Research Publication Under Pressure: Costs, Credibility and Machine-Assisted Peer Review

For most of the past two decades, the debate about research publication was mainly about access: who could read a paper, and after how long. That question is close to settled in principle. The harder questions now are about money and about trust — who pays for publication when funders start refusing to, and how a reader can tell whether a paper, or the review that validated it, was written by a person who understood it.

Three pressures are converging in 2026. Public funders that spent a decade mandating open access are now trying to cap what they will spend on it. Fraudulent manuscript production has industrialised, and generative AI has lowered its cost. And peer review — the quality-control layer the whole system rests on — is absorbing large-language-model assistance faster than venues can measure or govern it.

This article is written for people who have to make decisions inside that system: researchers choosing where to submit, research administrators forecasting budgets, editors and reviewers setting policy, and executives in publishing, pharma, healthcare and technology who rely on the published record as an input. It sets out what the evidence currently supports, what remains contested, and what a defensible strategy looks like while the rules are still moving.

---

Executive summary

• The 2022 US Office of Science and Technology Policy (OSTP) "Nelson memo" required federal agencies to end publication embargoes, with agency policies due in effect by 31 December 2025; most agencies implemented, but the memo's future is now uncertain (AIP FYI, 2026).
• The Trump administration's FY2027 budget request proposes a government-wide prohibition on using federal funds for expensive journal subscriptions and high publishing charges without prior approval; NASA and NIST budget documents state they will stop paying such costs (AIP FYI, 2026).
• The National Institutes of Health (NIH) proposed five options for limiting allowable publication costs in July 2025, ranging from a $2,000 per-article cap to disallowing publication costs entirely. As of 11 August 2026, no final NIH Guide notice setting a cap had been published.
• Estimated global spending on article processing charges (APCs) at six major publishers rose from roughly $910 million in 2019 to about $2.54 billion in 2023 (Schneider et al., 2024, preprint).
• Gold open access grew far faster than the literature overall between 2014 and 2024, at roughly 16% compound annual growth versus 4% for all article types (STM, 2025).
• Paper-mill retractions cluster in identifiable journal profiles: one cross-sectional study found 142 journals retracted 2,051 paper-mill articles between 2020 and 2022, concentrated in high-volume, mid-quartile medicine and health titles (Mascato Fontaíña et al., 2025).
• A vendor analysis of the ICLR 2026 review cycle classified about 21% of roughly 75,800 reviews as fully AI-generated, with over half showing some AI involvement — a striking figure that comes from an AI-detection company and is contested by other detectors.
• Publisher and society policy has converged on one firm rule — do not upload confidential manuscripts to third-party AI tools — while diverging widely on everything else.
• The practical risk for professionals is no longer only "can I read this?" but "can I fund this, and can I trust this?"

---

How research publication reached this point

Scholarly journals began in 1665 as a way to register priority and circulate correspondence. The structure most researchers now treat as natural — external refereeing before publication — became routine only after the mid-twentieth century, and the subscription-bundle economics that funded it emerged later still, as commercial publishers consolidated society titles.

The open-access movement, formalised in the Budapest Open Access Initiative of 2002, proposed a straightforward correction: research paid for by the public should be readable by the public. Two mechanisms followed. Green open access asks authors to deposit accepted manuscripts in repositories. Gold open access shifts payment from reader to author, usually through an APC. A third model, diamond open access, charges neither, and is instead underwritten by institutions, consortia or ministries.

Policy has steadily favoured immediacy. The 2013 OSTP memo allowed a 12-month embargo; the 2022 memo, issued on 25 August 2022 under then-acting director Alondra Nelson, removed it and extended coverage to all federal research-funding agencies, with policies required to be in effect no later than 31 December 2025. The NIH brought its own compliance date forward to 1 July 2025.

The unresolved question was never whether papers should be free to read. It was who absorbs the cost once they are.

---

The open-access economy meets fiscal reality

That question is now being answered, at least in the United States, by subtraction.

Independent estimates put APC spending at six large publishers at about $8.35 billion cumulatively between 2019 and 2023, with annual spend nearly tripling from roughly $910 million to $2.54 billion (Schneider et al., 2024). OSTP itself estimated in a 2022 congressional report that US taxpayer-covered open-access publishing costs fell somewhere between $390 million and $789 million a year — a range wide enough to signal how poor the underlying data is.

On 8 July 2025, NIH director Jay Bhattacharya announced an intention to cap allowable publication costs. The follow-up request for information (NOT-OD-25-138, 30 July 2025) put five options on the table, including disallowing publication costs altogether, a flat $2,000 per-article cap, a $3,000 cap available only at journals that compensate reviewers and publish reviewer reports, and a proportional limit of 0.8% of direct costs or $20,000 per award. NIH's own analysis found a median APC of $950 globally and $2,040 among US journals, against an average requested cost per publication of roughly $2,565 to $3,104 in FY2025 R01 budgets.

The comment period closed on 15 September 2025 and the proposed effective date was 1 January 2026, but as of this article's publication no final notice fixing a figure had appeared. Administrators should treat every number above as proposed, not binding.

The wider signal is less ambiguous. The FY2027 president's budget request calls for a government-wide prohibition on federal funds paying for expensive journal subscriptions and high publishing costs unless required by statute or approved in advance. NASA's and NIST's budget documents state they will no longer use federal funds for journal subscriptions or publication fees. Reporting by AIP's FYI service in May 2026 indicated OSTP was in the process of repealing the Nelson memo, with congressional appropriations documents referring to that process and the House Appropriations Committee asking the National Science Foundation to pause new public-access implementation pending OSTP guidance.

Callout — read this carefully Repealing the memo would not automatically repeal agency policies, which were issued separately and remain in force unless individually withdrawn. As of August 2026 the practical position is: comply with your funder's published policy, and assume the funding for compliance is less secure than the requirement itself.

Table 1 — Publication routes compared (original OneWise analysis)

Cost ranges reflect publicly listed 2025–2026 prices and vary widely by journal and waiver policy. Verify current figures with the publisher before budgeting.

---

Integrity at scale: paper mills and the fraud economy

Retractions have risen sharply, and the reasons have shifted. Where retraction once mostly signalled honest error, a substantial share now reflects organised misconduct: fabricated data, sold authorship, and manipulated or fake peer review.

The most rigorous published characterisation of the pattern comes from a 2025 cross-sectional study in Research Integrity and Peer Review, which identified 142 journals that retracted 2,051 paper-mill articles between January 2020 and December 2022. Half of those journals had a single retraction, but a small group had more than 100 each. The study concluded that paper-mill retractions concentrate in journals sharing a profile: high open-access rates, intermediate impact-factor quartiles, very high volumes of citable items, and classification in medicine and health.

Two qualifications matter. First, retraction data measures detection, not prevalence — a journal with many retractions may be screening well, not failing. Second, the largest single spike, the Hindawi retractions of 2023, reflected one publisher unwinding one compromised portfolio, which distorts year-over-year comparisons.

Publisher-side analyses point in a consistent direction. MDPI's own review of 2025 retractions across the industry attributed 43% of cases to compromised peer review or paper mills, with compromised peer review alone accounting for 37% — a figure that implicates editorial gatekeeping, not only author behaviour. Trade reporting drawing on the Retraction Watch database has put paper-mill activity at roughly a fifth of health and life-science retractions logged between 2016 and 2025. These are useful indicators, not peer-reviewed prevalence estimates.

Figure 1 — suggested original diagram

Title: Where integrity failures enter the publication pipeline Purpose: Show that fraud is not a single event at submission but a set of entry points across the workflow, each with a corresponding control. Layout: A horizontal five-stage pipeline, left to right: Study design & data → Manuscript preparation → Submission & editorial triage → Peer review → Post-publication record. Above the pipeline (red downward arrows, labelled "failure modes"): fabricated or reused images and data; paper-mill authorship sale and undisclosed AI generation; identity fraud and hijacked email domains; rogue editors, fake reviewer accounts, review-manipulation rings; citation cartels and undetected errors. Below the pipeline (blue upward arrows, labelled "controls"): data availability statements and protocol registration; image-integrity and text-similarity screening; ORCID and institutional-affiliation verification; reviewer-identity checks and structured review forms; post-publication commenting, corrections and retraction indexing. Visual hierarchy: Pipeline stages in equal-width blocks; failure arrows in a warm tone, controls in a cool tone; no stage left without at least one control. Caption: Integrity controls map to specific pipeline stages; gaps at any stage are exploitable.

---

AI in peer review: what the evidence actually shows

The most contested development of the past year is the scale of large-language-model use by reviewers, not authors.

The first systematic evidence came from a 2024 study estimating that between roughly 6.5% and 16.9% of review text at several major AI conferences showed signs of substantial LLM modification. In November 2025, the AI-detection company Pangram Labs analysed the ICLR 2026 cycle — approximately 19,500 submissions and 76,000 reviews — and reported that about 21% of reviews (roughly 15,900) were fully AI-generated, with more than half showing some AI involvement. Around 1% of submissions were classified as fully AI-generated.

That number deserves both attention and scepticism. Pangram sells detection tools and has a commercial interest in the finding. Detector disagreement is real: a 2026 methods preprint comparing detectors on a sample of ICLR 2026 reviews found one tool classifying 24% as AI-generated and 44% as human, while its own method returned 5% and 34% respectively — the same corpus, very different pictures. NeurIPS organisers publicly reported that 28.2% of 969 position-paper submissions received a maximum AI score from the same vendor tool, then examined how window size and permissible AI use could inflate that reading.

There is, however, corroboration that the phenomenon is not an artefact. An independent 2026 preprint searched roughly 160,000 author comments on ICLR 2026 submissions, identified 58 cases where authors specifically accused a review of being AI-generated, and found that 50 of those (86%) had been independently flagged, with authors citing concrete evidence such as hallucinated citations.

A reasonable reading of the evidence: undisclosed LLM use in peer review is widespread at large computing venues, the precise share is not reliably measurable with current detectors, and generalising these figures to biomedical or humanities journals is unsupported.

Adversarial behaviour has followed. Documented cases in 2025 involved authors embedding hidden instructions in manuscripts intended to manipulate an LLM reviewer into a favourable assessment — a prompt-injection attack against the review process itself.

Table 2 — How institutions have positioned themselves on AI in review

Synthesis of publicly stated policies as of mid-2026; policies change frequently, so verify against the venue's current instructions.

Table 3 — Confidence levels for the claims in this article

---

Latest developments

• 8 July 2025 — NIH announces intent to cap allowable publication costs from FY2026.
• 30 July 2025 — NIH publishes NOT-OD-25-138 with five cost-limiting options; comments closed 15 September 2025.
• November 2025 — Pangram Labs publishes its ICLR 2026 review analysis; the figure is reported by Nature's news team and triggers venue-level responses.
• 31 December 2025 — Deadline for federal agency public-access policies to be in effect under the 2022 OSTP memo.
• January 2026 — Congressional appropriations language asks OSTP to report on the status of its process of repealing the Nelson memo.
• March 2026 — COPE updates its guidance for editors who suspect reviewers are using AI.
• April 2026 — USDA issues an updated public-access policy that additionally restricts publication in journals whose publishers operate from countries of concern.
• May 2026 — FY2027 budget documents propose a government-wide restriction on federally funded subscription and publication costs; House appropriators ask NSF to pause new public-access implementation.
• June 2026 — Elsevier's generative-AI policy for journals is updated; NeurIPS publishes its own analysis of AI-detection results on a submission track.

---

Best practices, and mistakes to avoid

Practices with reasonable support

• Deposit the accepted manuscript in the designated repository on acceptance, every time. It is the cheapest reliable route to compliance and is unaffected by APC policy changes.
• Retain rights deliberately. Institutional rights-retention policies and CC BY licensing on the author accepted manuscript reduce dependence on any single publisher's terms.
• Screen venues before submission using journal-level signals: indexing in DOAJ, membership of COPE, transparent editorial board affiliations, and retraction history.
• Treat every unpublished manuscript as confidential data. Do not paste it into a general-purpose AI tool, regardless of the vendor's retention settings.
• Budget publication costs as a variable, not a fixed line, and document the basis of the estimate.

Common mistakes

• Assuming an embargo is still permitted because a publisher's default workflow implies one.
• Equating "open access" with "quality-assured" — the two are independent variables.
• Using AI-detection percentages as if they were measurements rather than estimates from proprietary classifiers.
• Citing a retracted paper because the reference was captured before retraction; check reference lists against retraction indexes before submission.
• Choosing a journal on impact factor alone when funder cost rules may make that choice unbudgetable next year.

---

Practical takeaways by role

Researchers and principal investigators. Build a two-tier venue list for every project: a preferred target and a compliant, low-cost fallback. Confirm your funder's current allowable-cost guidance at proposal stage, not at acceptance.

Research administrators and librarians. Model at least two scenarios for the next budget cycle — one where funder APC support continues, one where it is capped or removed. Scenario planning is cheap; a mid-year funding gap is not.

Editors and reviewers. Write your AI policy in terms of actions rather than intent: what may not be uploaded, what must be disclosed, and who remains accountable. Ambiguous policies are unenforceable.

Executives relying on published evidence. Treat single papers as hypotheses. For decisions with material consequences, require replication, registered protocols or systematic review evidence, and check retraction status as a routine control.

---

Key insights

1. Access is largely won as policy; the fight has moved to who funds it.
2. Zero-embargo mandates and shrinking publication budgets are pulling in opposite directions.
3. Repository deposit is the most policy-resilient publication strategy available today.
4. Hybrid open access is the model most exposed to funder cost caps.
5. Retraction volume measures detection capability as much as misconduct prevalence.
6. Paper-mill risk concentrates in high-volume journals with weak editorial control, not in open access as such.
7. Compromised peer review implicates editors and reviewer networks, not only authors.
8. Undisclosed AI use in review appears widespread at large computing venues; the exact scale is not reliably measurable.
9. AI detectors disagree substantially with one another and should not be treated as adjudicators.
10. The scarce resource in publishing is no longer distribution — it is credible expert attention.

---

Frequently asked questions

What is research publication? The formal process of registering, certifying, disseminating and archiving research findings — most commonly through peer-reviewed journal articles, conference papers, preprints and data publications.

What is an article processing charge? A fee paid by an author, institution or funder to make an article openly available at the point of publication. Listed prices in 2025–2026 range from under $1,000 to over $12,000 depending on the journal.

Did the US zero-embargo requirement take effect? Yes. Federal agency public-access policies were required to be in effect by 31 December 2025 under the 2022 OSTP memo, and most agencies implemented policies accordingly. The memo's own status is now under review.

Has NIH capped article processing charges? Not as a final published policy. NIH proposed five options in July 2025 with a proposed 1 January 2026 effective date, but as of 11 August 2026 no NIH Guide notice fixing a cap had been issued. Confirm current guidance with your research office.

What is green open access, and does it satisfy funders? Depositing the peer-reviewed accepted manuscript in a designated repository. It satisfies most public-access mandates and carries no author fee, which is why it is the most robust route under cost pressure.

What is diamond open access? A model in which neither readers nor authors pay, with costs covered by institutions, consortia or public funders. It removes author-side barriers but depends on sustained institutional commitment.

What is a paper mill? An organisation that sells fabricated manuscripts or authorship positions. Undisclosed use is misconduct, and paper-mill output is a leading cause of mass retractions.

How can I check whether a paper has been retracted? Check the publisher's page for the version of record, the Retraction Watch Database, and Crossref's Crossmark record. Reference-management tools increasingly surface retraction flags automatically.

Is it acceptable to use AI to write a peer review? Under most current publisher, society and conference policies, no. Uploading a confidential manuscript to a third-party AI tool is prohibited by major publishers and by COPE guidance; limited language assistance is sometimes permitted with disclosure.

Are AI detectors reliable enough to accuse someone? No. Independent comparisons in 2026 showed different detectors producing markedly different estimates on the same corpus. Detector output is a signal for investigation, not evidence of misconduct.

Should I post a preprint? In most quantitative fields, yes — it establishes priority and accelerates feedback. Check the target journal's preprint policy first, since a minority of journals still treat prior posting as prior publication.

Does open access increase citations? Some studies report a citation advantage, but the effect is confounded by article quality, discipline, funding and self-selection. Treat the claim as plausible but not settled.

---

Glossary

• APC (article processing charge) — Fee to publish an article openly.
• COPE — Committee on Publication Ethics; issues guidance and case advice for editors and publishers.
• DOAJ — Directory of Open Access Journals; a curated index used widely as a quality signal.
• Embargo — A delay between publication and free public availability.
• Green / gold / diamond OA — Repository-based, author-side-funded, and no-fee models of open access respectively.
• Hybrid journal — A subscription journal offering per-article open access for a fee.
• Paper mill — A commercial operation producing fraudulent manuscripts or selling authorship.
• Preprint — A manuscript posted publicly before formal peer review.
• Prompt injection — Hidden instructions embedded in content to manipulate an AI system processing it.
• Public access policy — A funder requirement that outputs of funded research be made freely available.
• Retraction — Formal withdrawal of a published article from the scholarly record.
• Rights retention — An institutional or funder policy under which authors keep re-use rights in their accepted manuscripts.
• Transformative agreement — A contract converting institutional subscription spend toward open-access publishing.
• Version of record — The final published version, as distinct from the accepted manuscript.

---

References

Government and funder sources

National Institutes of Health. (2024). 2024 NIH public access policy (NOT-OD-25-047). https://grants.nih.gov/grants/guide/notice-files/NOT-OD-25-047.html

National Institutes of Health. (2024). Supplemental guidance to the 2024 NIH public access policy: Publication costs (NOT-OD-25-048). https://grants.nih.gov/grants/guide/notice-files/NOT-OD-25-048.html

National Institutes of Health. (2025, July 8). NIH to establish new policies for allowable publication costs [Statement of the NIH Director]. https://www.nih.gov/about-nih/nih-director/statements/nih-establish-new-policies-allowable-publication-costs

National Institutes of Health. (2025, July 30). Request for information on maximizing research funds by limiting allowable publishing costs (NOT-OD-25-138). https://grants.nih.gov/grants/guide/notice-files/NOT-OD-25-138.html

Office of Science and Technology Policy. (2022). Ensuring free, immediate, and equitable access to federally funded research. Executive Office of the President.

Academic papers and preprints

Mascato Fontaíña, N., Candal-Pedreira, C., García, G., et al. (2025). Identifying common patterns in journals that retracted papers from paper mills: A cross-sectional study. Research Integrity and Peer Review, 10, 21. https://doi.org/10.1186/s41073-025-00177-9

Schneider, J., et al. (2024). Estimating global article processing charges paid to six publishers for open access between 2019 and 2023 [Preprint]. arXiv. https://arxiv.org/abs/2407.16551

Teixeira da Silva, J. A. (2025). The citation of retracted papers and impact on the integrity of the scientific biomedical literature. Learned Publishing. https://doi.org/10.1002/leap.1667

Industry, publisher and standards sources

Committee on Publication Ethics. (2026). Editors suspect that reviewers are using artificial intelligence [Case guidance, updated March 2026]. https://publicationethics.org/guidance/case/editors-suspect-reviewers-are-using-artificial-intelligence

Elsevier. (2026). Generative AI policies for journals [Policy updated June 2026]. https://www.elsevier.com/about/policies-and-standards/generative-ai-policies-for-journals

Pangram. (2025, November). Pangram predicts 21% of ICLR reviews are AI-generated. https://www.pangram.com/blog/pangram-predicts-21-of-iclr-reviews-are-ai-generated

STM Association. (2025). Uptake of open access [Open access dashboard]. https://stm-assoc.org/oa-dashboard/oa-dashboard-2/uptake-of-open-access/

Reporting

McKenzie, L. (2026, May 20). Scholarly publishing costs under scrutiny by Trump administration. FYI: Science Policy News, American Institute of Physics. https://www.aip.org/fyi/scholarly-publishing-costs-under-scrutiny-by-trump-administration

NeurIPS. (2026, June 2). AI-generated papers in the NeurIPS 2026 position paper track [Conference blog]. https://blog.neurips.cc/2026/06/02/ai-generated-papers-in-the-neurips-2026-position-paper-track/

---

Editorial note: This article synthesises publicly available policy documents, peer-reviewed literature, preprints and industry reporting. Where evidence is preliminary, vendor-supplied or contested, that status is stated in the text. Figures describing proposed policies were accurate as of 11 August 2026 and should be reverified before use in budgeting or compliance decisions.