — mapping courses and assessments to secure and open lanes and to learning outcomes, so program leaders can see coverage.
  • Oral assessment support — scheduling, randomized question generation from a student's own submission, and rubric-based marking.
  • Sanctioned AI assistance — tutors that work inside the course's rules, cite course material, and log their interactions for transparency. Grounding matters: AI tutors that hallucinate sources create a new integrity problem instead of solving one.

Platforms that still lead with detection scores are selling into a shrinking budget line and carry reputational risk every time a false accusation makes the news.

How to evaluate whether your institution or platform is ready

  1. Do you know which of your assessments are vulnerable? An audit against current models is the starting point; most institutions that run one are surprised.
  2. Is your AI policy set at the assignment level, not just the institution level? Blanket rules don't survive contact with real courses.
  3. Can your LMS or platform capture drafts and process data — and do you have a clear, published purpose and retention period for it?
  4. Do misconduct procedures rely on detector scores? If so, review them; a probability score alone is weak evidence and a legal and reputational liability.
  5. Do you have capacity for secure assessment at your scale? If not, what staged or oral alternatives are feasible?
  6. For edtech vendors: can your product show learning progress, not just final outputs? That is what buyers will ask for next.

Sources

  • Higher Education Policy Institute, Student Generative AI Survey 2025 (hepi.ac.uk)
  • Liang, W. et al., "GPT detectors are biased against non-native English writers," Patterns, 2023 (arxiv.org/abs/2304.02819)
  • University of Waterloo, "Discontinuing use of AI detection functionality in Turnitin," September 2025 (uwaterloo.ca)
  • EdTech Innovation Hub, "Curtin University to disable Turnitin AI detection tool in 2026"
  • University of Sydney, Teaching@Sydney, two-lane approach to assessment and FAQ (educational-innovation.sydney.edu.au)
  • Turnitin, "Turnitin Clarity: Bringing transparency to the student writing process" (turnitin.com)

Conclusion: where Syslabs fits

Syslabs builds custom software and SaaS product features for education providers and edtech platforms — staged submission workflows, LMS integrations, assessment-planning tools and AI tutoring features grounded in course content, designed with student data privacy built in rather than bolted on. The goal is simple: help institutions evidence real learning in a world where AI is part of how students work.