Academic peer review has always run on a quiet assumption: that the person whose name sits atop a paper actually did the thinking. That assumption is now under open challenge. At a major research venue, track chairs running submissions through an AI-text detector reported that a startling share scored as fully automated — and the response, a wave of desk-rejections and demands for proof of human authorship, reads less like a routine integrity sweep than a system improvising under pressure. It is a vivid early warning about what happens when the tools that generate text outrun the systems built to certify knowledge.
What happened
According to a June 2026 AI news roundup from devFlokers — a figure worth verifying against the venue’s own primary statement — track chairs found that roughly 28.2% of submissions originally scored as fully automated under standard detector analysis. That represents about a tenfold increase in highly automated submissions since 2025. In other words, the share of papers reading as machine-written did not creep up; it exploded in the span of two conference cycles.
The response was correspondingly drastic. The roundup reports 178 papers desk-rejected outright and 123 conditionally rejected, with affected authors asked to do something peer review has never historically required: submit version-history audit trails — the editing breadcrumbs from their drafting tools — to demonstrate that a human actually wrote the work. The burden of proof, in effect, flipped. For decades, the default was to assume good faith and scrutinise the science. Here, authorship itself became the thing on trial, and the accused were told to produce a paper trail or be presumed guilty.
That inversion is the real story. A 28% flag rate is shocking on its face, but the demand for audit trails is what signals a structural break. It concedes that the finished artefact — the polished PDF — can no longer be read as evidence of the labour behind it.
Why it matters beyond academia
Conferences and journals are credentialing machines. A peer-reviewed acceptance is a certificate that says: experts checked this, and it holds. When more than a quarter of submissions trip an automation flag, that certificate’s value wobbles — not because the underlying science is necessarily fake, but because the venue can no longer cheaply tell the difference between a researcher who used AI to polish prose and one who outsourced the intellectual work entirely.
And here the methodology cuts both ways. AI-text detectors are notoriously unreliable. They produce false positives, and they penalise predictable populations: non-native English writers, whose careful, formulaic phrasing reads as ‘machine-like’; researchers who lean on grammar tools or translation; and anyone whose plain, structured style happens to resemble a model’s house voice. A 28.2% flag rate is not the same as 28.2% fraud. Some unknown slice of those desk-rejected papers may be honest work caught by a blunt instrument. That is precisely why a detector score alone should never be a verdict — and why the venue’s pivot to audit trails, however clumsy, is the more defensible instinct.
That instinct points to the broader shift now underway. As the same industry reporting notes, the centre of gravity is moving away from after-the-fact AI detection and toward provenance: content credentials, signed edit histories, and audit trails that travel with a document and attest to how it was made. Detection asks an unanswerable question — ‘does this look AI-written?’ Provenance asks a tractable one — ‘can you show how this came to exist?’ The first is a guessing game that punishes the innocent alongside the guilty. The second is an accountability record. The whole episode is best understood as the moment the certification economy started reaching for the second model because the first one broke.
The knock-on effects
What is happening to peer review is a preview, not an exception. Every system that certifies a person through their output is exposed to the same strain.
- Hiring and portfolios. The cover letter, the take-home assignment, the writing sample, the GitHub portfolio — these were proxies for capability. When anyone can generate a fluent, plausible artefact in seconds, the proxy decouples from the person. Recruiters who lean on AI-detection tools to screen candidates will replicate the conference’s false-positive problem at scale, quietly filtering out competent applicants whose prose reads ‘too clean.’
- Credentials and assessment. Take-home exams, essays and certifications all assumed a friction that no longer exists. Institutions face a choice between performative crackdowns — unreliable detectors, suspicion by default — and harder structural fixes: oral defences, in-person components, and process-based evidence rather than artefact-based evidence.
- Disclosure norms and policy. The unresolved question is what honest AI use even looks like. Using a model to fix grammar is not the same as having it write your literature review, yet most venues and employers lack norms granular enough to distinguish them. Clear disclosure standards — what assistance is allowed, what must be declared — are now overdue, and the absence of them is what forces the crude binary of ‘flagged or clean.’
- Trust as the scarce resource. The throughline is that trust, not content, is becoming the bottleneck. Text is now infinite and nearly free; verified human authorship is finite and increasingly valuable. Whoever can credibly prove provenance — a signed edit history, a defensible process, a track record — holds the scarce asset.
This is the uncomfortable upshot: in a world of infinite plausible output, the artefact is cheap and the attestation is expensive. Systems that fail to make that switch will keep mistaking fluency for competence in both directions — accepting confident fakes and rejecting honest work.
The India read
For Indian academia and assessment, the stakes are sharpened by scale and by stakes. The country runs some of the world’s largest, highest-pressure examination and credentialing systems, where a marginal edge can reshape a life — and where the temptation to use AI, and the incentive to police it, are both intense. Indian universities and conferences are not insulated from the dynamics that produced the 28% flag rate; if anything, the volume of submissions and the prevalence of English-as-a-second-language writing make blunt detector-based crackdowns especially dangerous here, because the false-positive burden lands hardest on exactly the students this system should not be filtering out.
The constructive path is to skip the detector arms race and build authorship and provenance norms directly. A few things students and institutions can adopt now:
- Treat process as the evidence. Encourage drafting in tools that preserve version history, and normalise submitting that history alongside major work. The point is not surveillance; it is giving honest authors a way to prove their case before they are ever accused.
- Write disclosure rules that distinguish degrees of help. A clear policy that permits grammar and translation assistance while requiring declaration of substantive generation protects students from arbitrary judgement and gives institutions something enforceable.
- Shift assessment toward defensible formats. Vivas, in-class components, oral defences and iterative supervision reward understanding that cannot be outsourced — and they sidestep the unwinnable battle of inspecting a finished document for its ‘AI-ness.’
- Refuse detector scores as verdicts. Indian institutions should treat any AI-detection figure as a flag for human review, never as proof. The conference episode is a cautionary tale precisely because a probabilistic tool was close to functioning as a tribunal.
The venue that flagged a quarter of its papers did not solve the problem; it exposed it. The lesson travelling outward — to hiring desks, examination boards and publishers everywhere — is that we can no longer certify people by inspecting their output. We will have to certify them by recording their process. Provenance is not a perfect answer, and it carries its own privacy and access tradeoffs. But it is the honest one, and the institutions that build it first will be the ones still worth trusting when everything else can be generated on demand.
