AI as a Scientific Auditor: What Can Actually Be Verified?

Getting your Trinity Audio player ready...

Artificial intelligence can already perform useful scientific audits, but the word verify must be used carefully. An AI system can check whether a paper’s citations exist, reproduce calculations, execute available code, compare reported values, and identify contradictions or suspicious patterns. It cannot usually establish that an experiment really occurred, that unpublished data are authentic, or that a scientific theory correctly describes reality.

The practical distinction is this:

AI can verify properties of the available scientific record. It cannot automatically verify the reality that the record claims to represent.

This limitation does not make AI auditing unimportant. Most scientific publications contain numerous claims that are computationally, logically, or documentarily testable. Automating these checks could make research evaluation faster, cheaper, and more systematic—provided that an AI audit is not confused with proof of scientific truth.

What Is a Scientific Audit?

A scientific audit is a structured examination of the evidence, reasoning, methods, and research artifacts supporting a scientific claim.

Unlike ordinary peer review, an audit does not need to judge whether a paper is fashionable, interesting, or likely to be highly cited. It asks narrower questions:

• Do the cited sources support the claims attributed to them?
• Do the calculations produce the reported results?
• Is the statistical analysis appropriate and internally consistent?
• Can the supplied code run on the supplied data?
• Are figures, tables, methods, and conclusions mutually consistent?
• Is there evidence of fabrication, falsification, plagiarism, or manipulation?

The U.S. Office of Research Integrity defines research misconduct primarily in terms of fabrication, falsification, and plagiarism. These categories illustrate both the value and the limits of automated auditing: plagiarism may often be detected from documents, whereas fabrication may remain invisible unless the auditor has access to original records or independent observations.

What AI Can Verify Directly

Citation Existence and Bibliographic Accuracy

AI can check whether cited papers, datasets, books, software packages, and identifiers actually exist. It can compare titles, authors, publication dates, journal names, DOIs, and quoted passages against authoritative databases.

This is increasingly important because generative AI can produce plausible-looking but nonexistent references. Reports of hallucinated citations entering scientific literature show why citation validation should be an automatic preliminary check rather than an optional editorial service.

A citation audit can verify:

• whether the cited work exists;
• whether its metadata are correct;
• whether a quotation appears in the source;
• whether the source discusses the attributed topic;
• whether the citation has been retracted or corrected.

Existence, however, is not the same as support. Determining whether a source genuinely justifies a complex claim may require technical interpretation.

Arithmetic and Symbolic Calculations

Many errors in scientific papers are mechanically detectable. AI combined with calculators, computer-algebra systems, theorem provers, or domain-specific software can examine:

• arithmetic operations;
• algebraic transformations;
• unit conversions;
• dimensional consistency;
• propagation of uncertainty;
• numerical approximations;
• equations copied between sections;
• values transferred from tables into prose.

These are among the strongest candidates for automated verification because the expected result is often precisely defined.

An ordinary language model should not be trusted to perform this work entirely in its generated text. A stronger auditor delegates calculations to deterministic tools and records the commands and outputs.

Statistical Consistency

An AI auditor can inspect whether reported statistics are mutually compatible. For example, it may check relationships among:

• sample sizes;
• means and standard deviations;
• confidence intervals;
• test statistics;
• degrees of freedom;
• p-values;
• regression coefficients;
• effect sizes;
• percentages and absolute counts.

It can also flag multiple-comparison problems, inappropriate tests, missing-data inconsistencies, implausible distributions, and conclusions that overstate the reported evidence.

Some of these checks are exact. Others depend on methodological judgment. AI may reliably detect that a p-value does not correspond to the stated test statistic, while being less reliable when deciding whether the entire statistical model was scientifically appropriate.

Computational Reproducibility

When authors publish complete code, data, parameters, software versions, and execution instructions, an automated system can attempt to reproduce the reported output.

It can ask:

1. Does the code execute?
2. Does it use the documented data?
3. Does it generate the reported tables and figures?
4. Are random seeds and dependencies specified?
5. Do reasonable environment changes alter the result?
6. Are important processing steps hidden or undocumented?

A large-scale study of research code found that execution and reuse depend on practical details such as documentation, dependencies, licensing, and data accessibility. This is why merely attaching a source-code archive is not equivalent to achieving computational reproducibility.

Formal code review is already used by some journals when computational models are essential to the conclusions.

AI can substantially reduce the cost of such review by constructing environments, tracing errors, comparing outputs, and explaining discrepancies. Yet an AI-generated “successful reproduction” should include a machine-readable execution record. Otherwise, readers are merely being asked to trust another generated statement.

Internal Consistency

Scientific manuscripts often contradict themselves without containing any single obviously false sentence. AI is well suited to comparing information across a long document.

It can detect situations in which:

• the abstract reports a different sample size from the methods;
• figure labels disagree with the main text;
• inclusion criteria change between sections;
• a variable is defined in incompatible ways;
• reported dates or experimental groups do not match;
• the conclusion claims an effect not shown in the results;
• supplementary files use different parameters from the paper.

This is a document-level verification task. Large language models can be useful because they process semantic relationships rather than searching only for identical strings.

Reporting Completeness

An AI auditor can compare a manuscript against a reporting checklist or preregistered protocol. It can flag missing information about:

• participant selection;
• randomization;
• blinding;
• exclusions;
• outcome definitions;
• data preprocessing;
• model selection;
• conflicts of interest;
• code and data availability;
• deviations from a preregistration.

Nature Portfolio, for example, explicitly connects reporting standards and access to research materials with transparency and reproducibility.

This does not prove that the disclosed procedure was followed. It verifies that required information was reported and that the report is consistent with other accessible records.

Textual Reuse and Possible Plagiarism

AI can compare wording, argument structure, equations, figures, code, and unusual sequences of ideas across publications. It can identify probable unattributed reuse even where simple phrase matching fails.

The result should normally be treated as a lead, not a verdict. Legitimate overlap can arise from standard terminology, reused methods, common templates, translations, preprints, or an author’s earlier work. Establishing plagiarism requires attribution analysis and contextual judgment.

Image and Figure Anomalies

Computer vision systems can flag:

• duplicated image regions;
• repeated experimental panels;
• suspicious rotations or reflections;
• inconsistent backgrounds;
• abrupt boundaries;
• reused microscopy images;
• discrepancies between plotted points and reported values.

These systems may detect evidence compatible with manipulation, but they usually cannot establish intent. Image processing may be legitimate, accidental, negligent, or deceptive. The original files and laboratory records may be needed to distinguish these cases.

Provenance and Version History

Where cryptographic signatures, trusted timestamps, laboratory systems, repositories, or blockchains are used, AI can inspect chains of custody and document history.

It may verify that:

• a file existed before a stated date;
• a dataset has not changed since its hash was registered;
• a particular account uploaded a particular artifact;
• the manuscript refers to the same version of the data that was analyzed;
• modifications occurred after preregistration.

This verifies digital provenance, not physical truth. A cryptographically authentic file can still contain fabricated measurements.

What AI Can Only Flag as Suspicious

Some audits produce risk indicators rather than definitive answers.

AI may identify:

• data that appear too regular;
• distributions inconsistent with an asserted process;
• improbable similarities between independent experiments;
• unusually repeated decimal patterns;
• citations clustered around irrelevant or coordinated sources;
• text that appears automatically generated;
• conclusions that seem stronger than the evidence;
• images with signs of editing;
• implausibly rapid or extensive research output.

These findings can prioritize human investigation. They should not automatically trigger punishment.

A rare pattern is not necessarily fraudulent. A surprising dataset may reflect a real phenomenon, an unusual instrument, preprocessing, a transcription error, or a poorly documented method. When systems are deployed at scale, even a low false-positive rate can wrongly accuse many researchers.

The appropriate output is therefore not simply fraud or no fraud, but something closer to:

“This pattern is inconsistent with the stated model under these assumptions. Here are the relevant records, calculations, alternative explanations, and recommended checks.”

What AI Cannot Verify from a Paper Alone

Whether an Experiment Actually Happened

A publication can describe a perfectly coherent experiment that never occurred. Unless an auditor has access to independent records—instrument logs, samples, timestamps, laboratory notebooks, procurement records, witnesses, or replication attempts—the paper itself may contain no contradiction.

AI cannot infer physical occurrence merely from a convincing description.

Whether the Raw Data Are Genuine

AI can test whether data are plausible. It generally cannot determine whether plausible data were measured, simulated, copied, selectively edited, or invented.

Provenance systems improve the situation, but only when evidence is captured during the research process. Retrospective auditing cannot reconstruct records that were never created.

Whether Every Relevant Observation Was Disclosed

Published data may be authentic but selectively chosen. Researchers can omit failed experiments, inconvenient variables, participants, alternative models, or negative results.

Preregistration, registered reports, automatic instrument logging, and complete data deposits make selective reporting easier to detect. Without these mechanisms, an AI auditor sees only what it is given.

Whether a Claim Is True in the Real World

Reproducing a calculation verifies that the calculation follows from the inputs. It does not prove that the inputs represent nature or that the underlying assumptions are correct.

Likewise, reproducing an analysis is not the same as replicating an experiment. A replication collects new evidence. Computational reproduction normally reprocesses existing evidence.

This distinction is essential:

• Verification checks a claim or procedure against specified evidence and rules.
• Reproducibility checks whether the reported analysis can be repeated with the same materials.
• Replication tests whether a result appears again using new observations.
• Scientific truth is a broader conclusion supported by cumulative evidence, criticism, and successful prediction.

AI can assist with all four, but it cannot collapse them into one automatic judgment.

Whether a Research Question Is Important

Importance is partly normative and context-dependent. An auditor can map dependencies, estimate possible applications, identify neglected fields, or compare expert assessments. It cannot derive a uniquely correct social value from the manuscript.

This matters for research funding. A system such as AI Internet-Meritocracy should distinguish audit findings from evaluations of scientific merit. Methodological correctness, originality, usefulness, researcher contribution, and societal priority are related but different dimensions.

Whether the Authors Acted Intentionally

Research misconduct generally involves more than the presence of an error. Federal policy distinguishes misconduct from honest error and requires attention to whether conduct was intentional, knowing, or reckless.

AI may reconstruct what changed and when. It usually cannot directly determine a person’s state of mind. Intent must be inferred from records, explanations, repeated behavior, and institutional investigation.

AI Auditing Works Best When Research Is Auditable by Design

The greatest obstacle may not be inadequate artificial intelligence. It may be inadequate scientific infrastructure.

An AI auditor becomes much more effective when a project provides:

• persistent identifiers for people and artifacts;
• complete source data;
• executable code;
• declared software environments;
• timestamped preregistrations;
• machine-readable methods;
• version histories;
• structured contribution records;
• laboratory or instrument logs;
• explicit links between claims and supporting artifacts.

Without these materials, even a highly capable AI is forced to speculate. With them, many audit tasks become deterministic.

This suggests a policy principle:

Do not merely ask AI to understand scientific papers better. Publish science in forms that allow claims to be traced to evidence and computations.

Open and transparent models may also be preferable for high-stakes auditing because researchers need to know how evidence was retrieved, processed, and scored. Proprietary systems whose prompts, versions, retrieval sources, or decision procedures cannot be reconstructed create an additional reproducibility problem.

A Layered Model of AI Scientific Auditing

A credible auditing system should separate several levels of assurance.

Level 1: Documentary Checks

The system checks citations, identifiers, disclosure forms, metadata, terminology, and consistency among manuscript sections.

These checks are relatively inexpensive and can be applied to almost every submission.

Level 2: Analytical Checks

The system recomputes statistics, validates equations, extracts plotted values, examines assumptions, and checks whether conclusions follow from reported results.

This requires access to structured data or reliable extraction from the paper.

Level 3: Computational Reproduction

The auditor executes the authors’ code and attempts to regenerate the outputs. Failures are documented with environment information and error traces.

Level 4: Provenance Audit

The system examines original data histories, timestamps, instrument records, preregistrations, repository logs, and cryptographic commitments.

Level 5: Independent Replication

New researchers or automated laboratories repeat the relevant procedure using new samples or observations.

This is no longer merely a document audit. It is new scientific work.

Level 6: Cumulative Scientific Assessment

The result is compared with independent studies, alternative models, theoretical constraints, and evidence from related fields.

At this level, AI supports scientific judgment but does not replace the scientific community.

Why AI Should Not Be the Final Judge

Language models can generate unsupported claims, misread technical sources, follow malicious instructions, and produce highly confident but unstable assessments. Automated benchmarks themselves require controls for validity, transparency, and reproducibility—a problem emphasized in recent NIST work on language-model evaluation.

An auditing system therefore needs its own audit trail:

• the precise model and version used;
• the complete task instructions;
• retrieved evidence;
• tool calls and calculations;
• confidence or uncertainty estimates;
• alternative interpretations;
• independent model outputs;
• a procedure for challenge and appeal.

The broader issues are related to AI alignment and become more consequential as systems approach greater autonomy or possible superintelligence. A powerful evaluator that can influence funding, publication, or reputation must be resistant not only to ordinary mistakes but also to manipulation and concentrated control.

COPE maintains that human authors remain responsible for manuscript content produced with AI tools. The same principle should apply in reverse: organizations using AI audits remain responsible for the decisions made from them.

The Role of Human Review and Voting

Human review is still necessary where an audit involves context, intent, disputed assumptions, unpublished records, or sanctions against a researcher.

Humans are not automatically unbiased or accurate. They can follow authority, protect institutions, misunderstand unfamiliar work, or participate in conflicts of interest. But a properly designed collective process has one advantage over a single language model: participants can introduce genuinely independent evidence, interpretations, and moral judgments.

In AIIM, human voting can act as a governance layer for contested cases, especially bans, unbans, suspected manipulation, and appeals. The AI can assemble evidence and explain the inconsistency, while humans decide whether the evidence justifies a consequential action.

The useful division is not AI versus humans. It is:

• machines for exhaustive, repeatable checks;
• specialized tools for deterministic computation;
• independent experts for technical interpretation;
• broader governance for contested institutional decisions;
• replication for claims about physical reality.

AI Auditing Should Produce Evidence, Not Authority

The strongest AI scientific auditor would not imitate an all-knowing journal editor. It would create an inspectable map from claims to evidence.

For every significant conclusion, it should show:

1. the exact claim being evaluated;
2. the artifacts available to support it;
3. the checks that were performed;
4. which results were reproduced;
5. which assumptions were required;
6. which discrepancies were detected;
7. what remains unverifiable;
8. what additional evidence would resolve the uncertainty.

This output is more useful than a single trust score. A score compresses uncertainty and may conceal disagreement among checks. An evidence map allows researchers, funders, reviewers, and readers to inspect the reasons.

Conclusion

AI can already verify a substantial portion of the formal scientific record: citations, calculations, statistical relationships, code execution, data consistency, reporting completeness, provenance, and agreement between claims and accessible evidence.

It can also identify patterns that deserve investigation. But suspiciousness is not proof of fraud, computational reproduction is not experimental replication, and document consistency is not scientific truth.

The correct objective is therefore not an AI that declares papers “true” or “false.” It is an auditing infrastructure that makes each scientific claim more traceable, testable, reproducible, and open to challenge.

Used this way, AI does not replace science. It reduces the amount of scientific evaluation that must depend on reputation, limited reviewer attention, or institutional trust.