- with Inhouse Counsel
- with readers working within the Healthcare industries
- in European Union
For years, AI developers operated on a simple assumption: if content was publicly accessible online, it could be scraped, processed, and used to train AI systems. That assumption is now at the centre of one of the most consequential legal battles in the technology industry.
The issue is no longer limited to large foundation model companies. Enterprises deploying Generative AI systems, procurement teams integrating third-party AI tools, legal departments approving AI workflows, and boards overseeing digital transformation are now exposed to the same risk chain. Regulators, publishers, creators, and courts are drawing a distinction between content that is “available” and content that is “licensed.”
That distinction is reshaping how AI systems are built in 2026 and beyond.
The market has already moved past the early “move fast” phase of AI development. Courts are now examining whether AI companies copied copyrighted material without authorisation, whether datasets were lawfully obtained, whether outputs compete with original works, and whether AI developers can rely on fair use or text-and-data-mining exceptions at scale.
The result is a rapidly evolving compliance landscape in which data governance is as important as model performance.
Why the “publicly available” argument is failing
The internet was never designed as a free training repository for commercial AI systems.
Copyright law protects creative expression regardless of whether the content is freely accessible online. A newspaper article behind a paywall, an artist’s portfolio page, a GitHub repository, a blog post, or even publicly indexed website content may still carry licensing restrictions, contractual obligations, database rights, or copyright protections.
This is where many AI developers underestimated the legal risk.
The problem is not merely that AI systems “viewed” content. Courts are now evaluating whether companies copied, stored, reproduced, transformed, redistributed, or commercially benefited from protected works during model training.
The legal scrutiny intensified because many datasets used in foundational model development were allegedly sourced from pirated books, scraped journalism, copyrighted images, code repositories, and proprietary databases without consent or licensing agreements.
That changed the conversation from innovation to infringement.
The cases that changed the AI training debate
Several global disputes are now shaping the legal standards for AI training data governance.
The New York Times v. OpenAI and Microsoft1
The New York Times sued OpenAI and Microsoft, alleging that copyrighted news articles were used without authorisation to train generative AI systems. The lawsuit claims that AI-generated responses can reproduce or closely imitate original journalism, potentially reducing traffic and revenue for publishers. The presiding judge denied OpenAI’s motion to dismiss certain copyright infringement and trademark dilution claims, allowing core parts of the Times’ lawsuit to proceed. The case is now a major test of whether AI training qualifies as “transformative use” or creates unlawful market harm under copyright law.
Anthropic and the “pirated library” problem2
In 2025, a U.S. federal judge held that training AI systems on books could qualify as fair use under specific circumstances. However, the same decision also found serious problems with how the underlying data was obtained and stored.
While the training use was protected, the sourcing was not. The court reportedly concluded that maintaining millions of pirated books in a centralised training library was not protected.
Meta and Market Harm Arguments3
The litigation involving Meta’s LLaMA models exposed another major issue: the use of allegedly pirated datasets during AI development. The discourse around Meta and market harm generally falls into two distinct legal and economic arenas: copyright infringement in generative AI and antitrust/monopoly abuse. Both challenge Meta’s core business models.
Court filings and judicial comments increasingly focused on whether AI-generated outputs could dilute or compete with the original creative market. Ultimately, these legal battles signal a growing regulatory pushback against Meta’s strategy of leveraging unauthorised data and platform dominance for corporate growth.
Thomson Reuters v. ROSS Intelligence4
This case became a turning point because the court rejected a fair use defence involving copyrighted legal research content used to build a competing AI-driven platform. The decision is particularly important for enterprise AI because it demonstrates that courts may treat proprietary databases differently from broad internet content.
Businesses training internal AI systems on subscription databases, licensed repositories, customer datasets, or proprietary enterprise records should pay close attention to this distinction.
The assumption that “internal use” automatically reduces risk is no longer safe.
India’s defining AI copyright battle: ANI Media Pvt Ltd v OpenAI Inc & Anr (CS(COMM) 1028/2024)5
India’s most significant AI copyright dispute is currently unfolding before the Delhi High Court in ANI Media Pvt Ltd v OpenAI Inc & Anr (CS(COMM) 1028/2024). The case is expected to become a landmark precedent for how Indian law treats AI training data, web scraping, and the commercial use of copyrighted material in generative AI systems.
The dispute was initiated by Asian News International (ANI), one of India’s largest news agencies, against OpenAI, alleging unauthorised use of ANI’s copyrighted news content for training ChatGPT and related AI systems. ANI seeks a permanent injunction to prevent OpenAI from using and reproducing its material, and has requested the removal of its content from OpenAI’s training datasets
At the core of the dispute is a question that now sits at the centre of global AI regulation: does the use of copyrighted content for AI model training amount to copyright infringement, even when the content is publicly accessible online?
ANI argued before the Delhi High Court that its news reports, articles, and syndicated journalistic content constitute protected literary works under the Indian Copyright Act, 1957. According to ANI, OpenAI allegedly reproduced, stored, processed, and commercially exploited this material without authorisation or licensing consent. The news agency also contended that AI-generated responses could reproduce or closely resemble its original reporting, thereby affecting the commercial value of its content.
The matter is particularly important because ANI’s arguments go beyond simple copying. The case raises broader issues involving:
- whether AI training constitutes “reproduction” under Section 14 of the Copyright Act,
- whether temporary storage and ingestion of copyrighted works during model training amounts to infringement,
- whether AI-generated summaries or outputs create market substitution concerns,
- and whether global AI companies can rely on “publicly available” content as a defense in India.
OpenAI, in response, reportedly argued that ChatGPT does not reproduce articles in the traditional sense and that its systems process patterns and relationships in data rather than storing expressive copies for republication. The company also reportedly raised jurisdictional and operational arguments, including the claim that portions of model training infrastructure and processing occur outside India.
What makes litigation strategically important for businesses is that Indian copyright law currently lacks a dedicated AI or text-and-data-mining framework comparable to the European Union’s approach. This means the Delhi High Court is effectively being asked to interpret conventional copyright provisions in the context of large language model training.
The court’s interpretation of Section 52 of the Copyright Act, which deals with “fair dealing” exceptions, may become especially significant. Unlike the broader U.S. fair use doctrine, India’s fair dealing framework is narrower and applies only to specific statutory purposes such as private use, research, criticism, review, and reporting of current events.
That distinction matters. If Indian courts conclude that commercial AI training does not comfortably fit within existing fair dealing exceptions, AI developers and enterprise adopters could face a stricter compliance environment in India than in some other jurisdictions.
The case has also triggered wider participation from Indian publishing and media stakeholders, many of whom are closely watching the proceedings because the outcome could influence how AI systems interact with journalism, publishing, entertainment, and digital content industries across the country.
For enterprises, the ANI litigation sends a clear message: India is no longer merely observing global AI copyright disputes from the sidelines. Indian courts are now actively shaping the legal standards that may govern AI training, content ingestion, and generative AI deployment within one of the world’s fastest-growing digital markets.
The commercial implications are substantial. Businesses operating AI systems in India, particularly in media, publishing, analytics, legal technology, knowledge management, and enterprise AI automation, should prepare for increased scrutiny around dataset sourcing, licensing practices, vendor disclosures, and output governance.
The case also highlights a broader strategic shift taking place globally: copyright owners are increasingly challenging not only AI-generated outputs, but also the upstream training processes themselves. That means legal exposure may arise long before a problematic output is ever generated.
As India moves further into enterprise-scale AI adoption in 2026–27, ANI Media Pvt Ltd v OpenAI Inc & Anr is likely to become one of the country’s most closely watched technology law cases, not only for what it says about copyright, but for how it defines the future boundaries of lawful AI development in India.
The real enterprise risk is no longer the lawsuit alone
Many organisations still treat AI copyright disputes as a problem only for large model developers. That is becoming a dangerous assumption.
Enterprises now face downstream liability exposure through:
- Procurement of non-compliant AI vendors
- Integration of unverified foundation models
- Employee use of unauthorised generative AI systems
- Ingestion of copyrighted enterprise data into external AI tools
- AI-generated outputs that resemble protected content
- Insufficient contractual safeguards with AI providers
This is why AI governance is rapidly becoming a board-level issue rather than a narrow technology issue.
The challenge is no longer just legal compliance. It is operational resilience. A company deploying AI systems built on legally questionable datasets may face regulatory investigations, contract disputes, reputational damage, indemnity conflicts, or content takedown demands even if it did not directly scrape the data itself.
What companies should have done and what businesses must do now
The strongest pattern emerging across AI litigation is not merely that copyrighted content was used. It is that companies often failed to build defensible governance structures around data acquisition and model development. Several safeguards could have significantly reduced risk exposure.
Organisations should have implemented formal dataset provenance tracking from the beginning. Many companies cannot clearly identify where portions of their training data originated, whether licenses existed, or whether scraping violated platform terms.
They should also have separated legally obtained datasets from high-risk scraped material. Courts are increasingly distinguishing between lawfully acquired content and pirated or unauthorised repositories.
Another major gap involved vendor diligence. Enterprises adopted AI tools without demanding transparency regarding training sources, indemnification obligations, or copyright compliance frameworks. The industry also underestimated output monitoring. Several lawsuits intensified because AI systems allegedly reproduced protected content too closely. Stronger controls over memorisation, filtering mechanisms, and output auditing could have reduced exposure.
Most importantly, businesses should have recognised that copyright compliance cannot be retrofitted after deployment. AI governance must be in place at the model design and procurement stages.
The new standard for AI governance in 2026–27
Leading organisations are already changing how they approach AI development and procurement. The shift is visible in four major areas.
First, licensed datasets are becoming commercially valuable assets. Publishers, media houses, research providers, and creative platforms are increasingly negotiating structured AI licensing agreements rather than allowing unrestricted scraping.
Second, enterprises are demanding contractual transparency from AI vendors. Procurement teams are now asking whether training data was licensed, whether opt-out mechanisms exist, and whether indemnities cover copyright disputes.
Third, companies are adopting internal AI usage policies that restrict how employees upload confidential or copyrighted content into external AI systems.
Fourth, AI governance is moving into enterprise risk management frameworks alongside cybersecurity, privacy, and regulatory compliance.
This is where the market is heading. The companies that succeed in AI adoption over the next three years will not necessarily be those with the fastest deployment cycles. They will be the organisations capable of scaling AI systems without creating unmanageable legal exposure.
Why this matters for decision-makers
Senior leadership teams should stop viewing AI copyright litigation as an isolated dispute between technology companies and creators. These cases are establishing the operating rules for enterprise AI itself. The emerging legal principle is becoming increasingly clear: accessibility does not equal permission.
Just because content can be scraped does not mean it can be commercially exploited for AI training. Just because a model vendor claims compliance does not eliminate enterprise liability. And just because an AI system produces new outputs does not automatically shield the underlying training process from scrutiny.
The organisations preparing for 2026–27 are already building governance structures around lawful data sourcing, licensing, contractual risk allocation, transparency, and AI accountability. That is quickly becoming the difference between sustainable AI adoption and preventable litigation exposure.
FAQs
1. Is publicly available internet content free to use for AI training?
No. Public accessibility does not remove copyright protection. Content may still be protected by copyright law, licensing terms, database rights, platform restrictions, or contractual obligations.
2. Are AI companies allowed to rely on fair use defences?
In some jurisdictions, courts have accepted limited fair use arguments for AI training under specific conditions. However, recent cases also show that unlawful acquisition methods, piracy, and market harm can still create liability exposure.
3. Why are publishers and creators suing AI companies?
Most lawsuits argue that copyrighted works were copied, stored, or used without authorisation during AI training and that AI-generated outputs may compete with or substitute the original content.
4. Can enterprises face liability even if they did not train the AI model themselves?
Yes. Enterprises may face contractual disputes, regulatory scrutiny, or downstream infringement claims depending on how AI tools are deployed, the data employees upload, and the vendor obligations.
5. What is the biggest compliance mistake companies are making?
Many organisations still cannot verify the provenance of training datasets or explain whether data was lawfully sourced and licensed.
6. What should businesses ask AI vendors before procurement?
Businesses should assess dataset sourcing practices, licensing models, copyright indemnities, transparency measures, output filtering controls, audit capabilities, and regulatory compliance frameworks.
7. How is the EU approach different from the U.S. approach?
The U.S. debate currently centres on fair use analysis through litigation, while the EU framework emphasises transparency, text-and-data-mining restrictions, and regulatory obligations under evolving AI governance rules.
8. What is the safest long-term approach for AI training data?
The safest approach is a combination of licensed datasets, documented provenance tracking, contractual compliance controls, internal governance frameworks, and continuous legal review of AI deployment practices.
Footnotes
1. https://harvardlawreview.org/blog/2024/04/nyt-v-openai-the-timess-about-face/
3. https://indiankanoon.org/doc/43025115/
4. https://law.justia.com/cases/federal/district-courts/delaware/dedce/1:2020cv00613/72109/547/
5. https://www.nls.ac.in/wp-content/uploads/2024/11/ANI_vs_OPEN_AI.pdf
The content of this article is intended to provide a general guide to the subject matter. Specialist advice should be sought about your specific circumstances.
[View Source]