AI Peer Review and Research Provenance: Why Verifiable AI Contributions Are Now a Scientific Necessity

The Invisible Hand: When AI Writes Research Nobody Can Audit

Imagine submitting a manuscript to a top-tier journal, knowing that segments of its introduction were drafted by a large language model, its statistical tables were restructured by an AI refactoring tool, and its conclusions were partially shaped by an AI-assisted synthesis engine — yet none of these contributions appear anywhere in the paper's metadata, author statements, or supplementary files. This is not a hypothetical scenario. It is the operational reality of scientific publishing in 2025. A newly updated preprint from arXiv, titled F(AI)²R: Who Did What, and Who Checked? Verifiable AI Provenance as an Executable Skill (arXiv:2607.25637v2), confronts this reality directly, proposing a formal, machine-readable provenance model that tracks every AI contribution across the full lifecycle of a research artefact. For anyone working at the intersection of AI peer review, research integrity, and scholarly publishing, this work demands careful attention.

The paper extends the original F(AI)²R framework — which stood for FAIR research with AI in the loop — into a generalized provenance ontology called aiprov, built as an extension of the W3C PROV-O standard. The core argument is disarmingly simple: if AI systems are now drafting, refactoring, and verifying research artefacts, their contributions must be recorded in a form that a later human reviewer or automated machine can actually audit. Not as a vague author declaration buried in a cover letter, but as structured, executable metadata attached to every artefact. The implications for AI-powered peer review systems, automated manuscript analysis, and the broader infrastructure of scientific trust are substantial.

What F(AI)²R Actually Proposes — and Why It Matters

The FAIR principles — Findable, Accessible, Interoperable, Reusable — have guided open science data management since their 2016 formalization. F(AI)²R takes this framework and applies it twice: once during AI-assisted authoring, and once during a machine-readable audit pass over every research artefact produced. The result is a two-layer provenance architecture.

In practical terms, the aiprov ontology allows researchers to attach structured provenance records to individual research components — a figure, a paragraph, a dataset transformation, a statistical test — specifying which AI agent performed which action, at what time, using which model version, and under what human supervision. This is not merely a logging exercise. The authors frame provenance capture as an executable skill, meaning it can be automated, validated programmatically, and integrated into existing research workflows without requiring manual annotation at every step.

Consider a concrete example: a research team uses GPT-4o to generate an initial literature synthesis, then uses a separate AI tool to reformat their citation list, and finally runs an automated statistical verification agent over their results table. Under the current publishing norm, none of this appears in the paper. Under an aiprov-compliant workflow, each of these three contributions would generate a provenance record linking the specific AI agent, the human who invoked it, the input artefact, the output artefact, and the validation step that followed. The entire chain becomes auditable — by journal editors, by peer reviewers, and critically, by automated peer review systems.

This distinction between declaration and auditability is the intellectual core of the paper. Dozens of journals now require AI disclosure statements. But a disclosure statement is not an audit trail. It is an assertion. The F(AI)²R framework insists that assertions are insufficient for scientific integrity; what is needed is verifiable, machine-readable evidence.

Implications for AI Peer Review and Automated Manuscript Analysis

The rise of AI peer review tools has created a productive tension in scholarly publishing. On one hand, automated manuscript analysis can process structural consistency, citation validity, statistical reporting standards, and methodological completeness at a speed and scale no human reviewer panel can match. On the other hand, if the manuscript itself was substantially produced by AI systems whose contributions are undocumented, then even a sophisticated AI peer review process is working on incomplete information — evaluating an artefact whose true provenance is opaque.

The F(AI)²R framework effectively proposes a new input layer for AI-powered peer review systems. If manuscripts arrived at review with machine-readable aiprov records attached, automated analysis tools could do considerably more than they currently do. Instead of inferring AI involvement through stylometric detection — a method with well-documented false positive rates — a review system could directly query the provenance graph: which sections were AI-drafted, which figures were AI-generated, which statistical tests were AI-executed, and which of these received documented human verification.

This shifts automated peer review from detection to validation. The question changes from "did AI write this?" to "were AI contributions appropriately supervised, documented, and verified?" That is a more scientifically meaningful question, and it is one that tools built for automated research paper analysis are better positioned to answer than human reviewers working under time pressure.

Platforms like PeerReviewerAI already provide structured automated analysis of manuscripts — examining argument coherence, methodological rigor, citation integrity, and compliance with reporting standards. The integration of provenance metadata of the type proposed by F(AI)²R would extend this capability significantly, allowing the system to cross-reference declared AI contributions against the manuscript's structural and stylistic features, flagging inconsistencies between what authors claim AI did and what the text itself suggests.

More broadly, the paper raises a challenge for the peer review infrastructure as a whole: journals that invest in AI research validation tools must also begin requiring structured provenance records as a submission condition, not merely narrative disclosure. Without the former, the latter remains unverifiable.

The Broader Transformation of AI in Scientific Research

The F(AI)²R preprint is not an isolated intervention. It reflects a broader structural shift in how research is produced and how that production must be governed. Several converging trends make this framework timely.

First, AI tool usage in research has crossed a threshold of normalization. A 2024 survey by Nature found that more than 60% of researchers had used generative AI tools in some aspect of manuscript preparation in the preceding 12 months. This is no longer an edge case or an early-adopter phenomenon. It is the median researcher behavior. The infrastructure of research integrity has not kept pace.

Second, the granularity of AI involvement has increased. Early concerns focused on AI-generated text in methods or discussion sections. The current reality is more complex: AI agents are now involved in data cleaning, figure generation, code refactoring, systematic review screening, and statistical interpretation. Each of these represents a different category of epistemic contribution with different implications for reproducibility and validity.

Third, machine learning for scientific manuscripts is developing rapidly on both the production and review sides. NLP models fine-tuned on scientific corpora can now perform literature gap analysis, methodology critique, and statistical anomaly detection with meaningful accuracy. The value of these tools depends directly on the quality and transparency of the manuscripts they analyze. Provenance-poor manuscripts are a form of information poverty for AI review systems.

The aiprov ontology, by treating provenance capture as an executable component of the research workflow rather than a retrospective disclosure obligation, aligns with a model of AI in academia where transparency is built into the process rather than appended to the product.

Practical Takeaways for Researchers Using AI Research Tools

For researchers actively using AI tools in their work, the F(AI)²R framework offers several concrete operational implications.

Document at the Point of Use, Not at the Point of Submission

The most significant practical lesson from the aiprov model is temporal. Provenance records are most accurate and most useful when captured at the moment an AI tool is used — not reconstructed weeks later when a submission is being prepared. Researchers should adopt the habit of logging AI interactions in real time, noting the tool, version, prompt type, and human oversight step for each substantive use. Simple structured templates can make this sustainable without adding significant overhead.

Treat AI Contributions as You Would Third-Party Data

Researchers already understand that using third-party datasets requires documentation of their source, version, and processing history. The same discipline should apply to AI-generated or AI-modified research artefacts. A figure partially generated by an image synthesis model is analogous to a figure derived from a licensed dataset: it requires a provenance record that allows reviewers and readers to assess its validity.

Use Automated Manuscript Analysis as a Provenance Check

Before submission, running a manuscript through an automated research paper analysis tool can help identify sections where AI fingerprints may be present but undocumented. Tools such as PeerReviewerAI can provide structured feedback on manuscript consistency, argumentation quality, and methodological transparency — serving as an independent check that your own provenance records are complete and that the manuscript reads coherently as a unified scholarly work.

Engage With Emerging Provenance Standards Early

The aiprov ontology is at the preprint stage, but the direction it represents — structured, machine-readable AI provenance as a submission standard — is one that leading journals and funders are likely to formalize over the next two to three years. Researchers who build provenance-aware workflows now will face significantly less adaptation cost when these standards become mandatory. Early adoption also positions research groups as credible practitioners of responsible AI in academia, which carries weight with reviewers, editors, and funding bodies.

Distinguish Between AI Assistance and AI Autonomy

Not all AI contributions carry the same epistemic weight. An AI tool that formats a reference list is categorically different from one that generates a hypothesis or interprets experimental results. The F(AI)²R framework's granularity — tracking contributions at the artefact level rather than the paper level — allows researchers to make these distinctions explicit. This granularity is also what makes AI research validation meaningful: a blanket disclosure that "AI was used" tells reviewers very little; a structured record of which AI agent contributed to which specific claim tells them a great deal.

Conclusion: AI Peer Review Requires an Infrastructure of Trust

The F(AI)²R framework addresses a problem that sits at the foundation of scientific knowledge production in the current period: when AI systems contribute to research artefacts without leaving auditable traces, the epistemic chain from observation to published claim is broken in ways that neither human nor automated peer review can reliably repair. The aiprov ontology represents a technically rigorous, standards-compatible approach to restoring that chain.

For the field of AI peer review specifically, the implications are significant. Automated manuscript analysis tools are only as reliable as the manuscripts they analyze. As AI contributions to research production become more extensive and more varied, the absence of structured provenance metadata increasingly limits what AI research validation systems can verify. The F(AI)²R model points toward a future where machine-readable provenance records are a standard submission component — enabling AI-powered peer review systems to perform validation rather than detection, and allowing human reviewers to focus their attention on the interpretive and contextual judgments that automated systems cannot yet make.

The broader trajectory is clear: AI in scientific research will continue to deepen, and the question of accountability for AI contributions will become more pressing, not less. Researchers, journals, and developers of scientific AI tools who invest in provenance infrastructure now are not simply complying with emerging norms — they are helping to build the epistemic foundation on which trustworthy AI-assisted science depends.