Back to Basics: Why AI Disclosures in Publishing Should Center on Accountability, Not Reporting

This article is the second in a two-part series on AI disclosures in scholarly publishing. In Part 1, I argued that despite widespread AI use by researchers, author disclosure remains the exception rather than the norm. I explored why current disclosure guidelines are failing and why fear, ambiguity, and burden are driving AI use underground rather than making it transparent. In this follow-up, I turn to the more challenging question: what publishers should do about it. 

The central mistake publishers make is assuming that more detailed reporting of AI use will lead to greater research integrity. In practice, the opposite seems to be happening. Overly demanding and vague disclosure requirements have created a transparency vacuum while doing little to protect what matters most in science: replication, reproducibility, and trust in the scholarly record.

The way forward is not more surveillance of authors’ workflows, but a renewed emphasis on author responsibility and outcome-based integrity.

The Wrong Question: “How Did You Use AI?”

Most current publisher policies implicitly ask authors to answer a deceptively simple question: How did you use AI?

This framing creates three immediate problems.

First, it assumes that AI use is a discrete, identifiable event rather than something increasingly embedded in everyday research tools. Second, it treats all AI use as equally relevant, when in fact AI assistance in writing prose presents fundamentally different risks than AI involvement in data analysis or modeling. Third, it shifts attention away from scholarly outcomes and toward process documentation: prompt logs, screenshots, and appendices that editors don’t have the time or expertise to evaluate.

In short, publishers are asking for information they cannot meaningfully use, while failing to ask for assurances they genuinely need.

The Right Question: “Can This Work Be Trusted and Reproduced?”

AI governance in publishing should be built around a much older and sturdier foundation: the norms of research integrity that long predate LLMs.

From that perspective, the relevant questions are not:

• Did the author use ChatGPT?
• What prompts did the author use?
• Do peer reviewers have familiarity and access to declared tools?

But rather:

• Are the data reliable?
• Are the methods transparent?
• Can the results be reproduced?
• Is authorship responsibility clearly assigned?

Reframing AI policy around these questions immediately clarifies where disclosure matters and where it does not.

A Practical Framework for Publishers

1. Replace Narrative Disclosures with Structured, Low-Burden Declarations

Rather than asking authors to “describe all AI use,” journals should implement simple, task-based declarations embedded directly into submission systems.

Authors should be asked to indicate categories of AI use (e.g., literature discovery, data analysis, code generation, language editing), not narrate workflows or share prompts. This standardization reduces ambiguity, minimizes burden, and creates consistent signals for editors without inviting overinterpretation. Crucially, such declarations should be routine and neutral, not framed as exceptional or suspicious.

2. Escalate Requirements Only Where Reproducibility is at Stake

Not all AI use poses the same risk to the scholarly record, and policies should state this explicitly.

AI use in data processing, statistical analysis, modeling, or code generation has direct implications for reproducibility and therefore warrants additional disclosure and justification. AI use in language editing, translation, or prose drafting generally does not.

Publishers should therefore adopt conditional disclosure according to editor discretion: deeper and more detailed disclosures should be required when AI touches analytic or methodological components. This keeps policy aligned with scholarly risk rather than distracting issues about writing assistance.

(PDF of table below available here)

3. Introduce an Explicit Author Responsibility Statement

At the heart of effective AI governance should be a simple but powerful declaration of accountability.

Authors should be required to affirm clearly and unambiguously that:

☐ I take full responsibility for all data, analyses, and interpretations in this manuscript, regardless of the tools used.

☐ I confirm that AI tools did not replace my judgment in analytical decisions.

☐ I attest that the work meets the journal’s standards for reproducibility and research integrity.

This approach mirrors existing practices around conflicts of interest and data availability. It avoids technological micromanagement while making responsibility unmistakable. Most importantly, it restores a principle that has become oddly obscured in AI debates: tools do not bear responsibility, authors do.

Educating Editors Matters as Much as Instructing Authors

One reason authors are reluctant to disclose AI use is not policy language, but editorial culture.

If authors believe that disclosure will quietly bias reviewers against them or signal lower rigor or competence, they will continue to hide AI use, regardless of how policies are written.

Publishers must therefore do something simple but significant: state clearly that AI use is not grounds for rejection.

Internally, editors and reviewers should be trained to evaluate AI-related concerns only insofar as they affect methodological soundness, data integrity, or reproducibility. 

Editors and reviewers should be trained to ask:

• Does AI use affect methods validity?
• Does it obscure data provenance?
• Does it compromise replication?

This keeps AI evaluation where it belongs: inside scholarly assessment, not stylistic prejudice. Style, fluency, or perceived “AI-ness” of prose should be explicitly excluded from integrity judgments.

Carrots and Sticks: Incentivizing Responsible AI Use Without Policing

One objection to responsibility-based frameworks is that they “lack teeth.” This objection misunderstands both where enforcement should occur and how behavior actually changes in scholarly communities.

Sanctions should not be tied to using AI, but to failing core scholarly obligations. At the same time, publishers should recognize that transparency improves not only when poor behavior is punished, but when good behavior is visibly rewarded.

The “Carrot”: Normalizing and Recognizing Responsible Disclosure

Publishers have an opportunity to reframe AI transparency as a marker of good scholarly practice rather than a red flag. Even modest, symbolic recognition can shift author behavior meaningfully.

Examples of low-cost, high-signal incentives include:

• Standardized “Responsible AI Declaration” statements that appear alongside data availability or ethics statements, signaling good practice rather than suspicion.
• Editorial acknowledgments (visible to authors but not necessarily highlighted to reviewers) that thank authors for clear and complete AI-related declarations.
• Badging or labeling systems similar to “open data” or “transparent methods” indicators recognizing manuscripts that clearly articulate responsibility for AI-assisted analyses.
• Positive reinforcement in author guidelines, explicitly stating that transparent AI use is viewed as a strength, not a liability, when evaluating submissions.

None of these confer advantage in peer review, but all help normalize disclosure and reduce the perception that honesty carries risk.

The “Stick”: Proportionate, Outcome-Based Sanctions

Where sanctions are necessary, they should be proportionate and tied to scholarly outcomes and not to tool usage itself. A sensible enforcement regime might look like this:

• No sanction for undisclosed AI use in writing or language editing, absent evidence of misconduct or deception.
• Corrections or expressions of concern where AI involvement contributed to minor errors that do not invalidate results or conclusions.
• Retractions where AI use resulted in fabricated or hallucinated data, irreproducible analyses, falsified methods, or deceptive representation of results or provenance.

In such cases, AI is not the misconduct; it is merely the instrument. The violation remains what it has always been: a breach of research integrity. In cases where major research integrity issues are identified that include improper AI use, the matter should be pursued to the fullest extent.

Crucially, responsibility-based attestations give publishers a clear foundation for post-publication action without resorting to intrusive pre-submission surveillance, unreliable AI detection tools, or speculative judgments about authors’ workflows.

Why This Approach Encourages Transparency

When disclosure is framed as limited, routine, non-punitive, and clearly disconnected from stylistic judgment, authors have far less incentive to hide AI use. Transparency increases not because authors are forced to comply, but because honesty no longer feels risky.

This mirrors what we have learned slowly and imperfectly from data sharing, preregistration, and open methods: compliance follows clarity, incentives, and trust, not ever-expanding documentation requirements.

Conclusion: Governing Outcomes, Not Tools

AI has exposed weaknesses in scholarly publishing but it did not create them. The current disclosure impasse is less about technology than about misaligned incentives and misplaced attention.

Publishers do not need to legislate every interaction between authors and machines. They need to reaffirm something simpler and more durable: that authors are accountable for the work they submit, that reproducibility remains non-negotiable, and that transparency should be rewarded rather than punished.

The path forward lies not in tracking AI use more aggressively, but in doubling down on the foundations of scholarly trust that AI has not replaced and cannot.

Author’s note: I wrote a rough draft, including a mix of bullet points and more fleshed-out narrative paragraphs. I then ask GPT to turn my rough ideas into an essay draft form. I then reviewed the output, made substantive revisions, sent to colleagues for feedback, and then sent it off for publication.