Responsible Use of AI Image Detectors in Research Integrity Investigations

Short answer: use a detector score as a lead, never as the finding

Use an AI image detector only to decide where to look next, never to decide what happened. A score is a probability from a model, not evidence of how an image was made. Corroborate it with source files, independent methods and expert review, give the author a chance to respond, and keep unpublished or sensitive images off third-party services unless the terms allow it.

This guide is for research integrity officers, journal editors and their staff, lab heads and institutional committees. It covers policy and process for one class of tool. It does not teach forensic technique; for that, see CASRAI’s guide to detecting image manipulation in figures.

Third-party tools named on this page are examples only. They are not endorsed, and they have not been validated by CASRAI. The sources behind every claim are listed at the end, with the date each was read (2026-10-30).

What do AI image detectors do, and what do they not do?

An AI image detector is a classifier. You give it an image and it returns a probability, or a label derived from one, that the image was produced by a generative model. It does not recover the history of the file, identify the generator with certainty, or tell you whether a result in the figure is true. The score is a statistical output of a model trained on some set of real and synthetic images.

One example of a free web-based tool is the AI image detector at imagedetector.com. Its home page, read on 2026-10-30, says it returns a probability score for whether an image is AI-generated and accepts JPG, PNG and WEBP files. The home page publishes no accuracy figures and no false-positive or false-negative rates, and it does not say what score separates “AI” from “real”. That is typical of the problem: a score without a published error profile for your kind of image cannot be interpreted.

Three things that are easy to confuse

• AI-image detection asks whether an image looks machine-generated. It is the subject of this guide.
• Duplication and manipulation forensics asks whether regions were copied, spliced or selectively altered, within a figure, between figures or between papers. A fully synthetic blot has no source image to match, which is why duplication screening alone does not answer the AI question. See image duplication for the definition.
• Provenance records where a file came from. C2PA Content Credentials are the best-known standard: the C2PA explainer (version 2.2, 22 April 2025) describes provenance as the facts about the history of a piece of content and says it makes no value judgments about whether provenance data is “true”. It also says that assets without Content Credentials should not be assumed untrustworthy purely for lacking them. A missing credential is therefore not a detection result, and a present one is not proof of authenticity. CASRAI has a separate explainer on C2PA Content Credentials.

For the definition of the underlying technology, see AI image generation.

What are the known failure modes of AI image detectors?

Peer-reviewed and preprint studies describe several failure modes. Which of them applies to your case depends on the detector, the generator and the image type, so treat the list as reasons for caution rather than as measured error rates for any specific tool.

False positives and false negatives on scientific images

Gosselin (PeerJ, 20 February 2025) tested three free web-based detectors on 48 western blot images generated with ChatGPT 4 and 48 authentic blots from articles published before generative AI became widespread. Sensitivity ranged from 0.1875 (Hive Moderation) to 0.9583 (Is It AI?). Specificity ranged from 0.4167 (Illuminarty) to 0.8750 (Hive Moderation). By simple arithmetic from those figures, Illuminarty labelled about 58% of the authentic blots as AI-generated, and Hive Moderation missed about 81% of the generated ones. The author reports a very low positive predictive value at a realistic prevalence of AI-generated images and advises that free detectors be used with great caution in peer review or editorial decisions. The study notes a limit: subscription detectors might perform differently.

The practical lesson is about base rates. If most images you check are authentic, even a detector with decent specificity will produce many false alarms compared with true hits, so a flag is far weaker evidence than the headline accuracy figure suggests.

Resizing and compression

Corvi and colleagues (arXiv preprint, 2022) report that detectors built for GAN-generated images suffer a significant performance drop when image quality is impaired by resizing and compression, because the forensic traces are weak and easily removed by non-malicious processing. In the Gosselin study, reducing the size of the blots lowered sensitivity and raised specificity. In practice, figures in a PDF are usually recompressed and rescaled, so ask for the file as submitted or as acquired, not a screenshot of the article.

New generators

Ojha, Li and Lee (arXiv preprint, 2023, revised 2024) find that real-versus-fake classifiers trained on GAN images fail to detect images from newer generative models. Corvi and colleagues likewise find that generalization is the main hurdle and that detectors trained only on GAN images work poorly on diffusion-model images, calling their results preliminary. A tool that performed well last year may be uninformative against this year’s generator.

Adversarial editing

Carlini and Farid (arXiv preprint, 2020) showed that, for a classifier trained on one generator, attacks with full access to the classifier reduced the area under the ROC curve from 0.95 to as low as 0.0005, and a black-box attack reduced it to 0.22. The practical reading is that a person who wants to evade a given detector can sometimes do so. A low score is not clearance.

Domain shift for microscopy, gels and blots

Most public detectors are built around photographs of scenes and faces. The Gosselin results above are the direct evidence for scientific images: generic detectors performed unevenly on blots, and the author called for detectors trained on scientific content. Whether the same holds for micrographs, gels or plots is an extension we cannot confirm from the sources we read, so validate any tool on a labelled set of your own field’s images before relying on it.

What the publisher bodies say

STM’s page on the feasibility of technical solutions for detecting falsified images (posted 6 February 2025) says generative AI poses significant additional risks and describes AI-generated images as “very hard — if not impossible — to detect with current technologies.” STM’s 2021 recommendations say image screening cannot be expected to uncover all aberrations, and that not all anomalies imply intent to deceive.

What is a responsible step-by-step workflow for an inquiry?

The steps below follow the common structure of the US Public Health Service rules (42 CFR Part 93), the ORI generative-AI guidance of 2026, the STM recommendations for handling image integrity issues, and the principles in the UK Concordat and the European Code of Conduct. Where a step is our own editorial recommendation rather than a requirement in those documents, it says so. Regulations differ by country and funder; check your own policy first.

1. Triage the concern. Write down what specifically made someone look: a feature of a panel, an odd pattern across papers, missing methods, or a tool alert. Under 42 CFR 93.307(a), an inquiry is warranted when an allegation falls within the definition of research misconduct, is within the rule’s scope, and is sufficiently credible and specific that potential evidence can be identified. A bare score with no stated threshold is a weak fit for “specific” (editorial interpretation).
2. Preserve originals and the chain of custody. 42 CFR 93.305(a) requires institutions to promptly obtain, inventory and sequester research records and other evidence, and, whenever possible, to do so before or at the time the respondent is notified. Keep the images as received, with file hashes and a log of who handled them. The ORI guidance notes that institutions must generally retain the institutional record and sequestered evidence for seven years after a proceeding ends.
3. Request source data from the authors. STM defines source data as minimally processed “raw” data underlying a figure and says editors may request it for any figure before or after publication. It adds that the absence of compelling source data may undermine confidence in a figure. Instrument files, acquisition metadata and intermediate processing files are the kind of material that separates real experiments from synthetic ones. The ORI guidance lists purchasing records, intermediate files from figure processing and physical access logs for animal facilities as examples of a forensic evidence trail.
4. Run the detector as one test, and record it. Record the tool name, version or model date if shown, the date of the run, the exact file tested and its format, any resizing or conversion you applied, the raw output and any threshold the vendor states. This is an editorial practice; it follows from the need to be able to explain and repeat every item of evidence. Run the original file, not a re-encoded copy.
5. Corroborate with independent methods and human experts. The ORI guidance recommends against relying on automated or AI-enabled tools as the sole source of evidence in a research misconduct proceeding, and says investigation committees must include persons with appropriate scientific expertise in the respondent’s field (42 CFR 93.310(f)). It also suggests institutions consider including AI subject-matter experts on committees. Use methods that do not share the detector’s blind spots: manual inspection, duplication screening, comparison with source files, and a field expert’s judgment on whether the image is plausible for the stated experiment.
6. Treat the score as a lead, not a finding. A finding of research misconduct under Part 93 requires a significant departure from accepted practices, intentional, knowing or reckless conduct, and proof by a preponderance of the evidence (42 CFR 93.103). The institution carries the burden of proof, and the finder of fact must give due consideration to credible evidence of honest error or difference of opinion (93.105). A detector score says nothing about intent. Inquiry stage reports cannot make findings of misconduct at all (93.307(f)(2)).
7. Give a fair process and a chance to respond. The institution must make a good-faith effort to notify the respondent in writing at or before the start of an inquiry (93.307(c)) and give the respondent an opportunity to review and comment on the inquiry report (93.307(g)(3)). The ALLEA code adds that accused persons should be given full details of the allegations and a fair process for responding and presenting evidence, and are presumed innocent until proven otherwise. STM says editors should inform corresponding authors confidentially and give them an opportunity to respond.
8. Stay in your lane. STM states that journal editors and publishers cannot undertake formal investigations for research misconduct and should limit their assessment to the evidence and claims in the published article, referring questions about individuals to the authors’ institutions. Institutions decide misconduct; journals protect the record.
9. Decide escalation by corroboration, not by score. There is no source that sets a score above which a figure is “AI-generated”, and we do not offer one. A defensible pattern (editorial interpretation) is: no action on an unexplained score alone; a request for source data when the score is accompanied by an identifiable concern; formal escalation when independent methods or an expert agree, or when source data are missing or inconsistent. STM’s three-level classification reflects the same logic: it records the editor’s conclusion on all the available evidence, which may change as more information arrives. COPE’s flowchart on inappropriate image manipulation in a published article (version 2, 20 May 2024) points editors toward a correction, expression of concern or retraction depending on whether the problem affects the conclusions, and recommends an expression of concern while an institutional investigation is ongoing.
10. Document and review. Keep the full log with the case file, note what each method contributed, and record the honest-error explanations the author offered.

How do I handle confidentiality and data protection when uploading images?

Uploading an unpublished figure to a web tool sends it to a third party. Before you do that, check what the tool’s operator says it does with the file.

As an example of what to look for, we read the privacy policy and terms of the tool linked above (both marked “last updated 01/14/2026”). The privacy policy says uploaded images are sent to a third-party storage service for processing, and that retention and deletion are managed by external service providers under their own terms. It gives no retention period and describes no image-specific deletion process. The terms say the service is provided as-is, make no warranty about accuracy, and grant the operator a broad license to content users post as “Contributions”. The text we read does not say whether an image uploaded for detection counts as a Contribution. If you are considering a tool like this for unpublished work, ask the operator in writing; do not assume.

The rules around investigations are explicit about confidentiality:

• 42 CFR 93.106(a) limits disclosure of the identities of respondents, complainants and witnesses to those who need to know, and 93.106(b) requires confidentiality for any records or evidence from which research subjects might be identified.
• The ALLEA code says investigations are conducted confidentially to protect those involved.
• The UK Concordat (2025) expects employers to maintain fair and transparent processes, manage conflicts of interest, use external advisers where needed and offer appeals processes, and it expects funders to treat allegations with confidentiality and abide by data protection laws.
• STM says editors who consult experts and reviewers should ensure a confidential process.

Our recommendation (editorial interpretation): for unpublished images, images of patients or participants, images under embargo, or anything in an open case, prefer an on-premise tool or a vendor under a written agreement that covers retention, deletion, no secondary use and breach notification, and clear this with your data protection officer. For published figures, a public web tool carries less confidentiality risk, but you still reveal that a particular paper is under scrutiny, so consider whether that disclosure is itself a problem.

Decision checklist

Check before relying on the result	What to do if the answer is no	Source
Is the detector output the only evidence?	Add independent evidence before any finding; ORI recommends against sole reliance on such tools.	ORI generative-AI guidance (2026)
Does the vendor publish an error profile for your image type, and have you validated it?	Do not weight the score; free web detectors performed unevenly on blots, and a home page may state no accuracy rates.	Gosselin (2025); imagedetector.com home page
Are the original files and source data obtained and sequestered?	Obtain and inventory them, ideally before or when the respondent is notified.	42 CFR 93.305(a); STM (2021)
Was the exact tested file the original, not a screenshot or recompressed copy?	Rerun on the original; compression and resizing degrade detector performance.	Corvi et al. (2022); Gosselin (2025)
Has a field expert reviewed the image, with AI expertise available?	Add expert review; investigation committees must include appropriate scientific expertise.	ORI guidance (2026); 42 CFR 93.310(f)
Is tool, version, date, settings and raw output recorded?	Record them now and keep them in the case file.	ORI guidance (2026) on retaining the record
Is anyone inferring intent from the score?	Stop; intent must be established separately, and honest error considered.	42 CFR 93.103, 93.105; STM (2021)
Has the author been told and given a chance to respond?	Notify and invite comment according to your policy before conclusions are drawn.	42 CFR 93.307(c), (g)(3); ALLEA (2023); STM (2021)
Is the upload permitted by confidentiality rules and the vendor’s terms?	Use an on-premise tool or a contractual arrangement, or do not upload.	42 CFR 93.106; ALLEA (2023); vendor privacy policy and terms
Is a journal acting beyond its scope?	Limit the assessment to the published record and refer individual questions to the institution.	STM (2021)
Has provenance or disclosure been checked, without treating its absence as proof?	Check for Content Credentials and an AI-use disclosure, but do not treat their absence as a finding.	C2PA explainer 2.2; COPE position on authorship and AI tools

Common mistakes

• Treating a probability as a verdict. The score describes the model’s output on a file, not the history of the experiment.
• Testing a derivative. Screenshots, PDF extractions and platform-recompressed images change what the detector sees.
• Ignoring base rates. A tool with moderate specificity run across many authentic figures produces many false alarms.
• Using a tool on image types it was not tested for. Scientific images are a different domain from the photographs most detectors target.
• Reading a low score as clearance. Detectors can be evaded and can miss new generators.
• Uploading confidential material to a free service without reading its terms.
• Telling the author only the score. Share the specific concern and the evidence, and invite the underlying data.
• Skipping the record. If you cannot say which version of which tool produced the number, you cannot defend it.

Frequently asked questions

Can an AI image detector prove that a figure is AI-generated?

No. It returns a probability from a model with its own error rates, and the studies above show those rates can be poor on scientific images. Treat the result as a reason to request source data and seek expert review.

Is a detector score enough to support a finding of research misconduct?

Not under the US rules or the ORI guidance, which recommend against relying on such tools as the sole source of evidence. A finding also requires a significant departure from accepted practice, a culpable state of mind and proof by a preponderance of the evidence (42 CFR 93.103).

Should I upload unpublished figures to a free web detector?

Generally not. Read the operator’s privacy policy and terms first, and prefer an on-premise tool or a written agreement for unpublished, embargoed, patient or participant images.

What is the difference between an AI-image detector, duplication screening and content credentials?

A detector estimates whether an image looks machine-generated. Duplication screening looks for reused or altered regions. Content credentials record provenance, and say nothing about whether the content is true.

What should I record about the tool I used?

The name, version or model date, run date, the exact file and format tested, any conversion applied, the raw output, any stated threshold, and who ran it. Keep this with the case file.

Where does this fit with the rest of my integrity process?

It is one input among several. See CASRAI’s research integrity fundamentals for the wider framework and the US Office of Research Integrity entry for the US context.

Sources

All sources were read on 2026-10-30 unless noted.

1. US Office of Research Integrity, Guidance for PHS Policies on Research Misconduct, 42 CFR Part 93 (2024): Generative Artificial Intelligence, 2026 (nonbinding recommendations). https://ori.hhs.gov/sites/default/files/2026-08/ORI%20Generative%20Artificial%20Intelligence%20Guidance_final.pdf; announcement of 14 August 2026: https://ori.hhs.gov/blog/oris-guidance-generative-artificial-intelligence-released
2. Electronic Code of Federal Regulations, 42 CFR Part 93, Public Health Service Policies on Research Misconduct (sections 93.103, 93.105, 93.106, 93.305, 93.307, 93.312). https://www.ecfr.gov/current/title-42/chapter-I/subchapter-H/part-93
3. STM Working Group on Image Alteration and Duplication Detection, Recommendations for handling image integrity issues, Version 1.0, 6 December 2021. https://s3.eu-west-2.amazonaws.com/stm.offloadmedia/wp-content/uploads/2024/08/10032902/2021_12_07_Recommendations_for_handling_image_integrity_issues_V10-1.pdf
4. STM, Feasibility of technical solutions for the detection of falsified images in research, 6 February 2025. https://stm-assoc.org/document/feasibility-of-technical-solutions-for-the-detection-of-falsified-images-in-research/
5. COPE, Inappropriate image manipulation in a published article (flowchart, version 2, 20 May 2024). https://publicationethics.org/guidance/flowchart/inappropriate-image-manipulation-published-article
6. COPE, Authorship and AI tools (position statement). https://publicationethics.org/guidance/cope-position/authorship-and-ai-tools
7. The Concordat to Support Research Integrity, 2025 edition (UKCORI). https://ukcori.org/wp-content/uploads/2025/12/The-Concordat-to-Support-Research-Integrity-2025.pdf
8. ALLEA, The European Code of Conduct for Research Integrity, Revised Edition 2023, section 3.2. https://allea.org/wp-content/uploads/2023/06/European-Code-of-Conduct-Revised-Edition-2023.pdf
9. Gosselin R-D. AI detectors are poor western blot classifiers: a study of accuracy and predictive values. PeerJ 13:e18988, 20 February 2025. doi:10.7717/peerj.18988. https://peerj.com/articles/18988
10. Corvi R, Cozzolino D, Zingarini G, Poggi G, Nagano K, Verdoliva L. On the detection of synthetic images generated by diffusion models. arXiv:2211.00680, 2022 (preprint). https://arxiv.org/abs/2211.00680
11. Ojha U, Li Y, Lee YJ. Towards universal fake image detectors that generalize across generative models. arXiv:2302.10174, 2023, revised 2024 (preprint). https://arxiv.org/abs/2302.10174
12. Carlini N, Farid H. Evading deepfake-image detectors with white- and black-box attacks. arXiv:2004.00622, 2020 (preprint). https://arxiv.org/abs/2004.00622
13. C2PA and Content Credentials Explainer, version 2.2, 22 April 2025. https://spec.c2pa.org/specifications/specifications/2.2/explainer/_attachments/Explainer.pdf
14. ImageDetector.com home page, privacy policy and terms of use (both last updated 01/14/2026). https://imagedetector.com/privacy/ and https://imagedetector.com/terms/. The tool is named as an example only and is not endorsed or validated by CASRAI.