Industries / Media, arts and recreation / Media and entertainment / Publishing
Scholarly journals
Journal publishing collects research papers, has other researchers check them before publication, and makes them available to readers by subscription or, increasingly, free to read.
7
Milestones achieved, from 1665 to 2025.
2
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1665
Science gets a journal
Henry Oldenburg, secretary of London's Royal Society, began Philosophical Transactions in 1665, a periodical of letter extracts, book summaries and accounts of experiments that he chose and often wrote himself. When the Society took the journal over in 1753, a Committee of Papers voted on each paper by secret ballot and could call in a more expert member, which the Royal Society describes as the first provision for peer review written into its statutes. The Royal Society calls Philosophical Transactions the world's first and longest-running scientific periodical.
Credit
Henry Oldenburg; the Royal Society
1973
Nature requires outside referees
From 1939 the editors of Nature, a leading weekly science journal, sent papers to outside experts only some of the time, and their successor John Maddox printed papers he admired without outside review. The historian Melinda Baldwin found that it was not until 1973 that editor David Davies made external peer review, a check of each paper by researchers outside the journal, a requirement for publication.
LessonPeer review as a rule for every paper is younger than it looks; by Baldwin's account a journal could be thought scientifically respectable without it as late as the 1960s.
Credit
David Davies, editor of Nature
1991
Papers shared before review
The physicist Paul Ginsparg started arXiv in 1991, a free server where researchers post papers, called preprints, before or without journal review. Volunteer moderators check that a paper fits its subject and has scholarly value, but arXiv does not peer review it. arXiv says it now holds more than three million articles in eight subject areas, and it became an independent nonprofit in 2026 after decades at Cornell University. In October 2025 its computer science section began rejecting review articles and position papers that a journal or conference had not already accepted, citing an "unmanageable influx" that AI language models had made fast and easy to write.
LessonAn open door with no peer review works only while moderators can keep up with what comes through it.
Credit
Paul Ginsparg; Cornell University
2016
Duplicated images counted
Elisabeth Bik, Arturo Casadevall and Ferric Fang looked by eye at the images in 20,621 papers published in 40 journals from 1995 to 2014. They found problematic figures in 3.8% of the papers, at least half with features suggesting deliberate manipulation, and the share had risen markedly in the previous decade.
LessonThe rate varied widely between journals, which the authors read as a sign that screening images before publication makes a difference.
Credit
Elisabeth Bik, Arturo Casadevall, Ferric Fang (published in mBio)
2018
Funders demand immediate access
On 4 September 2018 a group of national research funders, backed by the European Commission and the European Research Council, launched cOAlition S and its Plan S. From 2021, papers from their grants must be free to read on publication, in open-access journals or platforms or in a repository with no embargo (a waiting period before free release), under an open licence, preferably Creative Commons Attribution (CC BY), which lets anyone reuse the work if they credit the authors. The funders do not pay for "hybrid" journals, which sell subscriptions and also charge for open articles, except within agreements to switch fully to open access.
Credit
cOAlition S funders; the European Commission; the European Research Council
2023
Paper mills close journals
Paper mills, organizations that subvert academic publishing for money with fabricated text, manipulated images and rigged peer review, got large numbers of papers into journals run by Hindawi, a publisher that the US firm Wiley bought in 2021. Wiley shut four Hindawi journals in May 2023 "to mitigate against systematic manipulation of the publishing process", dropped the Hindawi name that December, and told The Register in May 2024 that it had retracted more than 11,300 Hindawi papers in two years. Wiley put an $18 million fall in its research publishing revenue down mainly to the Hindawi affair.
LessonFraud organized outside a journal can take over its review process at scale, and the cleanup cost the publisher years, revenue and a brand.
Credit
Wiley; Hindawi
2025
Embargo ends, delays remain
The US National Institutes of Health (NIH) replaced its 2008 rule, which let journals keep an NIH-funded paper behind a paywall for up to 12 months, with a policy that took effect on 1 July 2025: the accepted manuscript must be public on the day of publication in PubMed Central (PMC), the National Library of Medicine's free archive of papers. Haining Wang compared 1,880 NIH-funded articles from January to April 2026 with 1,880 from 2025, each matched to an article in the same journal and month that PubMed does not list as US government supported. By four months after publication, availability had risen 19.4 percentage points more for NIH-funded articles than for the comparison articles, but only 0.1% of NIH-funded articles were in PMC on the day of publication, and the median wait for author manuscripts that had reached PMC fell from 374 days to 70.
LessonSetting a date in a rule does not make articles appear on it; Wang recommends that agencies count the days to public access, not only whether a paper is eventually deposited.
Credit
NIH; Haining Wang, who measured it
Watch for
AI reviews for every paper
The organizers of AAAI-26, a large AI conference, report that every one of its 22,977 main-track papers got one clearly labeled AI review alongside the human reviews, produced in less than a day by a system that combined frontier models, tool use and safeguards. In their survey of authors and program committee members, respondents found the AI reviews useful and preferred them to human reviews on technical accuracy and research suggestions.
Achieved whena group outside the organizing team checks a sample of the AI reviews' points against the papers and reports how often they were right and whether they changed decisions, or another conference repeats the trial and publishes the same measures.
Claimed by
Joydeep Biswas and 12 co-authors, reporting on the AAAI-26 pilot, April 2026
A fifth of reviews AI-written
Pangram Labs, which sells an AI-text detector, ran it on all 75,800 peer reviews and 19,490 papers submitted to ICLR 2026, a major AI conference, and reports that 21% of the reviews (15,899) were fully AI-generated and more than half showed some AI involvement. Run on ICLR's 2022 reviews as a check, it labeled none as fully AI-generated. ICLR said it would not base decisions on detection tools alone.
Achieved whena different detector, or people labeling a random sample of the reviews by hand, finds a share within a few points of Pangram's.
Claimed by
Pangram Labs, November 2025
Open
01
Corrections as a test set
A public test of AI reviewers built from real mistakes. Collect 1,000 open-access, peer-reviewed papers whose authors later published a correction notice naming a specific error (a wrong number in a table, a mislabeled figure, a faulty calculation). Give each AI reviewer the original, uncorrected paper and score whether it points to the passage the correction later fixed.
The briefCredit · unclaimed02
Days to public access
A monthly public count, journal by journal, of how many days each NIH-funded article takes to appear free in PubMed Central after it is published, covering every such article since the no-embargo rule began on 1 July 2025, with a matched comparison set as in Wang's study.
The briefCredit · unclaimed
Open 01 · the brief
Corrections as a test set
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public test of AI reviewers built from real mistakes. Collect 1,000 open-access, peer-reviewed papers whose authors later published a correction notice naming a specific error (a wrong number in a table, a mislabeled figure, a faulty calculation). Give each AI reviewer the original, uncorrected paper and score whether it points to the passage the correction later fixed.
Why it is still open
The tests built so far are small or rest on other records. SPOT (May 2025) pairs 83 published papers with 91 errors serious enough to prompt an erratum or a retraction, checked with the authors; no model it tested passed 21.1% recall, and across eight runs models rarely found the same errors again. FLAWS (November 2025) uses 713 errors that AI models inserted into papers accepted at ICML 2025; the best model, GPT-5, found the error among its top ten guesses 39.1% of the time. Tianmai Zhang and Neil Abernethy (2025) used papers withdrawn from arXiv, whose withdrawal notes often do not state the exact error. This research found no test drawn from the correction notices that publishers register with Crossref across all fields.
Built with what exists
The free Crossref API, which listed 218,740 records registered as corrections in November 2026, with a link from each notice to the paper it corrects; open-access full texts from PubMed Central or publishers' sites; AI models that read whole papers, including an open-weight model so others can rerun the test (see the Note "A trillion open parameters"); each paper run 20 times, since one pass can hide unstable answers (see the Note "All twenty times"); Python for scoring.
First step
Pull 2025 correction notices from the Crossref API, keep those whose original paper is open access, and hand-mark the error in the first 50.
Done looks like
1,000 paper-correction pairs, each with the corrected passage marked by a person; for each AI reviewer, the share of errors found in its top three and top ten flags, with how often the result repeats across 20 runs; the pairs and scoring code released.
Open 02 · the brief
Days to public access
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A monthly public count, journal by journal, of how many days each NIH-funded article takes to appear free in PubMed Central after it is published, covering every such article since the no-embargo rule began on 1 July 2025, with a matched comparison set as in Wang's study.
Why it is still open
Haining Wang's 2026 study measured the change once, on a sample of 1,880 NIH-funded articles from January to April 2026, and found only 0.1% public on the day of publication. Wang notes that whether ending the embargo had made articles available sooner had not been measured before, and recommends that agencies count days to public access. This research found no ongoing count by journal or publisher.
Built with what exists
PubMed's E-utilities interface for NIH-funded articles and their publication dates; the PMC ID Converter and PMC records for the date each article went live; the NIH RePORTER API to link articles to grants; Python and a static web page for the table.
First step
Repeat Wang's method on one month, January 2026, and compare the result with his figures.
Done looks like
Every NIH-funded article published since July 2025 with its days to PMC, grouped by journal and publisher and refreshed monthly; Wang's headline figures reproduced on his period to within one percentage point; 200 articles checked by hand against PMC's own dates.