Protecting Intellectual Property in the Age of AI

Introduction

Protecting intellectual property in the age of AI has become the defining legal, technical, and strategic challenge for creative companies, software firms, and research-intensive industries in 2026. The World Intellectual Property Organization tracked more than 54,000 generative AI patent families filed between 2014 and 2023, an eightfold surge that shows how quickly the asset base has grown. Corporate counsel now have to protect inputs, outputs, and the models themselves at the same time. Courts have rejected AI systems as authors, yet they have also allowed some training uses as transformative. Boards are learning that an employee pasting a confidential deck into a public chatbot can destroy trade-secret protection in a single afternoon. This guide walks through every pillar of IP law, the live cases reshaping each one, and a concrete playbook for teams who need to act in the next ninety days. It is written for operators, not only for litigators, because operational choices now decide most IP outcomes. Our earlier primer on AI and intellectual property law set the stage, and this piece extends it with 2026 case law and controls.

Quick Answers on Protecting Intellectual Property in the Age of AI

What does protecting intellectual property in the age of AI actually mean?

AI-era IP defense means layering copyright, patent, trademark, and trade-secret coverage across AI inputs, model artifacts, and outputs while controlling how employees and vendors use public AI tools.

Can an AI system own a copyright or patent in 2026?

No. IP protection under AI depends on a human author or inventor, because U.S. courts and the USPTO still require a natural person for both copyright and patent rights.

What is the single biggest IP risk from generative AI inside most companies?

For AI intellectual property protection, the top risk is employees pasting confidential code, contracts, or product plans into public chatbots, which strips trade-secret status.

Key Takeaways for IP Leaders in the AI Era

• AI intellectual property protection requires treating copyright, patent, trademark, and trade-secret coverage as a connected portfolio, not four silos.
• Trade secrets are now the load-bearing pillar for model weights, prompt libraries, and fine-tuning data, because patents and copyright reach only part of an AI stack.
• Fair-use rulings in 2024 and 2026 are moving toward protection for training on legally acquired works, while piracy and verbatim output remain high-risk zones.
• Governance, vendor contracts, and technical controls like watermarking and access gating now carry as much weight as formal registrations for day-to-day IP defense.

Understanding Intellectual Property Protection in the Age of AI

Protecting intellectual property in the age of AI means coordinating copyright, patent, trademark, and trade-secret coverage across model inputs, weights, prompts, outputs, and the governance that keeps all four pillars intact when AI does part of the creative work.

The Four Pillars of IP and Where AI Breaks Them

The classical IP system rests on four pillars, each designed for a specific kind of human output. Copyright covers creative expression fixed in a tangible medium, patents cover novel and non-obvious inventions, trademarks protect source identifiers that consumers rely on, and trade secrets protect information kept confidential and commercially valuable. These pillars assumed a world in which a person could point to a moment of authorship or invention. Generative models do not disappear that moment, but they blur it across prompt writers, training sets, model providers, and downstream editors. That blur is why boards and general counsel can no longer rely on registrations alone to defend a position.

Every pillar is under new pressure because AI shifts both the input and the output side of creation. On the input side, models are trained on huge corpora that may contain copyrighted works, patented disclosures, and third-party trade secrets, and provenance is often opaque. On the output side, generative systems can produce text, images, music, code, and even drug candidates that look protectable but may have no qualifying human author. Between these two sides sits the model itself, a piece of software plus weights that most AI companies treat as the crown jewel. The most useful mental model is that AI has turned IP from a set of four separate boxes into a single pipeline where every stage must be defended. Our overview of AI’s impact on intellectual property law explains why legacy concepts keep tripping over this pipeline.

Breaking the pillars also means breaking the reflex to over-rely on one regime. A company that patents its model architecture still has to decide whether disclosure in the patent destroys the trade-secret premium on its weights. A newsroom that copyrights its journalism still has to deal with scrapers that train chatbots on those stories without a license. A consumer brand that owns a trademark still has to police lookalike AI-generated content that confuses the origin of goods. The practical answer for AI IP protection is a layered portfolio that moves deliberately across registration, contracts, and secrecy. The goal is to make each pillar reinforce the others instead of leaving gaps where AI can slip through.

Copyright After Thaler and the Anthropic Ruling

Building on that pipeline view, copyright is the IP regime where the courts have moved fastest and most visibly in the AI era. The United States Copyright Office has repeatedly held that a work must have a human author to qualify for registration, and the D.C. Circuit affirmed that principle in Thaler v. Perlmutter when it rejected Stephen Thaler’s attempt to register a picture whose only author he listed as a machine. The Supreme Court denied certiorari in that case in April 2026, which locked in the human-authorship floor across the federal system. Agencies have added layers of nuance on top of that floor, but the floor itself is now very hard to move. AI IP defense therefore starts with proving human contribution during creation.

The Copyright Office has refined its guidance in three parts, covering digital replicas, copyrightability of AI outputs, and legal implications of training. Its view is that purely prompt-driven outputs are not protected because prompts function as instructions rather than controlled expression. A human can still claim copyright in the arrangement, selection, or meaningful edits layered on top of an AI output, but the AI-generated core itself is in the public domain. This rule creates a practical workflow question for every creative team. Teams that want to own what they ship must either generate from scratch, add substantial human expression on top of AI drafts, or move to fully human tooling for the protectable layer.

The fair-use front has moved in the other direction. In Bartz v. Anthropic, Judge William Alsup held in June 2025 that training large language models on legally acquired books was a transformative fair use, while piracy of books through shadow libraries was not. The Anthropic ruling is the most consequential training-data decision to date because it separated the input acquisition question from the training activity itself. A follow-on class settlement reported in September 2025 at $1.5 billion focused on pirated copies, not the training itself. This pair of outcomes tells IP owners two things: train on licensed content or risk the piracy count, and expect fair-use defenses to survive when input provenance is clean.

Beyond text, the AI-generated Ghibli-style images and ban risk shows that visual artists and studios are pushing their own copyright claims into the courts and the platforms. Andersen v. Stability AI survived a motion to dismiss and is now in discovery on direct infringement and induced infringement theories. The music industry has sued Suno and Udio over output similarity and unlicensed training. These lawsuits will not close out in a single ruling, but they are building a doctrine that treats verbatim or near-verbatim AI output of protected works as presumptively risky. Copyright is alive and well, it is just being enforced through output similarity more than through old doctrines about photocopying.

Patents, AI-Assisted Invention and the USPTO Framework

Turning from copyright to patents, the United States Patent and Trademark Office issued its first formal guidance on AI-assisted inventions in February 2024, and refined it through 2026. The core rule is that a named inventor must be a natural person who made a significant contribution to each claim of the application. The USPTO’s Federal Register notices on AI set out the five Pannu factors the Office uses to assess that contribution. Companies filing AI-assisted applications must therefore document which humans shaped claim scope, selected training data, and validated the invention. AI-era IP protection through the patent system requires keeping those records in the lab notebook, not reconstructing them later.

Patent law also grapples with subject matter eligibility for AI inventions themselves. Claims that recite generic application of a mathematical model to a computer continue to be rejected under Alice as abstract ideas, while claims that tie the AI to a concrete technical improvement in a system tend to survive. Applicants in drug discovery, materials, chip design, and industrial automation have had better outcomes than applicants in business methods or generic classification. A patent group inside a corporate IP team should therefore triage AI-related disclosures with eligibility in mind, not only novelty. Filing an abstract AI claim wastes money and tells competitors what you are working on. A rejected abstract application also removes the trade-secret option for the same invention, because the disclosure is now public.

Internationally the picture is uneven. The European Patent Office has signaled openness to AI-assisted inventions so long as a human inventor is named, while the UK Supreme Court in Thaler v. Comptroller-General held that an AI cannot be a named inventor under UK law. China continues to grant AI-related patents at a very high volume, and WIPO’s generative AI patent landscape report shows Chinese applicants leading the global count. For multinational filers this means harmonizing an inventor-attribution record across jurisdictions. It also means accepting that aggressive filing in China may be necessary even for defensive purposes. The patent system is still useful, it just rewards operators who treat AI-assisted invention as a discipline rather than an afterthought.

Trademarks, Brand Dilution and AI-Generated Lookalikes

Shifting to trademarks, generative AI has made it trivial to produce convincing lookalike packaging, logos, and brand voices in seconds. Trademark law still works the way it did before, around likelihood of confusion, dilution, and source identification. What has changed is enforcement volume. A platform that once saw a hundred counterfeit image uploads a day may now see tens of thousands, and brand teams cannot police them by hand. The strongest brands are shifting from case-by-case takedowns to automated detection paired with contractual commitments from AI platforms. Our write-up on the AI-generated Ghibli images and the ban-risk conversation captures how fast this shift has happened.

AI also creates new trademark risks around personality and voice. The Lanham Act false-association theory has become a tool for celebrities and voice actors to go after AI-generated imitations. Scarlett Johansson’s dispute with OpenAI over a voice reminiscent of her Her performance was an early flashpoint. State-level right-of-publicity statutes add another layer, particularly in Tennessee’s ELVIS Act and California’s AB 2602. Brand teams should map every registered mark, trade dress, and right-of-publicity position across jurisdictions. They should also assume that bad actors will try to misuse their brand in AI output, and plan monitoring and enforcement accordingly.

Trade Secrets as the Load-Bearing Pillar of AI IP

Looking across the three regimes above, trade secrets have quietly become the load-bearing pillar of AI intellectual property protection. Copyright will not touch model weights, patents disclose them to the public, and trademarks have nothing to say about them at all. Trade secrets, by contrast, can protect weights, training recipes, prompt libraries, fine-tuning datasets, and evaluation harnesses as long as reasonable secrecy measures are in place. IIPLA has tracked a clear shift toward trade-secret strategies across AI-heavy companies in 2024 through 2026. The flexibility of this regime is its biggest advantage, and its fragility is its biggest risk.

The Defend Trade Secrets Act and state Uniform Trade Secrets Act counterparts require two things from an owner: information with independent economic value from secrecy, and reasonable measures to keep it secret. Modern AI leakage happens not through corporate espionage but through everyday employee behavior with public chatbots. A Stradling survey found roughly 65% of organizations had no formal policy controlling such use. Samsung’s 2023 incident, where engineers pasted proprietary source code into ChatGPT, is the canonical example. Once information lands in a training corpus or a third-party log, courts can find that reasonable measures were never in place.

Beyond leakage, trade-secret protection depends on identification. Teams cannot protect what they have not inventoried. A good AI-era trade-secret program lists each protectable asset and ties it to concrete access controls, document classifications, and departure procedures. For models this often means tiering weights by sensitivity, blocking export of base weights, logging every inference request, and encrypting weights at rest. For datasets this means watermark-style canaries, access logs, and strict vendor isolation. For prompts and chains this means versioning in a secrets-safe repository instead of a shared notebook. These controls need to be defensible in court, not aspirational.

Trade-secret strategy also carries a timing cost. Unlike a patent, a trade secret can live forever, but it dies the moment the secret leaks. This changes incident response. A security event is now also an IP event, and legal must be on the breach call from minute one to preserve privilege and build the record. The companies that handle this well treat AI security and trade-secret defense as one program with two reporting lines rather than two separate programs that meet at an incident. Doing it that way prevents the gap where no single team owns both exposure and remediation.

Training Data, Fair Use and the Shadow Library Problem

Beyond trade secrets, training data has become the most litigated battleground in AI IP. The core doctrinal question is whether copying a protected work into a training set qualifies as fair use. Courts have begun to answer that in pieces. The Bartz v. Anthropic ruling treated training on legally acquired books as transformative, following a pattern consistent with Authors Guild v. Google. The same ruling treated ingestion of pirated books from shadow libraries as infringing, which is why the follow-on settlement focused on acquisition rather than training activity. This split matters because it tells IP owners how to position cases and lets AI developers know which workflow changes reduce risk most.

The New York Times lawsuit against OpenAI is the test case for whether verbatim or near-verbatim output changes the fair-use analysis. Memorization of training data, where the model can reproduce protected text on demand, is the crucial technical fact that could flip a case from transformative to infringing. Plaintiffs cite dozens of examples where ChatGPT reproduced Times articles nearly word for word. OpenAI counters that memorization is a bug being fixed, not a feature, and that normal use rarely surfaces it. The Supreme Court has not resolved the question, and lower courts are testing both theories in parallel. Our review of AI copyright crises on live platforms shows how quickly these questions spill into user-facing products.

Internationally the training-data question has produced different answers. The European Union AI Act and the related text and data mining exception under the Copyright in the Digital Single Market Directive allow training only when rightsholders have not reserved their rights. Japan’s AI copyright guidance permits training more broadly but limits use of outputs that imitate specific protected works. The UK has consulted on a text and data mining exception but backed off a broad version after creative-industry pushback. For a developer with a global user base the right default is licensing, with careful ingestion controls and a legal position on each jurisdiction. That default is more expensive but far cheaper than a long-running lawsuit.

Implementation: AI Governance, Contracts and Vendor Due Diligence

Building on the training-data analysis, governance and contracts have become where most AI IP outcomes are actually decided. A written policy that bars confidential information from public chatbots, paired with enforcement, is a near-zero-cost control that saves most trade-secret cases later. Governance programs that combine clear acceptable-use rules, approved-tool lists, and quarterly training are now table stakes rather than good-to-have. Our overview of AI governance trends and regulations explains how these programs fit into broader risk frameworks like NIST AI RMF and ISO 42001.

Vendor contracts are the second lever. Every enterprise agreement with a model provider should address training rights, output ownership, indemnification for third-party IP claims, data residency, and audit rights. Microsoft, Google, OpenAI, and Anthropic now offer IP indemnities for enterprise customers, but each indemnity has carve-outs that lawyers must understand. An AI vendor that refuses to allow audit of training provenance should be a red flag, not a procurement convenience. A good diligence checklist also probes model weight storage, fine-tuning retention, and the vendor’s own vendor chain. Every one of those links can carry an IP risk back to the enterprise buyer.

Employee contracts need to catch up too. Standard proprietary information and inventions agreements were written for a world where an employee typed on a company laptop. In the AI era an employee may generate work with an assistant that keeps logs, trains on inputs, and operates under its own terms. Updated agreements should address prompt ownership, output ownership, obligations when AI tools are used, and the duty to preserve confidentiality when interacting with third-party AI. Separation agreements should list AI systems and accounts the departing employee had access to. These are small contract edits that produce an outsized shift in enforceability.

Technical Controls: Watermarking, Provenance and Model Access

Stepping from contracts to engineering, technical controls close the gaps that policies cannot. Content provenance standards like the Coalition for Content Provenance and Authenticity, known as C2PA, let a publisher sign a creative asset with cryptographic metadata that survives downstream edits. Watermarking AI outputs and signing human-authored work is now the fastest way to build a defensible evidentiary record for both ownership and infringement. Google’s SynthID, OpenAI’s output signing, and Adobe’s Content Credentials show how quickly the standards ecosystem is maturing.

Model access control is the other technical anchor. Enterprises can protect model artifacts by restricting API access by role, logging every inference, blocking export of base weights, using AI cybersecurity controls, and using secure enclaves for fine-tuning. Our review of AI data privacy best practices for businesses maps the same controls to data protection, and the overlap with trade-secret defense is almost complete. Teams that pair these controls with data-labeling hygiene from training-data annotation workflows start to build a clean provenance chain. The chain is what a court looks for when a trade-secret claim comes up.

Ethical Questions Behind AI Authorship and Ownership

Beyond the mechanics, AI IP raises ethical questions that deserve their own weight. The first is distributive. When a model trained on millions of artists’ and writers’ work generates a near-substitute, the economic value flows to platforms and prompt users, not to the original creators. Societies that care about creative careers must decide whether existing opt-out and licensing regimes are enough. AI ethics and laws has sharpened on exactly this point since 2024.

The second question is consent. People’s voices, images, and writing styles are being used in AI training without explicit permission in many jurisdictions. State and national right-of-publicity rules give some protection, but coverage is patchy. Consent-based data commons and opt-in licensing pools are emerging as a middle path between pure open-web scraping and full gatekeeping. These structures let creators share in upside without blocking innovation outright. They also give enterprises a cleaner provenance story to tell their regulators and shareholders.

The third question is accountability. If an AI output infringes a protected work, who bears responsibility between the user, the deployer, and the model provider? Current doctrine treats humans in the loop as the responsible actors, but that cannot scale to fully autonomous agents. The AI deepfakes trust crisis is accelerating interest in strict-liability regimes for the most dangerous output categories. Any ethical IP framework has to allocate responsibility clearly enough that victims can get relief and developers can build.

Industry Impact: Media, Software, Life Sciences and Finance

Shifting from ethics to sectors, media companies face the most immediate AI IP pressure. News organizations fund reporting that AI assistants then summarize without a referral click, and the New York Times, Dow Jones, and others have sued or licensed in response. Music labels and film studios face deepfake voice and face risks that threaten talent contracts and theatrical value. Media companies that treat AI licensing as a new revenue line, with mandatory attribution and usage caps, are faring better than those that try to litigate every misuse. The AI music bots on streaming is a live example of this balance.

Software vendors sit in the opposite position, because they build with AI assistants and sell AI features, and they also need to protect their own code. GitHub Copilot’s license terms, the Microsoft Copilot customer agreements, and the Open Source Initiative’s stance on AI-generated code all affect how a software company can claim ownership. Mixed training sets that include permissive and copyleft code raise contamination risks. The practical defense is a disciplined use policy, a license-compliant AI coding assistant, and a scanner that checks AI-generated code against known repositories before merging. Software teams also need to stand up red-team processes to catch output that leaks internal prompts or weights.

Life sciences has become one of the fastest AI IP growth areas because of drug discovery and materials design. The USPTO’s AI-assisted invention guidance has direct bite here because many disclosures involve machine learning models proposing candidate molecules or structures. A life-sciences filer must document the human contribution at the claim level, maintain training data records, and watch for prior-art disclosures in model-generated pre-prints. Clinical trial plans now include AI documentation blocks. Our note on AI for regulatory filings and pharma factories walks through how companies track those records across groups.

Financial services firms treat AI models as crown-jewel trade secrets, in part because patent eligibility is narrow for financial methods. Hedge funds, trading firms, and credit scorers protect models through tiered access, strict employee covenants, and strong contractual terms with cloud providers. Regulatory pressure from the SEC, CFPB, and prudential regulators adds a layer of model risk management that overlaps with IP protection. The result is that the IP program and the model risk program share people and tooling. Finance has been doing this longer than many sectors, and other industries can learn a lot from the discipline it already uses.

Key Insights on Protecting Intellectual Property in the Age of AI

• WIPO’s 2024 generative AI patent landscape report counts more than 54,000 generative AI patent families filed from 2014 to 2023, an eightfold growth that signals how quickly AI inventions are reshaping the patent record. This asset base will soon dominate litigation calendars because disputes typically trail filings by five to seven years.
• The U.S. Copyright Office’s AI inquiry attracted more than 10,000 comments during its 2023 and 2024 listening phase, which is one of the largest responses in its history and shows that creators see copyright as an active front. The Office has translated that volume into a three-part guidance set that already shapes registration decisions.
• The Supreme Court denied certiorari in Thaler v. Perlmutter in April 2026, which locked in the human-authorship requirement for U.S. copyright registration and ended the most direct legal challenge to date. Enterprises can rely on this floor when drafting ownership clauses, but international positions still vary.
• Stradling’s data defense dilemma report found that roughly 65% of organizations lack a formal AI use policy covering public chatbots, which leaves most trade-secret claims vulnerable to a reasonable-measures challenge. A short written policy and a block list of unapproved tools would close most of this gap in a week.
• USPTO guidance issued February 13, 2024 requires that a natural person make a significant contribution to each claim of an AI-assisted invention, applying the long-standing Pannu factors. Teams that cannot document that contribution risk losing patents on invalidity grounds years later.
• Judge William Alsup’s June 2025 ruling in Bartz v. Anthropic held that training on legally acquired books is transformative fair use while ingestion through shadow libraries is not, which split the input question from the training activity. A reported $1.5 billion settlement focused on acquisition, not training, following the piracy count.
• Proskauer’s AI trade-secret framework highlights the exposure that comes when employees query public chatbots with confidential information, because the information may be retained for training and lose secrecy protection. Framework-aligned programs tie acceptable use, access controls, and incident response into a single policy stack.

Pulling those insights together, IP protection under AI is no longer a set of separate playbooks for copyright, patent, trademark, and trade secrets. The pipeline from training data to model weights to outputs forces each regime to interact, and gaps in one area now expose assets in another. Trade secrets have become the integrator because they can cover the parts of the pipeline the other regimes miss, while governance and contracts hold the whole thing together. A modern program looks less like an IP docket and more like an operating system that runs across legal, security, and engineering. The companies that treat it that way are already pulling ahead in both filings and litigation outcomes.

How the Four IP Pillars Compare Under AI Conditions

The four classical IP pillars behave very differently once AI touches the pipeline, and the table below maps the practical tradeoffs side by side. Read it as a decision aid before picking which regime to lean on for a given AI asset.

Dimension	Copyright	Patent	Trademark	Trade Secret
Transparency required	Deposit on registration	Full public disclosure	Registered mark on public register	Kept secret
Participation of AI	Human author required	Human inventor required	Human source of use required	Any information with value
Trust and provenance	Registration certificate	Issued patent and file history	Register plus common-law use	Secrecy measures and NDA chain
Decision making	Copyright Office examines	Patent examiner reviews claims	Examiner reviews mark	No agency, owner self-decides
Misinformation and fakes	Output similarity cases growing	Rarely a vector	Lookalike packaging surging	Chatbot leakage is top risk
Service delivery	Enables licensing markets	Supports exclusivity licensing	Protects consumer trust	Enables model-as-a-service scale
Accountability	Infringement suits	Infringement suits	Confusion and dilution claims	Civil misappropriation claims

Real-World Examples of AI and IP Protection in Action

Three operational examples show how major firms are applying the legal frameworks and controls above to live product decisions right now. Each illustrates a different combination of licensing, governance, and incident response under AI conditions.

Adobe’s Firefly Trained on Licensed Stock

Adobe trained and deployed its Firefly generative image model on a 100 percent licensed corpus using Adobe Stock, openly licensed content, and public-domain material when it rolled it out and launched Firefly in March 2023. Adobe also pledged a commercial IP indemnification covering enterprise outputs, which effectively buys legal cover for business users. The company then paid contributing Adobe Stock photographers roughly $300 million in a one-time bonus tied to training-set inclusion by late 2024, according to multiple business reports. The outcome has been fewer lawsuits and stronger enterprise adoption, because risk-averse buyers prefer the licensed-training story. The limitation is that Adobe’s catalog, while large, is narrower than open-web scraped sets, and output quality on some niches lags competitors. The example nonetheless shows that a licensed-training strategy can be commercially viable at scale.

Shutterstock’s Contributor Fund

Shutterstock deployed its contributor fund and paid out roughly $4.1 million to contributors over six months (a six-figure payout per week) which lifted contributor retention measurably of its contributor compensation fund after launching its AI image generator and an OpenAI partnership in 2023, according to Shutterstock’s own press materials. Contributors receive a share of revenue when their work is used for training and ongoing earnings as generated images enter the library. The fund is a middle path between full opt-out and uncompensated training, and it allows Shutterstock to position its outputs as cleaner than rivals. The limitation is that the per-contributor payout can be small for individual images, and some photographers argue the ratio favors the platform. Even with that critique, the fund is one of the few working licensing mechanisms at scale in generative AI today. It points toward where rightsholder deals may land across media more broadly.

Samsung’s Trade-Secret Exposure Through ChatGPT

Samsung’s semiconductor arm reported in April 2023 that engineers had pasted confidential source code and meeting notes into ChatGPT on at least three occasions in a three-week window, prompting a company-wide ban on generative AI tools reported by Bloomberg. The leaks were not acts of espionage, they were ordinary engineers trying to work faster. The company moved quickly to an internal assistant to meet the demand without the leakage path, and it tightened its device-level policy. The outcome within weeks was a 40 percent reduction in uncontrolled public-AI use, but Samsung’s ability to assert trade-secret protection for the leaked code was already damaged. The limitation is that the ban created friction and some workflows quietly moved to personal devices. The episode is the most cited example in enterprise AI acceptable-use policy debates today.

Case Studies in AI-Era Intellectual Property Disputes

Three live disputes anchor the doctrinal shifts above and show where courts are drawing the lines on training, outputs, and jurisdiction. Each case has produced early findings that already shape risk management choices across industries.

Case Study: Bartz v. Anthropic on Training Data Fair Use

The Northern District of California handled the first major decision on generative AI training fair use in Bartz v. Anthropic. The authors faced a core problem: Anthropic had allegedly copied their books to train Claude without a license, and some copies allegedly drawn from pirated shadow libraries. Judge William Alsup held in June 2025 that training on legally acquired books was transformative and therefore fair use, but that ingestion of pirated copies did not benefit from fair use. The ruling walked through more than 20 pages of fair-use analysis and split the input question from the training activity cleanly. Observers across the IP bar immediately treated the opinion as the first durable 2025 doctrine on generative AI training.

The court’s solution deployed a split analysis, and Anthropic later reached a reported class settlement of about $1.5 billion covering pirated-copy claims across approximately 500,000 books, with per-work payments reported near $3,000. The impact split neatly along acquisition lines: clean sourcing was treated as transformative, while piracy was not, which gave model builders a clear operational signal. A limitation of the ruling is that the training question for creative works outside books, such as music and images, remains open and lower courts are still testing output-similarity theories. The case nonetheless anchored a doctrine that most practitioners now treat as the baseline for training-data risk management, with roughly 500,000 affected books and settlement payouts reported near $3,000 per work. That precedent now guides how enterprise developers architect their licensing pipelines.

Case Study: Andersen v. Stability AI and Visual Artist Claims

Sarah Andersen and a group of visual artists faced the problem that Stability AI, Midjourney, and DeviantArt had built models on their work, and they sued in early 2023, alleging that the companies trained diffusion models on billions of copyrighted images scraped from the LAION dataset without permission. The docket in Andersen v. Stability AI shows the court allowed direct and induced-infringement claims to proceed past multiple motions to dismiss, which pushed the case into discovery. By mid 2025 more than 5 billion images were reportedly implicated in the training discovery requests, and plaintiffs had filed over 20 declarations supporting the harm claim. The judge narrowed some theories while keeping the core infringement counts alive.

The court’s solution was to let the direct and induced-infringement claims proceed, which deployed real discovery power. The impact has been large even before any trial because defendants have had to produce detailed training-set manifests and discovery on prompt behavior. Artists have used the case to push platforms toward opt-outs and compensation structures. A limitation is that the plaintiffs must still prove substantial similarity at the output level, and courts are divided on how much a model’s capacity to imitate a style counts as infringement when individual outputs are not direct copies. The case will likely frame how output-similarity theories interact with training inputs for years, and artists have cited it when pushing more than ten AI platforms toward opt-out mechanisms since late 2024.

Case Study: Getty Images v. Stability AI UK Judgment

Getty Images faced the problem that Stability AI had built Stable Diffusion by allegedly copying millions of Getty-watermarked images, so Getty sued in the UK High Court, arguing that the company deployed those copies for training and that Stable Diffusion outputs sometimes reproduced the Getty watermark. In November 2025 the court delivered a partial judgment that rejected the direct copyright claim on training in the UK because training did not take place there, while finding limited trademark infringement based on the watermark appearance in generated output. The dispute lasted roughly 3 years before judgment and implicated over 12 million images in the discovery record. The judgment also flagged more than 5 distinct output categories where the Getty watermark survived.

The solution the court introduced was to split the training and output questions by location, and the impact extended beyond the UK because the opinion read as a signal that jurisdiction over training workflows will matter as much as the training itself. Getty continues to press parallel litigation in the United States, which turns on the fair-use doctrine and is more friendly to large discovery into training sets. A limitation of the UK ruling is that it rested on highly specific facts about where model training occurred, which plaintiffs will try to distinguish in future cases. The case still shows how output-side trademark theories can succeed where training-side copyright theories fail, and it will influence how AI companies choose jurisdictions for their training infrastructure in 2026 and beyond.

The Risks of Doing Nothing About AI IP Exposure

Looking at the sector-level view, inaction carries measurable costs across every industry. A company that fails to update its policies may lose the ability to assert trade-secret status after a leak, even when the underlying information was obviously confidential. Boards that approve generative AI deployments without legal review can find themselves sued by content owners, employees, or customers whose data entered a training set. Doing nothing is not a neutral choice; it is a decision to accept the worst-case outcome of every open risk vector at once. Our overview of the AI deepfakes trust crisis shows how this exposure can escalate quickly in a public setting.

Doing nothing also creates a competitive handicap. Competitors that have trained employees, locked down vendor terms, and deployed watermarking can move faster on new AI use cases because they have a lower residual risk. They can indemnify customers more confidently. They can also respond to incidents without panic, which reassures regulators and insurers. A company that lags on these controls often finds its AI roadmap stalled by legal reviews that competitors finished a year earlier. The cost of inaction shows up in the roadmap, not just the lawsuit docket.

The reputational dimension is harder to price but just as real. Media coverage of AI-driven IP breaches travels fast in the trade press and on social platforms. Enterprise customers routinely ask for AI governance documentation before signing contracts, and smaller organizations often fail those questionnaires. Investors increasingly treat AI governance as part of ESG risk. The companies that move now can get ahead of these expectations without a crisis. Those that wait will be measured against the ones that moved, not against the companies that waited too.

The Future of Intellectual Property Law in an AI World

Building on the risk view, the next three years will reshape AI IP law more than the last ten. WIPO has been convening member states to discuss training-data transparency and the possibility of a new international instrument on AI and IP. The U.S. Congress has advanced the NO FAKES Act to create a federal right against unauthorized AI replicas of real people, and bipartisan interest suggests it may pass in some form. The EU AI Act’s transparency requirements will drive developers to publish training-data summaries, and that documentation will fuel more litigation and more licensing. AI intellectual property protection will look very different by 2028, even if the legal foundations remain recognizable today.

Technical standards are moving with the law. C2PA and related provenance frameworks are being baked into camera firmware, newsroom publishing systems, and major AI platforms. Watermarking methods are improving even as adversarial research shows they can be stripped. Model licensing marketplaces for text, image, music, and code are emerging to monetize what was previously scraped for free. These standards will intersect with regulation to create a new baseline for AI IP hygiene, including cleaner training data annotation practices. Companies that help build the standards gain early influence over their shape.

Looking further ahead, the enduring question is whether existing IP doctrines can accommodate AI without a fresh statutory layer. Some scholars argue that modest doctrinal updates and enforcement will suffice. Others argue that a new sui generis right may be needed for AI-assisted works. The commercial reality is probably a hybrid: licensing deals, revised registration standards, AI lawyer roles, and sector-specific statutes will do most of the work, with doctrine evolving in the gaps. AI IP protection is not a problem that gets solved once. It is a program that needs renewal every year as technology and law pull each other forward.

Frequently Asked Questions on Protecting Intellectual Property in the Age of AI