1. Introduction

Data scraping for biomedical artificial intelligence (AI) refers to the automated collection of medical research papers, clinical-trial records, genomic datasets, pathology images, radiology scans, pharmaceutical databases, electronic health information, and other biomedical materials from websites, repositories, and digital platforms for training, validating, or evaluating AI models.

The central legal question is whether a developer can collect and use these materials without obtaining permission from every copyright owner, relying on doctrines such as fair use, fair dealing, research exceptions, scientific-data exceptions, or text-and-data-mining (TDM) exceptions.

Biomedical AI creates distinctive legal problems because the same dataset may contain copyright-protected articles, uncopyrightable scientific facts, confidential clinical records, personal health information, database rights, and contractually restricted material.

The key principle is that scientific purpose does not automatically make scraping lawful. Legality depends on the jurisdiction, the material collected, how it was acquired, the purpose of processing, privacy restrictions, contractual terms, and the effect on legitimate licensing markets.

The discussion below distinguishes binding judgments from persuasive foreign authorities and explains their relevance to biomedical AI. The law is stated as of 5 November 2026.

2. What activities constitute biomedical AI data scraping?

News: Uniklinik Mannheim

A. Scientific literature scraping

Downloading journal articles, clinical guidelines, medical textbooks, systematic reviews, and pharmaceutical publications to train a biomedical language model. Copyright in the article's expression may remain protected even if the underlying scientific findings are not.

ᐈ Чим МРТ відрізняється від КТ і що робити в тому чи іншому випадку

B. Medical-image scraping

Collecting X-rays, CT scans, MRI images, histopathology slides, and associated annotations. Copyright, database rights, patient confidentiality, and data-protection law may apply independently.

Представлена Evo 2 от Nvidia и партнёров — ИИ-модель, обученная на 9,3 трлн нуклеотидов геномов более 128 тысяч видов

C. Genomic and biological data collection

Acquiring genome sequences, protein structures, biomarker data, and gene-expression datasets. Raw facts and sequences are not necessarily copyrightable, but original annotations, databases, software, and associated personal data may be protected.

Ovid Biotech, Pharma and CRO Research | Wolters Kluwer

D. Clinical-trial and patient-data collection

Gathering trial protocols, adverse-event reports, participant-level data, and electronic health records. Copyright is only one issue; consent, confidentiality, ethics approvals, and restrictions on secondary use may be more significant.

3. The legal framework: what must be established?

A scraping dispute should be analysed through separate legal questions rather than treating all collected information as one category.

A successful copyright defence does not necessarily defeat a privacy claim, a breach-of-contract claim, or a database-right claim.

4. Fair use under United States law

The United States provides one of the most developed frameworks for assessing whether copying protected material for computational research is lawful.

Section 107 of the US Copyright Act establishes four factors for determining fair use:

Purpose and character of the use, including commercial or nonprofit purposes.

Nature of the copyrighted work.

Amount and substantiality of the material used.

Effect on the potential market for, or value of, the original work.

No single factor automatically determines the outcome. A biomedical AI developer must establish fair use on the facts of the particular dispute.

LII / Legal Information Institute+1

Application to biomedical AI

Biomedical AI may have a compelling public-interest purpose, but a beneficial medical application is not an automatic exemption from copyright law. Equally, commercial deployment does not, by itself, establish infringement.

5. Important case laws and their relevance to biomedical AI

The following cases provide the principal legal authorities for evaluating scraping, computational research, and AI training. The cases are not all biomedical disputes; their relevance lies in the legal principles they establish or illustrate.

Case 1. Authors Guild v. Google, Inc., 804 F.3d 202 (2d Cir. 2015)

Court: United States Court of Appeals for the Second Circuit.

Facts: Google digitised millions of copyrighted books to create a searchable index. Users could search for terms and view limited snippets rather than access complete books.

Decision: The court upheld fair use. It considered the search and information-location function transformative and found that the project did not unlawfully substitute for the books themselves.

UC Berkeley Law+1

Application to biomedical AI:

Indexing medical publications to locate studies relevant to a rare disease may resemble the searchable-index function.

Extracting information about drug efficacy or adverse events may serve a distinct analytical purpose.

A diagnostic research tool that returns citations and limited relevant excerpts presents a stronger analogy than a chatbot that reproduces entire paywalled papers.

Limitation: Google Books did not decide whether every form of generative AI training is fair use. A biomedical model that generates near-verbatim copies of medical publications raises different issues from a searchable index.

Case 2. Authors Guild v. HathiTrust, 755 F.3d 87 (2d Cir. 2014)

Court: United States Court of Appeals for the Second Circuit, 2014.

Facts: A consortium of research libraries digitised books for full-text searching and accessible formats for people with print disabilities.

Decision: The court upheld important uses of the digital library under fair use, including full-text search and certain accessibility uses.

Legal principle: Digitisation and computational analysis may serve a purpose distinct from ordinary consumption of the original work.

Biomedical relevance: A research consortium creating a searchable corpus of scientific publications or using articles to analyse medical terminology can draw on this reasoning. The analogy becomes weaker if the corpus is distributed as a replacement for the original licensed publications.

Case 3. Andy Warhol Foundation for the Visual Arts, Inc. v. Goldsmith, 598 U.S. 508 (2023)

Court: Supreme Court of the United States.

Facts: The dispute concerned the commercial licensing of an altered portrait based on a photograph originally created by Lynn Goldsmith.

Decision: The Supreme Court held that the specific commercial licensing use at issue was not fair use. It emphasised that a new meaning or message does not alone make a use transformative; the court must examine the particular purpose of the challenged use.

LII / Legal Information Institute

Biomedical relevance: A developer cannot simply argue that a medical AI model is transformative because it uses data to produce predictions.

For example, a company that copies a publisher's medical illustrations to create a competing commercial image library may face a different analysis from a university that uses those images to study image-recognition performance.

The specific challenged use, not merely the technology's overall purpose, matters.

Case 4. Thomson Reuters Enterprise Centre GmbH v. ROSS Intelligence Inc. (D. Del. 2025; subsequent appellate proceedings)

Court: United States District Court for the District of Delaware; subsequent proceedings in the Third Circuit.

Facts: Thomson Reuters alleged that ROSS Intelligence used proprietary Westlaw editorial headnotes in developing a competing legal-research system.

Decision: The district court rejected fair use for the challenged copying, emphasising the protected editorial expression, the commercial and competitive purpose, and harm to actual or potential markets. The appellate proceedings provide further authority on the protection of editorial headnotes and the fair-use analysis.

AP News+1

Biomedical relevance: Consider a company that scrapes a proprietary clinical decision-support database containing editorially curated drug-interaction explanations, then trains a rival clinical decision-support product.

The case supports several arguments for the database owner:

Original editorial summaries may be copyrightable.

A competing commercial service may weigh against fair use.

The absence of verbatim copying in the final model does not automatically resolve whether the training-stage copying was lawful.

Important distinction: Raw medical facts, such as a measured laboratory value or the existence of a documented drug interaction, are not automatically protected just because they appear in a proprietary database. The original selection, arrangement, annotation, and explanatory expression may receive separate protection.

Case 5. Bartz v. Anthropic PBC, 787 F. Supp. 3d 1007 (N.D. Cal. 2025)

Court: United States District Court for the Northern District of California.

Decision date: 23 June 2025.

Facts: Authors challenged Anthropic's use of copyrighted books in developing its AI models. The case involved both lawfully acquired books that were digitised and books obtained from pirated sources.

Decision: The court found that training on the works in the circumstances before it was transformative and fair use, while distinguishing that activity from building a permanent library using pirated copies. The acquisition and retention of pirated works created separate infringement concerns.

WIPO+1

Biomedical relevance: This is particularly important for research datasets obtained from different sources.

Suppose a biotechnology company trains a drug-discovery model using:

Openly licensed scientific articles;

Articles purchased under institutional subscriptions;

Medical datasets obtained from a legitimate repository; and

Journal articles downloaded from unauthorised file-sharing websites.

The legal analysis should not treat all four sources identically. A potentially transformative training use does not automatically legalise an unlawful acquisition or the creation of an unauthorised permanent archive.

The case is a district-court decision, not a universal rule that all AI training is fair use.

Case 6. Kadrey v. Meta Platforms, Inc., 788 F. Supp. 3d 1026 (N.D. Cal. 2025)

Court: United States District Court for the Northern District of California.

Decision date: 25 June 2025.

Facts: Authors alleged that Meta used their copyrighted books to train its Llama models without authorisation.

Decision: On the evidentiary record before it, the court granted Meta summary judgment on the fair-use issue. It stressed that the ruling did not establish that AI training is necessarily fair use in other cases. The plaintiffs' market-harm evidence was insufficient in the circumstances.

FindLaw+1

Biomedical relevance: The case highlights the importance of evidence concerning market substitution.

A biomedical publisher challenging AI training should document matters such as:

Lost institutional subscriptions or database licences;

A decline in sales of clinical reference products;

Substitution of a commercial AI system for licensed medical content;

Unauthorised generation of protected article text; and

The existence and economic significance of a market for AI-training licences.

A general assertion that AI might reduce future revenue may be less persuasive than concrete evidence connecting the challenged use to an actual or reasonably supported potential market.

Comparative lesson from the US cases

Potentially stronger fair-use argument

Lawfully acquired scientific materials used for a genuinely analytical purpose, with reasonable security and little evidence of substitution for the original content.

Higher infringement risk

Pirated corpora, circumvention of access restrictions, commercial copying of proprietary databases, or model outputs that substitute for licensed medical publications.

These are risk indicators, not automatic legal outcomes. Fair use remains fact-specific.

6. Indian law: fair dealing and biomedical AI scraping

India is particularly important because its Copyright Act, 1957 does not contain a general, express text-and-data-mining exception equivalent to the European Union's statutory framework.

The relevant provisions include:

Section 14: Defines the exclusive rights associated with copyright, including reproduction rights.

Section 51: Sets out circumstances in which copyright is infringed.

Section 52(1)(a): Provides fair-dealing exceptions for specified purposes, including private or personal use, research, criticism or review, and reporting current events, subject to the statutory wording.

Section 65A: Addresses circumvention of effective technological measures in the circumstances specified by the Act.

Section 65B: Addresses removal or alteration of rights-management information in specified circumstances.

The precise application of these provisions depends on the work, the acts performed, and the available exception.

Chambers and Partners+1

Case 7. ANI Media Private Limited v. OpenAI OpCo LLC, CS(COMM) 1028/2024

Court: Delhi High Court.

Key interim ruling: 24 July 2026, as reported in the available case materials.

Facts: Asian News International (ANI) alleged that OpenAI used its copyrighted news content to train AI models without permission and raised concerns about the use of its material in generated responses.

Interim decision: The court refused ANI's request for an interim injunction. At the prima facie stage, it considered the challenged use to fall within the research limb of the fair-dealing exception under Section 52(1)(a)(i), while addressing reproduction, jurisdiction, and output-related issues. The ruling was an interim determination rather than a universal exemption for AI training.

WIPO Lex+2

Relevance to biomedical AI

The reasoning may be relevant where an Indian developer scrapes scientific publications, clinical-trial reports, or other copyrighted material to train a medical AI model.

However, several distinctions are critical:

ANI concerned news content, not clinical records, genomic databases, or patient-level datasets.

A research-related purpose does not automatically resolve the legality of every copying act.

The ruling does not create an unrestricted right to scrape subscription databases, circumvent paywalls, or collect confidential patient information.

Whether the same analysis applies to a commercial biomedical model will depend on the evidence, the statutory interpretation, and the nature of the material.

Privacy and confidentiality claims remain analytically separate from copyright fair dealing.

The judgment is a significant Indian authority for the AI-training debate, but its interim character must be preserved when citing it in legal submissions.

Case 8. Civic Chandran v. Ammini Amma, 1996 PTC (16) 329 (Ker)

Court: Kerala High Court.

Facts: The dispute concerned a dramatic work that criticised and reproduced elements of an earlier play.

Legal significance: The court examined the purpose and character of the use, the amount taken, and the impact on the original work when considering fair dealing.

Biomedical application: A researcher who reproduces selected passages from a medical publication to critique its methodology may have a stronger fair-dealing argument than a developer who copies an entire collection of articles to build a commercial substitute.

The case does not decide AI training directly. It is relevant as an Indian authority illustrating that the fairness of a use depends on its purpose, extent, and effect.

Case 9. Eastern Book Company v. D.B. Modak, (2008) 1 SCC 1

Court: Supreme Court of India.

Facts: The dispute concerned copyright in judicial-reporting materials, including editorial additions to judgments.

Decision: The Supreme Court applied a standard of originality requiring skill and judgment beyond mere labour or mechanical copying.

Biomedical application: The case helps distinguish between scientific facts and original expression.

For example:

A raw DNA sequence or a numerical laboratory result does not become copyrightable merely because it is valuable.

Original annotations, explanatory medical summaries, illustrations, and creative database arrangements may attract copyright protection where the statutory originality requirements are satisfied.

A developer may be able to use unprotected scientific facts while still infringing copyright by reproducing protected expression.

The case is relevant to determining what material in a biomedical dataset is protected, rather than establishing a general exception for scraping.

Position under Indian law

India's fair-dealing framework requires particular caution because it does not provide the same express TDM structure as the EU. The ANI interim ruling is important, but developers should not assume that all automated copying for medical research is lawful.

7. European Union: statutory text-and-data-mining exceptions

The EU provides a more explicit framework under Directive (EU) 2019/790 on copyright and related rights in the Digital Single Market.

Official legal text: Directive (EU) 2019/790 on EUR-Lex.

Article 3: Scientific-research TDM

Article 3 requires an exception for qualifying reproductions and extractions made by research organisations and cultural heritage institutions for scientific research, where they have lawful access to the material.

The provision also addresses secure storage and retention for scientific research, including verification of results.

EUR-Lex

For biomedical AI, a qualifying research organisation could potentially use this framework to mine lawfully accessible medical publications or datasets for scientific research, subject to applicable national implementation and other legal requirements.

Article 4: General TDM exception

Article 4 provides a broader TDM exception for lawfully accessible material, subject to conditions including the rights holder's ability to expressly reserve relevant rights in an appropriate manner.

For online material, machine-readable rights reservations may be relevant. Copies may be retained for as long as necessary for TDM.

Important distinction: Article 3 and Article 4 are not interchangeable. Article 3 is designed for qualifying research organisations and cultural heritage institutions conducting scientific research. Article 4 is broader but incorporates the rights-reservation mechanism.

A commercial pharmaceutical company should not assume that it qualifies for Article 3 merely because its project involves scientific research.

Biomedical examples

The TDM exceptions concern copyright and related rights. They do not independently authorise unlawful access to data, eliminate GDPR obligations, or grant permission to process confidential medical records.

8. Biomedical datasets: copyright is only one part of the dispute

A dataset can be lawfully reusable for one purpose and still be restricted for another. The following distinctions are especially important.

Scientific articles

Copyright can protect original text, figures, photographs, and illustrations. Open access does not necessarily mean unrestricted reuse; the specific licence governs the permissions granted.

Clinical-trial databases

Individual factual results may be unprotected by copyright, but the database's selection, arrangement, annotations, or substantial investment may attract protection under applicable law. The EU also has a sui generis database right.

Genomic information

Raw sequence data and scientific discoveries should be distinguished from original explanatory materials and protected database structures. Identifiable genomic data can also create serious privacy and re-identification risks.

Patient-level data

De-identification does not automatically eliminate every legal restriction. Consent limitations, confidentiality duties, ethics approvals, data-protection rules, and contractual controls may continue to apply.

Privacy and confidentiality

In the EU, the General Data Protection Regulation (GDPR) governs the processing of personal data, including health data under its special-category framework. A lawful basis under Article 6 and a relevant condition under Article 9 must be established where applicable, alongside other requirements.

In India, the Digital Personal Data Protection Act, 2023 and applicable rules and commencement notifications must be considered when processing digital personal data. The applicable legal position depends on the provisions in force, the processing circumstances, and any relevant exemptions. Biomedical researchers must also assess other applicable health, confidentiality, and sector-specific requirements.

In the United States, HIPAA may apply to covered entities, business associates, and relevant protected health information; it does not automatically cover every biomedical dataset or every AI developer.

Consequently, a developer may have a defensible copyright argument but still lack lawful authority to scrape or reuse a dataset containing identifiable health information.

9. Major grounds on which biomedical AI scraping disputes arise

A. Unauthorised reproduction

A developer may reproduce protected material when downloading articles, creating training corpora, making temporary or persistent copies, or storing images for repeated training.

The rights holder may argue that these activities fall within its exclusive reproduction rights. The developer may respond that the copying is licensed or covered by a statutory exception.

The key question is not simply whether the model's final parameters contain a readable copy of the original work. The legality of intermediate copying must also be examined.

B. Commercial substitution

A medical AI platform could compete with a publisher's clinical reference service or a proprietary biomedical database.

For example, a model that generates summaries of medical guidelines might reduce the need to access the licensed source. Whether this creates legally significant market harm depends on the actual use, the evidence, and the applicable law.

The reasoning in Warhol, Kadrey, and Thomson Reuters v. ROSS Intelligence helps explain why commercial purpose and market effects may matter.

LII / Legal Information Institute+2

C. Circumvention of access restrictions

Scraping public webpages is not legally identical to bypassing authentication, defeating technical restrictions, or obtaining content from unauthorised repositories.

Potentially relevant issues include:

Breach of contractual access conditions;

Unauthorised access under applicable computer-misuse legislation;

Circumvention of technological protection measures;

Misappropriation of confidential information; and

Breach of privacy or data-security requirements.

The consequences depend on the jurisdiction, the particular access method, and the relevant legal provisions.

D. Memorisation and infringing outputs

Even if a developer has a plausible defence for training, the model may generate passages, medical illustrations, or other material that substantially reproduces protected works.

A developer should therefore assess two separate questions:

Was the material lawfully acquired and copied for training?

Do the resulting model outputs independently infringe protected rights?

These questions can produce different answers. The lawfulness of training does not automatically immunise every output.

E. Patent and trade-secret concerns

Copyright does not generally protect scientific ideas or discoveries as such. However, scraping may expose a developer to separate claims if it involves confidential drug-development information, unpublished experimental results, protected software, or proprietary methods.

Patent infringement is a separate question: obtaining information about a patented invention is not, by itself, the same as practising the patented invention.

10. Defences available to biomedical AI developers

A developer facing a scraping claim may consider the following arguments, depending on the evidence.

A defence must be assessed against the specific claim. For instance, showing that scientific facts are not protected by copyright does not necessarily defeat a claim involving protected article text or database rights.

11. Compliance framework for biomedical AI developers

The strongest approach is to establish data provenance and legal authority before building the training corpus.

Biomedical AI scraping compliance checklist

Progress

0 of 10

Identify the owner and source of every dataset.

Classify articles, raw facts, images, annotations, and patient records separately.

Record licences, access conditions, and rights reservations.

Document the statutory exception or licence relied upon for each source.

Exclude unlawfully acquired or confidential material unless lawful authority is established.

Review consent, privacy, ethics, and secondary-use permissions for health data.

Maintain source records, timestamps, hashes, and acquisition logs.

Implement security controls and access restrictions for retained datasets.

Test model outputs for substantial reproduction of protected material.

Establish a process for rights-holder complaints, removal requests, and remediation.

Copy checklist

For a research institution, the checklist should also address publication, reproducibility, and data-sharing obligations. For a commercial developer, it should include licence negotiations, indemnities, model-output controls, and documented approval of high-risk sources.

12. Remedies and potential liability

Depending on the jurisdiction and the claim established, a rights holder may seek:

An injunction restricting continued copying or use;

Damages or an account of profits where legally available;

Statutory damages where the relevant statute permits them;

Destruction, deletion, or other treatment of infringing copies;

Disclosure or preservation of relevant evidence; and

Contractual remedies or other relief for independently established violations.

In biomedical AI disputes, the requested remedy requires particular care. Removing a dataset from a future training pipeline is not necessarily the same as deleting existing copies, retraining a model, or addressing potentially infringing outputs. Courts must consider the applicable legal standards and evidence when deciding what relief is justified.

A university may also face research-integrity or institutional consequences, while a commercial company may face licensing disputes, contractual claims, and reputational harm in addition to copyright litigation.

13. Comparative legal position

These systems should not be treated as interchangeable. A scraping operation may be lawful under one country's TDM rules but infringing under another country's copyright law. Cross-border projects must assess the relevant acts and applicable laws rather than assume that one country's exception governs the entire pipeline.

EUR-Lex+2

14. Conclusion

Data scraping for biomedical AI sits at the intersection of copyright, scientific research, database protection, privacy, and innovation policy.

The principal lessons from the case law are:

Computational research may support a fair-use or fair-dealing argument, but its purpose alone does not settle the issue.

The source and acquisition method matter. Bartz illustrates the distinction between training on lawfully acquired material and building a library from pirated works.

Market effects are important. Google Books, Kadrey, and Thomson Reuters v. ROSS Intelligence illustrate different approaches to transformation, competition, and the value of protected material.

Indian law requires a separate statutory analysis. The interim ruling in ANI v. OpenAI is significant, but it should not be interpreted as blanket permission to scrape all biomedical information.

Copyright permission is not a substitute for privacy compliance. Patient-level information, genomic data, and confidential clinical records require independent legal assessment.

The most defensible biomedical AI strategy is to combine lawful data acquisition, clear licensing or exception analysis, documented provenance, appropriate privacy safeguards, and output-level infringement testing.

Legal note: This is a comparative research overview, not an opinion on a particular dispute. For a litigation-ready analysis, the applicable jurisdiction, source datasets, licence terms, and intended commercial or non-commercial use must be identified. The status of ongoing proceedings and subsequent appellate decisions should also be verified before relying on any case as current binding authority.