Science is founded upon data that are reproducible, reliable, and authentic. However, generative artificial intelligence (GenAI) technologies are blurring the line between authentic data that are casually linked to real phenomena and synthetic data that are created by computer algorithms. The ability of individuals or organizations (such as paper mills) to use GenAI tools to produce realistic-looking data and misrepresent them as genuine threatens the integrity and trustworthiness of the research enterprise. Data attestation statements can play an important role in helping to protect science from the emerging threat posed by GenAI synthetic data and images.

GenAI’s ability to mimic real data has advanced dramatically in the last few years.Reference Resnik¹ GenAI tools can now produce realistic radiologic, fluorescence, and microscopy images; tabular data and data tables; electronic health records; and time-series records.Reference Kim² Studies have shown that radiologists have difficulty distinguishing between synthetic chest radiographs and genuine radiographic images and most people cannot tell the difference between AI-generated images of human faces and real images of faces.Reference Gibney³

Since GenAI synthetic data have become highly realistic, the scientific community now faces a data authenticity crisis. Provenance, that is, the traceable history of how something came to be and reached its current state, is essential for distinguishing between authentic and synthetic data.Reference Resnik⁴ Authentic scientific data can be linked to real phenomena observed by human beings or recorded by machines.Reference Resnik⁵ Original research records provide evidence of data provenance and are therefore essential for determining whether data (or images) are real or have been fabricated or manipulated.Reference Shamoo and Resnik⁶ Sequestering research records (i.e., seizing them and keeping them safe) is one of the steps that an institution takes when it initiates a research misconduct inquiry. The purpose of sequestration is to preserve provenance by preventing original records from being tampered with or destroyed. It also serves as the trigger for the person accused of misconduct (the respondent) to state what happened to the original data. In addition to helping to assure data integrity, preservation of research records helps to promote other aspects of good science, including quality control, data sharing, collaboration, and replication of results.

Today, distinguishing between GenAI data and images and authentic ones will be increasingly difficult for institutions, funding organizations, or journals. Examination of the research records, especially the original data, usually plays a crucial role in determining whether fabrication or falsification has occurred. However, if the data or images have been produced by GenAI, there will be no records that an institution could use to authenticate them. When this happens, the respondent could claim that the records have been lost or destroyed. If loss or destruction of records is not treated as evidence of research misconduct, as is the case in current US Public Health Service regulations, the investigating institution could find itself in a difficult position when a researcher claims that they have lost the records.⁷ Although losing research records may be reckless behavior, it does not constitute misconduct unless the institution can independently demonstrate, by forensic analysis or other means, that fabrication or falsification occurred.

Thus, not treating failure to produce research records as evidence of misconduct creates a legal loophole that undermines the enforcement of research misconduct policies. The point here is not that every missing record indicates misconduct; records do sometimes get lost or corrupted. The point is that deceptive use of powerful GenAI tools, combined with weak legal consequences for absent records, may create a new form of plausible deniability.

Technical solutions, such as provenance frameworks and blockchain-based certification systems, can partially address this problem.Reference Missier, Belhajjame and Cheney⁸ However, these tools have technical and practical limitations. It can be very costly to implement and maintain the information technology infrastructure required to operate these systems, making them infeasible for smaller institutions. More importantly, provenance systems only certify documented history, not the authenticity of the underlying data itself. If fabricated data are introduced before provenance capture begins, a provenance system may faithfully preserve the history of a fraud rather than prevent it.

A stronger response to the data authentication problem the scientific community now faces would be to combine technical provenance systems with enforceable attestations to the authenticity of the data or images. Attestations are used by most journals to ensure that submitted manuscripts are original, have not been previously published, and contain no material for which required copyright permissions have not been secured. We propose that in light of GenAI impacts on data integrity, scientific organizations, including journals, academic institutions, and research sponsors, should consider using attestation statements to help ensure that submitted data or images are authentic. To help promote research integrity, these statements could also include statements about use of GenAI, recordkeeping, and data sharing. (See Box 1 for an example of a data attestation statement.)

Some journals already require authors to submit original images and deposit data in publicly accessible websites. What is needed now is a more comprehensive accountability framework that ties publication or funding to a documented commitment that the underlying evidentiary record exists and remains available for review. In this way, data attestation would not replace provenance but would give it institutional oversight and legal force. Attestation shifts research integrity policy from post-hoc suspicion to prospective accountability. If concerns emerge about stand-alone published data or data associated with a manuscript, journals and institutions would no longer be left asking whether the necessary material ever existed. Authors would already have attested that it did and that they can provide it, so the burden of proof would fall on them to show that their data are authentic.

Such a framework would not eliminate misconduct, nor would it be a substitute for fair investigation of misconduct allegations. However, it would close a legal loophole in which plausibility and guile can override authenticity and accountability. The response to the data integrity issues created by GenAI tools should not be to reject them outright but to strengthen provenance standards for scientific data, require clearer stage-specific disclosure of GenAI use, and adopt submission requirements that preserve the link between published evidence and raw data. Without these links, scientific credibility becomes easier to simulate than to earn.