The most telling thing about Frontiers' new AI guidance framework, released this week, is not what it says. It's that they're publishing it in April 2026.

Frontiers has been one of the largest open-access publishers on the planet for years. They process tens of thousands of manuscripts annually. Their editors, reviewers, and authors have been using AI tools — at varying intensities, with varying levels of disclosure, and with varying degrees of awareness that they were doing something anyone cared about — for the better part of three years. And now, in spring 2026, they have released what they are calling "first-of-its-kind" guidance covering the full publication lifecycle.

I want to be clear: the guidance itself is genuinely worth something. Frontiers has moved past the blunt instrument of allowed/not-allowed to something more architecturally thoughtful — a structured framework that acknowledges AI use differs across roles and contexts, and that asks for transparency rather than prohibition. That is progress. Real progress.

But "first-of-its-kind" in April 2026 is not a boast. It is an inadvertent confession.

This is the pattern that has defined scholarly publishing's relationship with artificial intelligence since 2022: the technology arrives through the front door of researcher practice, occupies the house for two or three years, and the landlord shows up with a lease agreement sometime after the second lease period has already expired. The Knowledgespeak editorial that ran yesterday put it plainly: AI is everywhere in scholarly publishing, but not yet aligned. The framing — "arrived through use, not design" — is exactly right, and it deserves more scrutiny than it usually gets.

The Adoption Curve Nobody Planned For

Let's be precise about what "arrived through use" actually means in operational terms.

Springer Nature processed 3.1 million manuscript submissions through its Snapp platform in 2025, with AI integrity screening embedded at the workflow level. That number — 3.1 million — represents a scale of AI-mediated editorial decision-making that most publishers' policy documents do not come close to addressing. The 2025 annual report, released earlier this month, frames this as a win. And it largely is: AI augmenting a 75-person research integrity team is a sensible model, and the author satisfaction data Springer Nature is reporting suggests the friction reduction is real.

But here's the thing Springer Nature is understandably quiet about: the tools that flagged 25,000-odd papers for integrity concerns in recent years — the same class of tools now embedded at 3.1 million submission events — were not built into the system from the start. They were retrofitted in. The policies governing how those flags get acted upon, who reviews them, and what the escalation path looks like were written on the fly, after the tools were already running.

Aries Systems, maker of Editorial Manager — the submission platform that sits underneath a significant fraction of the world's peer-reviewed journals — announced a partnership with Integra earlier this month to bring AI-driven quality and integrity checks directly into the editorial workflow. The announcement describes this as AI moving from "bolt-on to built-in." That is precisely the right way to describe it. And that phrase alone reveals how long AI has been a bolt-on: a feature someone glued onto a system that was designed without it.

AAAS, publisher of the flagship Science family of journals — one of the most prestigious and scrutinized publishing operations on the planet — is now piloting full AI automation of MDAR reporting checklists via DataSeer. Automated compliance verification for data availability, code availability, and statistical reporting standards, applied at submission, at one of the world's top journals. The significance of that is hard to overstate. It also raises a question nobody seems to be asking publicly: what were MDAR compliance rates before the automation? The answer, I'd wager, is "lower than the current numbers will look, and we'd rather not compare."

The Governance Always Arrives Late

CAS — the Chemical Abstracts Service, which manages one of the most authoritative chemistry and life sciences literature databases in existence — launched CAS Newton this week: an agentic AI system that autonomously reasons across the scientific literature rather than just searching it. This is a genuine inflection point. Moving from "AI helps you find things" to "AI reasons across the corpus on your behalf" is not an incremental step. It is a different product with a different risk profile, operating at a scale that no editorial governance committee has yet written policy for.

That is not a criticism of CAS specifically. It is an observation about the sequence of events: the technology ships, the governance comes later. Always later.

The pattern is so consistent it has become the background assumption of the industry. Liverpool University Press announced this week that it is selecting Wiley Research Exchange to modernize its journal publishing infrastructure — and the announcement explicitly cites AI-enabled integrity screening as a primary driver of the platform decision. Meaning: a publisher is now choosing its core infrastructure based partly on the AI screening capabilities embedded within it. That is a procurement decision with policy implications that will outlast any guidance document.

Meanwhile, in the United Kingdom, the government reversed its earlier proposal to allow AI companies broad text-and-data mining rights over published content, backing away under pressure from publishers and rights-holders. That reversal happened after deals had already been explored, after training datasets had already been assembled, after the commercial relationships had already been partially formed. The policy, again, arrived after the practice.

The Counterargument Worth Taking Seriously

There is a version of this story where the sequencing is fine — where practice necessarily precedes policy because nobody knows what the rules should be until the tools exist to be governed. You do not write traffic laws before you have cars. You do not write AI peer review guidelines before you know what AI peer review looks like in practice.

I have some sympathy for this view. Governance that precedes technology tends to be poorly calibrated: either so restrictive it kills useful applications, or so broad it functions as a permission slip for anything. There is a reasonable argument that Frontiers' framework is better because it was written after three years of watching what actually happens when researchers, editors, and reviewers use AI, rather than based on imagined scenarios.

That argument is true as far as it goes.

But it stops well short of where the problem actually lies. The issue is not that governance follows practice — the issue is the gap between the two, and who bears the cost of that gap in the interim.

BMJ's Journal of Medical Genetics retracted seven of eight papers from a 2019 special issue this week, after AI-powered integrity detection tools surfaced compromised peer review that the human editorial process missed at the time. The detection tools — the kind that are now standard issue in the Snapp platform and the Aries/Integra integration and the AAAS/DataSeer pipeline — did not exist in their current form when those papers were published. They do now. And so 2019 content is being evaluated against 2026 detection capabilities, with authors facing consequences for work that cleared the bar at the time.

This is not a reason to avoid using better tools. It is a reason to think harder about the accountability gap that governance-after-practice creates. Papers published under the old screening regime will keep getting re-evaluated as the tools improve. The backlog is enormous. The policies for handling that retroactive scrutiny are almost nowhere to be found.

What Good Governance Actually Looks Like

What is notable about the Frontiers framework is that it explicitly rejects the binary logic that dominated early AI policy in publishing. Allowed/not-allowed is easy to write and nearly impossible to implement fairly, because it ignores the reality that AI use exists on a spectrum — from spell-check to wholesale generation — and that the relevant question is not whether AI was used but how, to what degree, and with what level of transparency.

Frontiers is asking for structured disclosure. The framework distinguishes between use by authors, editors, and reviewers because those roles carry different responsibilities and different risks. An author who uses an AI tool to restructure a draft paragraph is doing something categorically different from an editor who uses an AI tool to flag statistical anomalies in submitted data. The governance document should reflect that difference. Apparently, until this week, most didn't.

PLOS CEO Alison Mudditt made an argument last month that I keep returning to: that the AI era doesn't just use open access content, it depends on the trustworthiness signals that good OA practice produces — open data, rich metadata, transparent retraction workflows. AI tools are only as reliable as the corpus they reason across. If that corpus is riddled with undisclosed integrity issues, the AI reasoning will propagate those issues at scale. This is what the "arrived through use, not design" problem looks like at the infrastructure level.

The governance gap in scholarly publishing is not just a compliance problem. It is a data quality problem. And the two are connected: every year that publishers deferred writing clear AI use policies was a year in which the published record accumulated undisclosed AI involvement, inconsistent disclosure standards, and a growing mismatch between what the literature claims to represent and what it actually contains.

What to Watch

The Scholarly Kitchen published a piece this week arguing that AI rollout is fundamentally a people problem, not a technology problem. Todd Toler and Angela Cochran's framing — organizational inertia, change management resistance, cultural friction — is right in ways that are uncomfortable for an industry that would prefer to treat this as a tooling question.

The tools are here. CAS Newton. DataSeer. Snapp's integrity screening. Aries' built-in checks. The infrastructure for responsible AI use in scholarly publishing exists in a form that would have seemed science fiction five years ago.

What does not yet exist, at most organizations, is the institutional will to move governance from aspiration to standard practice at the speed the technology has already moved. Frontiers' framework is genuinely good news. The fact that it is news at all in April 2026 is the problem worth fixing.

The governance gap will close — eventually. The question is what accumulates in the interim, and who gets to clean it up later.