AI is writing more of our science-and disclosure isn't keeping up
An analysis of 5.2 million papers across 5000 journals shows a steep rise in AI-assisted academic writing since 2023, with physics leading the trend.
Most journals now require authors to disclose AI use. Yet in a full-text sample of 75,000 papers from 2023 onward, only 76 articles (about 0.1%) admitted to using AI. Policies exist; compliance doesn't. That's a clear transparency gap.
What the data shows
- Scope: 2021-2025 publications indexed in OpenAlex.
- Detection: AI-based screening indicates sharp growth in AI-written or AI-edited text across disciplines since 2023.
- Policy: About 70% of journals now have AI policies, mainly focused on disclosure.
- Reality: Disclosures are rare-0.1% in a 75k paper full-text check-regardless of whether journals have policies.
- Field effect: Physics shows the fastest increase.
- Language effect: Authors from non-English-speaking countries lean on AI tools more frequently.
- Access effect: Uptick is strongest in journals with high open-access publishing.
Why this matters to researchers and writers
Undisclosed AI use creates uncertainty in authorship, clarity of contribution, and the reliability of claims and citations. It also makes peer review harder and reproducibility fuzzier.
The takeaway: disclosure rules alone aren't moving behavior. We need practical workflows that make responsible AI use the default-not an afterthought.
Practical steps for researchers and science writers
- Create a simple team policy: where AI is allowed (e.g., outlining, language polishing, translation), where it's not (e.g., generating references, fabricating data), and mandatory human verification steps.
- Maintain an "AI log" per manuscript: tools/models used, versions, dates, tasks, and what you verified manually. Keep it in your repo alongside methods and analysis notes.
- Disclose with specifics: tasks (ideation, editing, translation), model names and versions, dates, and guardrails (fact-checking, citation audits, plagiarism checks).
- Ban generated references. Audit citations by checking DOIs and full-text relevance. If the tool suggests a source, you still verify it end to end.
- Protect data: avoid pasting sensitive or unpublished material into third-party tools unless your institution has an approved, private instance.
- For non-native English authors: use AI for clarity and grammar, but keep original meaning, and confirm technical terms. Document this as "language editing."
- Flag figures, tables, or images touched by AI-even for layout or denoising-if your journal policy requires it.
What editors and publishers can do now
- Replace vague bans with enforceable workflows: structured AI-use statements at submission, with checkboxes for tasks, model/version, and verification steps.
- Add process checkpoints: automated detection plus targeted editorial queries when risk signals are high.
- Publish AI-use metadata with the article record and DOI so readers and indexers can see it.
- Offer language support or editing credits to reduce dependence on undisclosed AI for authors who need it.
- Provide clear templates and examples of acceptable AI use for common scenarios (translation, copyediting, outline drafting).
Ready-to-use AI disclosure template
AI-use statement: The authors used [tool/model, version] on [date(s)] for [tasks: grammar editing, translation, outline drafting]. No text, data, figures, or references were generated without human review. All claims and citations were checked manually. No proprietary, confidential, or personal data were entered into third-party systems.
Ethics and guidance
Stronger norms start with clarity. Journals can pair disclosure with verification steps, and authors can keep a simple audit trail. For community standards, see the Committee on Publication Ethics (COPE) guidance as it evolves.
Bottom line
AI use in scientific writing is rising fast; disclosure isn't. Policing isn't enough. Build workflows that make transparency easy, verification routine, and credit clear. That's how we keep the literature credible while still benefitting from the tools.
Further resources
Your membership also unlocks:
