A custom AI reviewer for ethics and data sharing statements

PerfectPaper’s ethics agent reads a manuscript for the compliance statements editors check before review and funders check after publication: ethical approval naming an approving body, consent whose scope actually covers the sharing being proposed, de-identification claims that match what the data contain, and a data availability plan consistent with funder policy. Copy the brief below into a custom agent slot.

These are the least interesting findings in any review and the worst ones to receive late. The consent-scope item is the one that occasionally stops a paper: data cannot be deposited under a consent that did not contemplate deposition.

When to use this agent

The ethics agent earns its slot whenever the manuscript asserts a permission — approval, consent, deposition — that a reader outside your group cannot verify from the text as written.

• Your study involves human participants, human tissue, or animals
• You plan to deposit human genomic or individual-level data
• Consent was obtained years before the current analysis
• Your funder requires a data management and sharing plan
• The manuscript combines cohorts consented under different protocols
• The corresponding author did not personally hold the protocol under which the data were collected
• A previous submission was returned at editorial screening with a query about approval, consent or data availability

The last two describe most of the failures. A compliance statement is usually written by whoever is drafting the manuscript, from memory, about a protocol approved years earlier by someone else — and the approving body, the protocol identifier, and the consent language are exactly the details memory reconstructs wrongly. Ethics gaps are a leading reason a paper is desk rejected without review, and unlike a statistical objection they carry no argument you can win on revision.

The agent brief

The brief below covers approval, consent, consent scope for sharing, de-identification, controlled access, the data management plan, vulnerable populations, and competing interests in one pass. Paste it into a custom agent slot. Suggested settings: work type verification, skill level graduate, no tools required.

Name: Ethics, consent and data sharing

You are reviewing a manuscript for research ethics and data sharing
statements. Read the methods, the declarations, and any data availability
statement together.

Determine and report:

1. Approval. Determine whether an ethical approval statement exists naming the
   approving body and, ideally, a protocol identifier. For animal work,
   determine whether an institutional animal care committee approval is named.
   Report an approval statement that names no body, and report the absence of
   any statement where the work plainly requires one.

2. Consent. Determine whether informed consent is described, whether it was
   written, and whether a waiver was granted with a stated reason. For
   secondary analyses, determine whether the original consent is described at
   all. Report where consent is asserted without any account of what
   participants agreed to.

3. Consent scope for sharing. Where the manuscript proposes depositing
   individual-level or genomic data, determine whether the described consent
   covers broad sharing, controlled-access sharing, or neither. Report a
   deposition plan that exceeds the described consent scope, and state that
   this cannot be resolved after the fact. Treat this as the highest priority
   item in this brief.

4. De-identification. Determine what identifiers were removed and whether the
   remaining variables could re-identify participants in combination: rare
   diagnoses, small geographic units, exact dates, ages above ninety. Report a
   claim of de-identification alongside variables that undermine it.

5. Controlled access. Where human genomic data are deposited, determine
   whether the repository chosen matches the consent: an open repository for
   broadly consented data, a controlled-access repository otherwise. Report
   mismatches.

6. Data management plan. Determine whether the manuscript states what will be
   shared, where, when and under what conditions. Report a statement that
   promises availability without naming a repository or a timeline.

7. Vulnerable populations. Where participants are children, prisoners, or
   people unable to consent for themselves, determine whether the additional
   protections are described.

8. Competing interests and funding. Determine whether both are declared, and
   report the absence of either.

Report each gap with its location and the specific statement required. Be
brief; these findings are checklist items rather than arguments.

Item 3 is marked highest priority inside the brief on purpose. Without that instruction, a model orders findings by how much text it can generate about them, which puts de-identification and data management first — they are discussable — and buries the one item that is not discussable. The instruction to “state that this cannot be resolved after the fact” exists for the same reason: the default output for a permission gap is “consider clarifying the consent statement”, which is wrong advice, because there is nothing to clarify when the permission was never granted.

Why consent scope is the item that stops a paper

Consent scope stops a paper because a sharing permission is granted at the moment of collection and no later step manufactures it. The Declaration of Helsinki, as revised at the 75th WMA General Assembly in Helsinki in October 2024, requires researchers to obtain informed consent for the collection, processing, storage and foreseeable secondary use of biological material and of identifiable or re-identifiable data. The operative word is foreseeable: a use that was not foreseeable to the participant in 2011 is not covered by the sentence they signed in 2011.

The NIH Genomic Data Sharing Policy, effective for competing applications from the 25 January 2015 due date, is explicit about the same asymmetry: “For studies initiated after the effective date of the GDS Policy, NIH expects investigators to obtain participants consent for their genomic and phenotypic data to be used for future research purposes and to be shared broadly.” It extends that expectation to material you might assume is outside it — for cell lines or clinical specimens created after the effective date, “NIH expects that informed consent for future research use and broad data sharing will have been obtained even if the cell lines or clinical specimens are de-identified.”

A worked instance. A 2012 consent for a single-site cohort says samples “will be used to study diabetes at this institution”. The 2026 manuscript reports whole-genome sequencing and proposes deposition in an open repository. Three separate limitations sit in that one sentence — disease-specific, institution-specific, and silent on genomic data — and each of them independently blocks open deposition. The agent’s job is to surface all three at draft stage, when re-consent, a formal waiver determination, or a narrower deposition route are all still available options.

The two major consent regimes do not require the same degree of specificity, which is a genuine source of cross-border confusion rather than a simple compliance checklist. NIH’s Genomic Data Sharing Policy expects the original consent to name future research use and broad sharing in terms a participant could read and understand. The EU’s General Data Protection Regulation takes a different structural position for scientific research: Recital 33 records that it is “often not possible to fully identify the purpose of personal data processing for scientific research purposes at the time of data collection”, and so allows data subjects to consent to “certain areas of scientific research” rather than to an itemised list of analyses. Both are legitimate consent designs. What is not legitimate is a manuscript that treats them as interchangeable — citing a broad-consent form written for one regime to justify a deposition standard the other regime would require to be stated explicitly. The agent flags the mismatch; it does not decide which regime should have governed the cohort.

Questions about consent scope for data sharing

Can I share my data if the consent form never mentioned data sharing?

Not by default. A sharing permission is granted at collection, and if the form describes only the original study with no mention of deposition or future use, that silence does not become permission later. The available routes are a formal waiver determination, a controlled-access repository matched to the narrower consent, or a data availability statement that names the restriction directly instead of promising sharing the consent does not support.

My IRB approved this study back in 2015 and the consent form never says anything about putting data online — can I still deposit it in a public repository now?

Only if the wording, read plainly, covers broad sharing — most 2015-era single-site consent language does not, since it typically describes use “at this institution” for one stated purpose. Depositing anyway is the exact failure this brief exists to catch. The fixes are a waiver determination, a controlled-access repository under a data use agreement, or an aggregate-only release — not a reinterpretation of the 2015 wording.

The dataset is public and I didn’t collect it — how do I write the ethics statement?

State who approved the original collection, what the consent covered, and under what licence or agreement you obtained the data. Public availability is not a permission that replaces those three facts; it is a distribution route that inherited them. Where the source publishes no ethics statement of its own, say so plainly and name whatever determination your own committee made about the secondary use — the brief’s item 2 flags a re-analysis that says only “data were obtained from [source]” and stops.

Does GDPR’s “broad consent” for research mean I don’t need separate permission to share the data?

No. Broad consent under GDPR Recital 33 covers analyses that could not be specified at collection time, but it is still a bounded scope, not blanket permission, and it does not automatically satisfy a funder such as NIH that expects sharing named explicitly. State which regime governed the cohort: a broad-consent form valid under one framework is not interchangeable with the specific-sharing statement the other requires.

The vocabulary that makes consent scope checkable

Consent scope becomes checkable once it is written as a data use term rather than as prose. The GA4GH Data Use Ontology, approved as a GA4GH standard in 2019 and described in Cell Genomics in 2021, supplies the standard vocabulary: GRU (general research use, DUO:0000042), HMB (health, medical or biomedical research, DUO:0000006), and DS (disease-specific research, DUO:0000007), with modifier terms — not-for-profit use only, non-commercial use only, publication required, collaboration required, geographical restriction — layered on top of whichever of the three applies.

Naming the applicable term in the methods does two things. It tells a data access committee what it may approve without re-reading your consent form, and it forces the author to decide, before submission, which category the consent actually falls in. A common consent-scope failure is DS consent presented as GRU deposition — the study collected samples to study one disease and the manuscript proposes to release them for anything.

The corresponding NIH mechanism is the Institutional Certification, in which a signing official assures that any limitations on research use expressed in the informed consent documents have been delineated, and that submission and subsequent sharing are consistent with what participants agreed to. If your institutional signing official cannot make that assurance, the manuscript’s deposition sentence is a promise your institution will decline to keep — and the place to discover that is at drafting, not at deposit.

Controlled access is not simply slower distribution; the consent-scope question recurs downstream of deposit rather than being settled by it. At NIH’s dbGaP, the primary controlled-access repository for genomic data, a request is routed to the requesting investigator’s own institutional signing official for approval and co-signature, and then to a Data Access Committee, which checks the proposed research against the data use limitations the dataset carries — the limitations that came from the original participants’ consent. Approved access runs for a fixed term rather than indefinitely, after which the investigator renews or files a close-out report, so the same consistency question is put to the same dataset again. Naming the applicable Data Use Ontology term at deposit is what lets that committee answer it quickly, from the term alone, rather than re-reading the original consent form for every future request.

What a data availability statement has to state

A data availability statement has to answer five questions, and ICMJE enumerates them. Since 1 July 2018, manuscripts submitted to ICMJE journals reporting the results of clinical trials must contain a data sharing statement indicating whether deidentified individual participant data, including data dictionaries, will be shared; what data in particular; whether related documents such as the protocol and the statistical analysis plan will be available; when the data will become available and for how long; and by what access criteria — with whom, for what types of analyses, and by what mechanism. The statement is required whether or not data will be shared, so declining to share is an answer and silence is not. Separately, clinical trials that began enrolling participants on or after 1 January 2019 must include a data sharing plan in the trial’s registration record — which means the registry entry and the manuscript can now contradict each other, and reviewers do check.

The NIH Data Management and Sharing Policy, effective 25 January 2023, applies the same discipline to NIH-funded research that generates scientific data, and pushes the question upstream by asking researchers to consider at the outset how sharing will be addressed in the informed consent process. That is the ordering the whole area assumes: sharing terms are settled with participants, then written into a plan, then reported in a manuscript. A statement first drafted to satisfy a journal form at submission is being drafted several years too late, and the agent’s findings are largely the visible symptoms of that inversion.

Accession-level checking — whether the repository, the identifier format, the genome build and the tool versions are actually present — belongs to a different agent; see data availability and accessions, which is briefed on the repository landscape this one is not. For trial-specific consistency between the registry record and the manuscript, see trial registration and endpoints.

How de-identification claims fail

De-identification claims fail on combinations, not on direct identifiers, and the arithmetic is unforgiving. Rocher, Hendrickx and de Montjoye estimated in Nature Communications (2019;10:3069) that “99.98% of Americans would be correctly re-identified in any dataset using 15 demographic attributes”. Removing name, address and record number does not touch that estimate, because none of those were among the 15.

The variables that carry re-identification risk in a supplementary table are the ones authors think of as clinical rather than identifying:

• Exact dates. Date of admission, date of surgery and date of death are each near-unique in combination with a diagnosis and a region.
• Small geographic units. A postcode sector, a census tract or a named clinic narrows the population to a few thousand people before any other variable is applied.
• Rare diagnosis or rare variant. A pathogenic variant with a handful of reported carriers is an identifier in itself, which is why genomic data is treated as inherently re-identifiable rather than as data that can be de-identified.
• Extreme ages. Ages recorded above 89 are conventionally top-coded for exactly this reason.
• Small cell counts. A cross-tabulation cell containing one or two participants discloses those participants to anyone who knows one of the margins.

The agent reports a de-identification claim made alongside variables that undermine it, and it reports the claim’s location, because the remedy differs by location: a supplementary table can be aggregated or top-coded before resubmission, whereas a deposited dataset described in the methods needs a different access route. The same reasoning applies to registry and electronic health record analyses, where the deposited extract is frequently richer than the analysis actually required.

“De-identified” is not one fixed bar across every framework that uses the word, and the gap between the strictest and loosest readings is exactly the combination problem above. Some de-identification standards are satisfied once a fixed list of named identifiers — name, address, exact dates, record numbers — has been stripped, regardless of what combinations remain in the file. Other guidance frames the question as whether re-identification is reasonably likely at all: GDPR Recital 26 asks what “all the means reasonably likely to be used” by any party would achieve, weighed against the cost, the time required and the technology available at the time of processing — which is the standard a determined combination attack actually has to clear. A dataset that passes an identifier-removal checklist can still fail the combination-aware test with every remaining column untouched, which is why the agent checks what remains rather than what was removed.

Common questions about de-identification

Is stripping names and addresses enough to call a dataset de-identified?

No. Rocher and colleagues (Nature Communications, 2019;10:3069) estimated that 99.98% of Americans would be correctly re-identified in a dataset built from 15 ordinary demographic attributes, none of which need be a name or address. Exact dates, small geographic units, rare diagnoses and extreme ages carry re-identification risk that direct-identifier removal never touches.

I removed every direct identifier from my supplementary table — why would a reviewer still say it could re-identify patients?

Because re-identification risk comes from combinations, not single fields. A rare diagnosis, an exact surgery date and a three-digit postcode together can narrow a population to one person even with no name attached anywhere in the table. The usual fix is aggregation or top-coding — widening date ranges, generalising geography, or suppressing small cell counts — applied to the specific columns the reviewer named.

How many variables does it take before an “anonymous” dataset stops being anonymous?

There is no fixed count; it depends which variables are involved and how rare their combination is in the underlying population. The often-cited benchmark — 15 demographic attributes reaching 99.98% re-identification — used ordinary fields like birth date, sex and postcode, not exotic ones, which is why the agent checks combinations of clinical and geographic variables rather than counting how many fields a table contains.

Combined cohorts, secondary analyses and inherited permissions

Combined cohorts fail differently from single-protocol studies: the permissions do not merge, they intersect. A pooled analysis of three cohorts can share only what the most restrictive of the three consents allows, and a manuscript that describes a single deposition for the pooled dataset is almost always describing a permission that only one or two of the contributing studies granted.

One route around the intersection problem avoids pooling the individual-level records at all. Federated analysis platforms such as DataSHIELD compute the requested statistics at each contributing site and return only the aggregate result, so no site’s records ever cross that site’s own consent boundary; the analysis runs under each cohort’s original permission rather than under a newly negotiated joint one. This relocates the ethics question from “what can we deposit” to “what can we compute without moving anything” rather than eliminating it, and the methods still need to state which model was used and which site retained which records.

Secondary analyses carry a second, subtler problem. The manuscript typically says “data were obtained from [cohort]” and stops, which leaves the reader unable to determine whether the original consent contemplated this analysis at all. The brief’s item 2 targets exactly this: for secondary analyses, determine whether the original consent is described at all. The required sentence is short — who approved the original collection, what the consent covered, and under what agreement the current team received the data — and its absence is a frequent ethics finding in re-analysis manuscripts.

Where the pooled cohorts differ systematically in who was recruited and under what protections, the permission question and the interpretation question converge; see health equity reporting for the reporting side of the same problem. Qualitative and mixed-methods components add a further constraint, because verbatim quotations can identify a participant in a small purposive sample even when every named field has been stripped.

Animal work and vulnerable populations

Animal work is checked against a different statement than human work: the named institutional animal care and use committee, its approval identifier, and compliance with the ARRIVE 2.0 Essential 10, whose item 8 is “Experimental animals” and whose companion explanation was published in PLOS Biology in 2020. An approval sentence that names a country’s legislation but no committee is the usual gap, and it is caught at screening rather than at review. See the ARRIVE guidelines for the item list, and in vivo rigor for the design questions that sit alongside them.

Vulnerable populations require the manuscript to describe the additional protection, not merely to note the population. For children, that means parental permission plus age-appropriate assent, and for a longitudinal paediatric cohort it means stating whether participants were re-consented on reaching the age of majority — a step that is frequently performed and almost never reported. For adults unable to consent for themselves, it means naming the legally authorised representative mechanism and any assent process. For prisoners, employees and students, it means addressing the coercion question directly rather than asserting voluntariness. Assent thresholds are not uniform across jurisdictions or even across ethics committees within one country, so the agent does not assume a specific age; it reports whether an age-appropriate assent process is described at all, and leaves the applicable threshold to the committee that reviewed the protocol.

The agent reports the absence of the protection statement, not the ethics of the study. That distinction matters: the finding is always “the manuscript does not describe X”, which is a fixable reporting gap, and never a judgement about whether the study should have been conducted.

Questions about animal ethics and vulnerable participants

Does my IACUC approval statement need a protocol number, or is naming the committee enough?

Naming the committee is necessary but not sufficient by itself, though the finding that stops a paper at editorial screening is usually the missing committee, not a missing number. A protocol identifier is the stronger form and sits alongside ARRIVE 2.0’s “Experimental animals” item, but an approval sentence that names a country’s legislation and no committee at all is the gap the agent flags first.

Do I need to re-consent kids in my study once they turn 18?

For a longitudinal paediatric cohort, most ethics protocols expect re-consent or re-assent at the age of majority, and it is a step research teams frequently complete but rarely describe in the manuscript. The agent looks for that statement specifically, because its absence reads as a reporting gap in the text even in cases where the re-consent itself was properly carried out.

How the agent reads the manuscript

The agent reads three parts of the document against each other rather than checking each in isolation, which is why the brief’s first instruction is to read methods, declarations and the data availability statement together. Nearly every finding in this brief is a contradiction between two sections that are each internally plausible.

The three cross-references that produce most findings:

• Methods against declarations. The methods describe collecting a variable the declarations do not cover, or describe recruiting a population whose protections are never mentioned.
• Methods against the data availability statement. The statement promises “all data” while the methods describe a linked administrative dataset the team does not own and cannot redistribute.
• Data availability statement against consent. The statement names an open repository while the consent, as described, is disease-specific or institution-specific.

A generic model prompt does not find these, because each section reads correctly on its own and the defect exists only in the relationship between them. PerfectPaper reads the whole manuscript on every review rather than sampling sections, which is the precondition for a cross-reference check of this kind; on the confidentiality question that this raises for an unpublished manuscript, see secure AI for unpublished research.

What editors and reviewers actually write

Editors and reviewers do not write “ethics problem”. They write the specific missing sentence, and these are the recurring forms:

“The ethics statement does not name the approving committee or a protocol number.” “Please state whether consent was written or oral, and if a waiver was granted, by which body and on what grounds.” “The data availability statement promises deposition, but the described consent does not appear to permit sharing beyond the original study.” “Individual-level genomic data cannot be deposited in an open repository; please confirm the consent permits this or move to a controlled-access repository.” “The authors state the data are de-identified; Supplementary Table 3 reports exact dates of diagnosis and three-digit postcode.” “For the secondary analysis, please describe the original consent and the agreement under which the data were transferred.” “Competing interests are declared as none for all authors; the funding statement names an industry sponsor whose role is not described.” “Data are stated to be available on request from the corresponding author, which does not meet the journal’s policy.”

The last one is the most common and the most avoidable. Availability on request is treated as non-compliant by a growing number of journal policies; see journal policies for AI and reporting.

What to do when the obvious remedy is unavailable

The obvious remedy — re-consent — is unavailable in most of the cases the agent flags, because the cohort is closed, dispersed or deceased. Five routes remain, and a manuscript is far better served by naming one of them than by restating the promise:

1. Seek a formal waiver determination. An ethics committee can determine that secondary use without re-consent is permissible; the manuscript then names the committee and the determination rather than asserting that consent covered it.
2. Move to controlled access with a data use agreement. A controlled-access repository plus a data access committee can honour a narrower consent that an open repository cannot, and the deposition sentence changes from “openly available” to “available under [access mechanism]”.
3. Deposit metadata only, with a persistent identifier. A record describing the dataset, its variables and its access contact is citable and discoverable even when the data themselves cannot move. This is a genuine deposition, not a placeholder.
4. Release derived or aggregate data. Summary statistics, aggregated tables or model coefficients often answer the reproducibility question without individual-level release. State what was aggregated and at what granularity, and note that some aggregate outputs — allele frequencies in small cohorts, for instance — still carry disclosure risk.
5. State the restriction plainly. A data availability statement that says the data cannot be shared, names the specific legal or consent restriction, and gives the contact for access enquiries satisfies ICMJE’s fifth component. A statement that says data are “available on request” satisfies nothing, because it names neither the criteria nor the mechanism.

Route 5 is the one authors resist and the one editors accept most readily. Naming a restriction is a specification; promising availability you cannot deliver is a commitment that fails at the first request.

What this agent does not do

The ethics agent reports gaps in what a manuscript states, not gaps in the underlying research. It does not determine whether a study was ethically conducted, does not verify that a named protocol identifier exists, does not contact an ethics committee or a repository, and does not substitute for your institution’s research compliance office or a data protection assessment. Every finding takes the form “the manuscript does not state X, and X is required here” — which is a claim about the document, verifiable by reading the document.

That boundary is deliberate. A model asked to judge whether consent was adequate will produce a confident judgement it has no basis for, and the resulting finding is unusable by an author and indefensible to an editor. A model asked whether the manuscript describes the consent is answering a question the text can settle.

What this agent catches

The table below lists the ten gaps this brief is written to surface, each mapped to the point in the publication process where it would otherwise surface on its own — usually later, and usually with fewer options left.

How this differs from the built-in review

The standing team includes a submission-requirements specialist that reads journal-facing requirements. This agent covers the compliance layer beneath that: not what the journal’s form asks for, but whether the study as described is consistent with the permissions it was conducted under. The consent-scope check in particular has no counterpart in a general review, and it is the one with consequences that cannot be remedied later.

The difference is visible in what each one flags on the same manuscript. A submission-requirements check reports that the data availability statement is missing from the required position in the file. This agent reports that the statement, wherever it sits, promises open deposition of individual-level data collected under a disease-specific consent. The first is a formatting finding fixed in a minute; the second changes what the paper can claim to offer.

Applies across every field, which is why it sits with the general agents rather than in a field cluster. Field-specific deposition questions still belong to their own clusters: the repository landscape for omics data sits in genomics review, and participant-level trial data sharing sits in clinical trials review.

Review my manuscript

Related agents

Three agents share a review, so the ethics brief runs alongside two others. See Custom AI reviewer agents for the full library, organised by research area.

• How to write your own — use it when you are writing your own brief rather than pasting one.
• Overclaim check — use it for the distance between what was shown and what the abstract says.
• Study section lens — use it when you need the grant reviewer’s question, not the referee’s.

Frequently asked questions

How do I check my ethics and data sharing statements before submission?

Read the methods, the declarations and the data availability statement against each other, and confirm that every permission the manuscript asserts — approval, consent, deposition — is described somewhere in the text rather than assumed. The agent brief above automates that cross-reference and reports each gap with its location and the specific sentence required.

Can an AI reviewer check ethical approval and consent statements?

An AI reviewer can check whether the manuscript states approval, names the approving body and a protocol identifier, describes what participants agreed to, and matches the proposed deposition to that description. It cannot verify that a named protocol identifier exists, contact a committee, or judge whether a study was ethically conducted — so this agent is briefed to report document gaps only.

I’m about to submit and I’m worried I’ve missed something in my ethics section — can an AI actually catch that before a reviewer does?

It can, for gaps checkable from the text: an approval statement missing its approving body, a consent description silent on future use, a deposition plan wider than the described consent, or a data availability statement naming no repository. It cannot verify the underlying facts are true — that a protocol number is real, or that a study was conducted ethically — only that the manuscript states what it needs to.

Can an AI reviewer check my data availability statement against journal requirements?

The agent checks the statement for the five things ICMJE enumerates: whether deidentified individual participant data will be shared, which data, whether related documents such as the protocol will be available, when and for how long, and by what access criteria and mechanism. A statement offering data “on request from the corresponding author” names neither criteria nor mechanism, and the agent flags it.