AI and Intellectual Property Law

Introduction

AI and intellectual property law has become the fastest moving corner of commercial law in 2026 and the ground is shifting under every creator, engineer, and in-house counsel. A September 2025 settlement in Bartz against Anthropic sent about 1.5 billion dollars to authors whose books were pirated into a training set. The settlement is reshaping how foundation models are built across the frontier labs. US courts have split on fair use while the USPTO has issued binding guidance on inventorship for AI-assisted work. The EU AI Act now compels foundation-model providers to summarize what they ingested from the open web and licensed sources. This guide walks through the rights at stake and gives a working playbook that any in-house team can start using this quarter. The old playbook written for human-only creativity does not map cleanly onto diffusion models or large language models. The practical answer lies in the current wave of AI copyright lawsuits now defining the field, which carries actionable lessons for product and legal leaders.

Quick Answers on AI Copyright and IP Rights

Who owns the copyright in a purely AI-generated image under AI and intellectual property law?

No one owns the raw diffusion-model output under US law. The US Copyright Office requires human authorship, so pure machine output sits in the public domain. A human can register selective arrangement or edits, but not raw output.

Can an AI system be listed as the inventor on a US patent?

No. The Thaler decision and the USPTO 2024 Guidance confirm that only a natural person can be named as inventor. A human must make a significant contribution to each claim.

Is training a large language model on copyrighted books fair use?

Partly. The Thomson Reuters case rejected fair use for a competing dataset in 2025. The Bartz class action settled for about 1.5 billion dollars over pirated books. Licensing is now the safer commercial path.

Key Takeaways for AI IP Teams

• AI and intellectual property law now spans copyright, patents, trademarks, and trade secrets at the same time, and every enterprise AI product touches all four.
• The 2024 USPTO Inventorship Guidance and the Bartz settlement mean documentation of human contribution and training-data provenance is no longer optional.
• The EU AI Act Article 53 training-data summary obligation is in force for general-purpose models, and the published template exposes US copyright holders to new claims.
• Licensing deals between publishers and AI labs are becoming the default commercial path, which reshapes editorial economics for the next decade.

Understanding AI and Intellectual Property Law in 2026

AI and intellectual property law covers the rules that decide who owns, controls, and profits from outputs produced by a trained model across copyright, patents, trademarks, and trade secrets as applied to AI inputs and outputs.

The Core IP Rights Being Reshaped by Generative AI

Building on the shift above, four classical rights now carry the weight of AI-era disputes. Copyright is being tested on both the input side through training-data lawsuits and the output side through authorship questions. Patents are being reshaped by the significant contribution test for AI-assisted inventions and by Section 101 subject-matter rulings. Trademarks are confronting deepfake endorsements, synthetic voice clones, and generative mimicry of brand identity across consumer segments. Trade secrets are becoming the dominant protection vehicle because they do not require disclosure of model internals the way patents do.

Shifting from doctrine to implementation, the scope of each right now overlaps in ways they did not before. A single AI product release can trigger copyright, publicity, trademark, and trade-secret issues in one afternoon of commercial launch. Consider an AI music tool that lets a user clone a vocalist’s style from a prompt and release the result commercially. The training set raises copyright exposure while the voice itself raises right-of-publicity under state laws. The artist’s name in the prompt raises trademark exposure, and the voice-model weights raise trade-secret considerations about storage and access.

Beyond the single-product view, enterprises are reorganizing around this overlap in operational structure. Many firms combine patent, copyright, and trade-secret counsel into a single AI legal pod rather than keeping them in separate practice groups. This mirrors the way product teams moved from siloed engineering tribes to full-stack feature pods in the previous decade of software development. It also acknowledges that AI legal questions rarely stop at one doctrine or one practice area. Fragmented advice tends to miss cross-right exposures that keep landing in litigation across the sector.

Looking ahead, the World Intellectual Property Organization has been tracking this convergence in its ongoing AI dialogue with member states. WIPO working papers suggest the direction of travel for member-state law. Most member states still require a human author, but many are considering sui generis protection for substantial human investment in AI output. The database directive model used in Europe is one template under study by national legislatures. A company planning a 2030 global product has to prepare for different authorship fictions in different countries, with no single safe harbor that works everywhere today.

Copyright Ownership When the Author Is a Model

Turning to the clearest American rule, it is also the one most people get wrong in practice. The US Copyright Office refuses registration for anything made only by a machine, no matter how expressive the output appears. A human must contribute copyrightable expression, which can be selection, arrangement, or substantial editing of generated material. The Zarya of the Dawn and Theatre D’opera Spatial decisions set the ceiling for registration. Everything above that human-authorship line sits in the public domain under current doctrine. That rule may surprise anyone who assumed a prompt alone makes a work theirs.

On the open international front, other jurisdictions have taken different turns on authorship. The United Kingdom still formally recognizes computer-generated works under the CDPA 1988, assigning authorship to the person who made the arrangements. Japan allows some AI output into its registration system when a human made creative choices in the process. China has swung both ways in different Beijing Internet Court rulings, but recently extended protection where a user’s prompts and iterations show creative effort. The practical approach for a cross-border product is to treat every AI output as unregistrable in the United States until a human adds real expression. Log that human contribution carefully in a way that would survive discovery and audit.

Patent Inventorship After the USPTO 2024 Guidance

In practice, the United States Patent and Trademark Office published its Inventorship Guidance for AI-assisted inventions on February 13, 2024, after extensive public comment. That guidance has quietly reshaped how claims are drafted across the US patent bar. The guidance confirms that an AI system cannot be listed as inventor, which was already settled by the Federal Circuit in Thaler. The guidance also sets a workable test for human contribution across AI-touched claims. At least one natural person must have made a significant contribution to each claim of the patent. The Pannu factors adapted from case law on multi-inventor patents provide the measurement framework used in practice.

Beyond the formal test, significant contribution is not a bright line, and that is the deliberate design choice. The guidance says that merely owning the AI tool does not qualify as inventorship under current doctrine. Merely instructing a general-purpose model to attempt a task does not qualify either. Merely recognizing that the model’s output happens to be useful does not count toward inventorship. A human must contribute to the conception of the specific solution being claimed. Companies with mature pipelines now require inventors to keep a traceable notebook of their contribution before any patent counsel reviews the disclosure.

Moving on to enforcement, the downstream risk is a bigger worry than the application process itself. A patent that issues without a documented human contribution to each claim can be invalidated years later during litigation discovery. The practical fix is to run contribution interviews before filing with the responsible inventors. Record the AI tool used, the specific prompts, the human refinements, and the moment of inventive insight for each claim. That paper trail is what will survive a post-grant proceeding when challengers probe inventorship. The court ruling redefining AI patent standards underlined the stakes.

The Training Data Problem and the Fair Use Fight

Despite the many pre-2024 predictions, the training-data lawsuits are where the biggest money sits and the biggest doctrinal uncertainty remains. Thomson Reuters against Ross Intelligence became the first US summary judgment to find that copying a competitor’s commercial dataset to train an AI competitor was not transformative fair use. The ruling was narrow and limited to a specific legal-research dataset context. It signaled that fair use is not an automatic shield for training on copyrighted material at commercial scale. The Disney lawsuit challenging AI copyright boundaries is testing similar questions for creative content and media libraries.

Shifting to the money side, the Bartz against Anthropic settlement went the other direction procedurally. The deal still cost roughly 1.5 billion dollars in total. The parties settled before trial over around 500,000 pirated books in the training corpus and allied materials. No fair-use ruling emerged on the merits of that case for either side. The settlement did not endorse fair use, it priced the risk under one specific set of facts. Anthropic chose to pay about 3,000 dollars per book rather than roll the dice on a Northern District of California jury trial. Other labs read that number and started cutting deals with major publishers rather than defending similar suits.

Looking at the broader map, the doctrinal split is now visible across jurisdictions and judges. The Southern District of New York, the Northern District of California, and the District of Delaware have reached different conclusions on summary judgment. The Second Circuit and Ninth Circuit will likely weigh in during 2026 to 2027 on appeal. In the meantime, defending a US training-data suit now costs enough that licensing is often cheaper than litigation at scale. The Meta training-data piracy lawsuit is the next big case to watch in this active litigation sector.

Beyond US borders, the picture is different because national IP regimes diverge sharply on training. The EU AI Act compels disclosure of training data without granting permission to use any specific work. A disclosed training set makes it much easier for European rights holders to bring parallel suits in their own courts. The United Kingdom introduced a controversial text and data mining exception consultation that was repeatedly revised after artist pushback. Japan has one of the broadest statutory exceptions for training, which has made it a favored jurisdiction for compute-heavy pretraining runs. Any enterprise shipping a model trained on scraped data needs a jurisdiction map and a licensing roadmap for every target market.

Trademark Law in the Age of Deepfakes and Synthetic Brands

Turning to trademark, the doctrine was built for logos, slogans, and brand identifiers rather than for synthetic media. Now it is being pointed at voices, faces, and personas by every major rights holder. Deepfake endorsements that use a celebrity’s likeness without consent trigger Lanham Act false endorsement claims in federal court. State right-of-publicity claims apply in parallel, and in Tennessee the ELVIS Act specifically protects voice. The pending federal NO FAKES Act proposes a national right of publicity in a single federal statute. The 2025 Sora launch made this urgent for every rights holder with a public face or recognizable voice. The Sora deepfake wave accelerated enforcement work across the entertainment industry.

Rounding out the trademark picture, brands themselves face a different problem than celebrities. Generative models that mimic a brand’s visual identity or house style can create a flood of look-alike merchandise and marketing material online. Trademark owners now run routine model probes to see whether their marks appear in outputs across popular generative products. A product that reliably reproduces a protected logo is exposed to contributory infringement claims that follow the model, not just the end user. The Hollywood versus AI copyright showdown is partially a trademark fight dressed in copyright clothes for strategic reasons.

Trade Secrets, Model Weights, and Leakage Risk

Stepping back to look at the quietest workhorse of AI protection, trade secrets carry much of the practical enforcement weight in 2026. Model weights are not patentable in most jurisdictions, and copyright protection for weights is narrow and contested. Trade-secret law carries most of the enforcement weight for the one asset labs really care about commercially. The Defend Trade Secrets Act gives US holders a federal civil cause of action across state lines. Most leading labs now layer internal access controls, employee NDAs, non-compete and non-solicit terms, and output monitoring. A weight set that gets leaked to a torrent site is practically unrecoverable, so prevention is the only workable strategy for frontier labs.

Shifting to inputs, the counter-risk is more subtle than the leaked-weights scenario everyone fears. When enterprise customers upload confidential data as prompts or fine-tuning examples, that data can end up embedded in model behavior. A pharmaceutical company that fine-tunes a model on internal trial data may create a path for a competitor to extract proprietary information through carefully crafted queries. The defense is a layered combination of contractual carve-outs, data residency controls, output filters, and periodic red-team probes. Treat the model like an untrusted employee rather than a filing cabinet, and apply the same principles to it. Vendor contracts must disclaim training on customer inputs, retain audit rights, and allow deletion of fine-tuned adapters on request.

On top of secrecy, weight-level security intersects with export controls in several significant ways. US BIS rules now restrict transfer of frontier model weights above certain compute thresholds to specified foreign nationals. Trade-secret protection and export compliance share a common operational layer in most sophisticated labs. An enterprise serious about keeping its model behind a defensible moat should treat weights as controlled technology, not as ordinary software files. That reclassification changes how weights are stored, who may access them, and what rights of inspection the government reserves. The result is a tighter operational perimeter than labs typically maintained before the current regulatory wave.

How Companies Are Protecting AI-Assisted Inventions in Practice

For teams working on AI-assisted inventions, mature research organizations have replaced the standard invention disclosure form with an AI-aware version. The new form captures the model used, the prompts, the human iteration, and the moment of inventive insight for each claim. Patent attorneys review that record before claims are drafted, flagging any claim limitation that lacks a documented human contribution. The USPTO’s 2024 guidance put a premium on this implementation paper trail across every AI-touched filing. Companies that treat it as busywork are the ones that will see their patents invalidated during litigation discovery. A good AI invention tracker also doubles as a prior-art log for subsequent filings.

In practice, workflow matters as much as policy for AI-assisted patent programs. The best teams run inventor interviews within two weeks of a disclosure to capture fresh memory. The engineer still remembers which prompt variations led to the breakthrough and the sequence of decisions. They also keep the AI tool’s version pinned in the record, because a model refresh can change the inventorship analysis. Some pharmaceutical and semiconductor firms now run mock post-grant proceedings on every AI-assisted claim before filing. The AI governance trends and regulations field is tracking this as a new best practice.

Beyond the mechanics, many organizations are restructuring their inventor incentive programs around the implementation effort. Human engineers and AI tool stewards share credit for successful filings, which keeps the human contribution visible and documented. This is partly cultural and partly defensive in intent across most research-heavy firms. If the people who ran the model never get recognized, they stop bothering to document their involvement at all. The record of significant contribution then evaporates and leaves a weak paper trail. A small bonus for the engineer who prompted the model keeps everyone in the loop and in the file.

The EU AI Act Training Data Disclosure Rules

Turning to Europe, Article 53(1)(d) of the EU AI Act requires providers of general-purpose AI models to publish training-data summaries. The European Commission published the template on July 24, 2025, and compliance obligations are now in force. The summary is not the raw dataset, but it is detailed enough to identify categories of works, sources, and licensing status. The published template exposes US rights holders to new discovery opportunities they could not otherwise access. The Bloomberg Law analysis of the training disclosure exposure covers the cross-border risk implications in detail.

In practice, any lab serving European users must publish information that was previously treated as trade secret. Rights holders then use that disclosure to probe for infringement under their national laws across jurisdictions. Parallel suits in both the EU and the US are becoming routine for well-known models. A model trained on scraped photography websites, for example, becomes a target for the photographers whose work appears there. Compliance teams are now building a disclosure workflow that aligns public summaries, internal licensing records, and defensive documentation strategy. An inconsistency between those three sources is itself evidentiary and valuable to plaintiffs looking for weak spots.

The US Copyright Office Position on Generative Works

Turning to the registration side, the US Copyright Office issued a three-part report between 2023 and 2025. Part 1 handled digital replicas and voice clones, including the first federal position on deepfake output. Part 2 handled copyrightability of AI-assisted works and confirmed the human-authorship requirement for registration. Part 3 addressed training and licensing, taking the view that training on expressive works implicates copyright protections. The Office prefers voluntary licensing markets as the primary remedy for creators affected by model training. The agency has refused registration for purely machine-generated images while granting registration for human-selected, human-arranged compositions that include AI elements.

Shifting to real cases, the who owns AI art analysis walks through the practical consequences for working artists. A visual artist can register a composition where a diffusion model generated some elements and the artist made significant creative choices. The registration covers the human expression, not the raw AI output or the model’s unaided contribution. The documentation burden is on the artist to show which parts of the work carry human authorship clearly. That documentation is what the Office examines on review and uses to decide registration.

Beyond the technical rulings, the Office is facing political pressure to clarify what counts as sufficient human contribution. Working artists want a clear registration pathway, publishers want standardization, and software vendors want predictability for product features. The Office has resisted bright-line rules, preferring a case-by-case analysis that mirrors other difficult authorship questions. That approach is defensible on principle, but it leaves every registrant guessing about where the line sits. Many artists simply register the human-edited version of the work without disclosing the AI involvement at all. That practice creates a different and larger problem down the line in enforcement proceedings.

Licensing Markets Emerging Between Publishers and AI Labs

Moving on to commercial deals, licensing is now the preferred path for most frontier labs that previously relied on open scraping. OpenAI, Google, and Anthropic have signed multi-year deals with major news publishers, image agencies, and stock-video platforms. The deals pay a mix of upfront fees, per-query fees, and revenue shares across the term. The Canadian news lawsuit against OpenAI is one of the pressure points pushing labs toward negotiated agreements. Scraped data will become a smaller share of production training sets over the next three years. The commercial logic is now driving the compliance posture more directly than any ruling could.

Looking at the economics, these deals are still being discovered through negotiation and precedent. A top-tier news publisher with strong archival content can command deals in the tens of millions of dollars per year. A boutique academic publisher may receive seven figures in annual payments for its content. A single-creator photography blog receives nothing, because transaction costs make individual licensing uneconomic for aggregators. Collective licensing bodies are forming to represent the long tail, modeled loosely on music industry performance-rights organizations. These bodies will set rates, aggregate payouts, and reduce transaction friction for both sides over time.

For teams buying AI products, enterprises need to understand what licensing actually covers in practice. Many AI vendors offer customer indemnity for copyright claims, but those indemnities often exclude claims arising from training data. That exclusion is where the biggest commercial risk sits for enterprise AI deployments at scale. A careful contract will require the vendor to disclose its training-data sources and to indemnify against training-data claims. Without that structure, the enterprise is left holding the bag on a dispute whose root cause lives inside the vendor’s model.

Beyond the primary deals, secondary markets are starting to form around curated datasets for commercial use. Datasets with cleaned, licensed, high-quality content now trade at prices that would have seemed absurd two years ago. A curated set of licensed news articles can sell for the low-millions range per million tokens today. Specialized datasets for medicine, law, and finance command even higher premiums in private marketplace deals. Pretraining is no longer purely a compute cost under the implementation economics facing frontier labs today. It is a compute cost plus a data cost, and the data cost is rising faster than compute.

Ethics and the Economic Fairness Question for Creators

Stepping back to the broader argument, the ethical case against the first-wave training approach is simple and increasingly uncontroversial across the industry. Models were trained on vast corpora of unlicensed human creative work, and the resulting products now compete with the creators whose work was ingested. The economic consequence has been documented across journalism, illustration, voice acting, translation, and music over several years of market data. The UK arts icons demanding AI copyright safeguards campaign made the political argument visible across major media. The empty book protest turned it into a public symbol of what unlicensed training does to writers over time. Creator organizations have moved from petitions to lawsuits during the past 18 months.

From there, fairness in this context has three components that practitioners debate separately in licensing talks. Consent is whether creators were asked before their work was ingested into a training corpus. Compensation is whether they were paid for the use of their material by AI developers. Credit is whether their contribution is visible in the output of the model at all. A policy that solves one does not necessarily solve the others, and labs often address only one. A licensing deal that compensates publishers but strips creator credit from aggregated training data solves only two-thirds of the problem. The next generation of licensing frameworks will need to address all three ethics dimensions together.

Risks for Enterprises Deploying Generative AI Without Clearance

Turning to the buyer side, enterprises are now the biggest customers of generative AI across all commercial sectors. They inherit the training-data risk whenever they deploy a model whose provenance they have not vetted carefully. A marketing team that generates ad imagery with an unvetted model may distribute a work that infringes a protected illustration. The costs are not evenly distributed across different procurement patterns in the market. An enterprise that bought a licensed vendor product with a strong indemnity carries less exposure than open-weights users. The implementation choice itself shifts the risk profile of the AI deployment in significant ways.

Shifting to inputs, operational exposures also come from the data employees paste into tools without checking. Confidential documents pasted into a public chatbot can waive privilege and leak trade secrets in one keystroke of error. Even enterprise tiers can have logging and retention policies that put sensitive content into third-party storage systems. The practical fix is a data-classification policy that explicitly lists which categories of data may be entered where. Enforce it through browser proxies or SSO scoping, not just a training deck or awareness poster. The AI ethics and laws overview walks through the broader governance frame for enterprise programs.

Beyond the inputs problem, litigation exposure is the final layer that enterprises must plan for carefully. A plaintiff suing an AI vendor can and often does join customers who used the vendor’s outputs commercially. Enterprises reduce that risk by maintaining audit logs of prompts and outputs across their deployments. Choose vendors with sufficient public disclosures about their training data and licensing arrangements before committing. Carry indemnity-backed cyber and tech errors and omissions coverage to backstop residual exposure. Run periodic model-replacement drills so the organization is not locked into a single provider during a legal shock.

Implementation Playbook for an In-House IP Team

For teams starting the implementation work, a workable 90-day plan begins with an inventory, not with a policy document. The team maps every AI tool in use, every model embedded in a product, and every vendor contract with AI terms. Every piece of training or fine-tuning data touched by employees also lands in the same inventory. That inventory alone usually surfaces more exposure than the team expected, including shadow deployments that no one documented. A policy written before the inventory is complete tends to miss the actual risks that live in the organization. The inventory becomes the backbone of every downstream decision in the program.

From there, the team layers in four control domains for practical implementation work. Patent documentation uses the AI-aware invention disclosure form and inventorship interviews within two weeks of filing. Copyright hygiene adds vendor diligence, output filters, and employee training on registration limitations and the human-authorship rule. Trademark enforcement runs periodic model probes against the company’s protected marks across popular generative products. The notice-and-takedown workflow for deepfake endorsements becomes a routine operational item for the team to manage. Trade-secret protection covers vendor flow-downs on no-training-on-inputs clauses, employee NDAs, and periodic red-team probes against the model.

Looking ahead, the team should build a quarterly ruling digest and vendor update delivered in a single page of text. New rulings or EU AI Act clarifications materially change what counts as acceptable practice from quarter to quarter. A once-a-year refresh is far too slow for the pace of change in 2026 across these regimes. A monthly digest is too heavy for most executive audiences and consumes team bandwidth unnecessarily. Quarterly matches the pace at which guidance and rulings are landing, and it fits naturally into executive planning cycles. The livestream copyright crises show how fast a situation can escalate when nobody is tracking the cadence closely.

Comparing Jurisdictions: US, EU, UK, China, and Japan

Choosing among jurisdictions, each major one has staked out a position that reflects its policy priorities. The United States has leaned on existing copyright and patent statutes with new agency guidance across both the USPTO and the Copyright Office. The country maintains a strong fair-use tradition that is now being tested in several active appellate cases. The European Union has gone the furthest on disclosure and compliance obligations through the AI Act, while leaving substantive IP rules to member-state law. The United Kingdom has oscillated between pro-AI exceptions and creator-protective measures through consultation and consultation revisions. The UK copyright regime prioritizing AI firms framing drew significant pushback from creator organizations.

Turning east, China has moved aggressively through regulations rather than through legislation at the National People’s Congress. The Interim Measures for Generative AI Services require content provenance for all AI products placed on the Chinese market. The Beijing Internet Court has extended copyright to AI output where a user showed creative effort through prompts and iteration. Japan has one of the most permissive statutory text and data mining exceptions, which has made it a favored jurisdiction for compute-heavy pretraining. South Korea, India, and Brazil are tracking the US and EU debates and are not yet drivers of global rulemaking. The practical takeaway is that a global AI product in 2026 must have a jurisdiction-aware legal strategy, not a single opinion.

The Future of AI and Intellectual Property Law

Looking ahead to the 2026 to 2030 horizon, three forces will shape the regulatory and commercial landscape. Statutory reform is likely in the US around training-data standards, and probably moving faster at the state level than the federal. The NO FAKES Act or a successor is likely to become federal law, creating a national right of publicity for voice and likeness. Statutory damages for AI-related infringement will almost certainly be codified to reduce the uncertainty that currently drives both settlements and refusals to settle. The direction of travel is clear even if the exact sequence remains uncertain among the active bills. Industry coalitions are already lobbying on the specific language that would appear in such a statute.

Shifting to market structure, licensing will mature further over the coming three to five years. Collective licensing bodies for creators will emerge in parallel to the established music industry models. Model-vendor contracts will converge on standardized clauses for training-data provenance, indemnity scope, and no-training-on-inputs promises. Open-weights releases will carry more legal risk, and labs will diverge further between permissive and conservative release strategies. Legal strategy becomes a competitive advantage for every company building AI products at commercial scale. The implementation work of today becomes the moat of tomorrow for the companies that invest early.

Beyond the near term, the deepest structural change is harder to see because it moves slowly over many years. The old copyright and patent framework was built for human creativity at human speed and human scale. Generative AI operates at machine speed and machine scale, and the law is adapting in piecemeal increments across jurisdictions. A cleaner redesign is eventually inevitable for the system to handle the volume and velocity of claims. Sui generis protection for substantial human investment in AI output, modeled on the EU database right, is one plausible path. A statutory licensing scheme for training data, modeled on the music industry mechanical license, is another option that policy teams are studying. Which path wins will depend less on legal theory and more on which political coalition pushes a bill through.

Key Insights on AI and Intellectual Property Law

• The Bartz settlement of about 1.5 billion dollars covered approximately 500,000 pirated books and priced the risk at about 3,000 dollars per infringed title across the corpus. The implied cost has flipped most frontier labs from scraping toward licensing deals with major publishers and image agencies worldwide.
• USPTO filings using AI assistance climbed sharply after the February 13, 2024 Inventorship Guidance, and documenting human contribution for each claim has become the dominant drafting activity. Law firms now train associates specifically on the significant-contribution analysis demanded by the new agency framework for every AI touched filing.
• The EU AI Act Article 53 training-data summary template went live on July 24, 2025, and the first published summaries have already produced cross-border discovery openings. US rights holders are now using those disclosures in parallel copyright suits across both American and European federal courts simultaneously today.
• Thomson Reuters against Ross Intelligence became the first US summary judgment to reject fair use for commercial-competing training in February 2025. The ruling reshaped the risk model for enterprise AI deployments relying on scraped datasets without clear licenses in any form today.
• The US Copyright Office refused to register in multiple 2023 to 2025 decisions on the record. Its three-part AI report confirmed that voluntary licensing is the Office's preferred remedy rather than statutory reform of the authorship test or sui generis rights for machine output.
• Licensing-based training-data agreements between labs and major publishers now reach the tens of millions of dollars in annual value. Curated licensed datasets for specialized domains such as medicine and law trade at a premium multiple of unlicensed scraped alternatives and generic web dumps on the market today.
• Deepfake and voice-clone exposure prompted Tennessee to pass the ELVIS Act in 2024, and the federal NO FAKES Act proposal gained renewed momentum after Sora launched. That consumer tool proved how easily deepfake replicas can now be produced at scale by any user with a browser online.
• Enterprises deploying generative AI without vendor diligence now face joint-defendant exposure in training-data suits across the US and EU. Vendor contracts without indemnity against training-data claims have become uninsurable at commercial scale in every major cyber and tech errors policy today.

The pattern across these rulings and settlements has become unmistakable to any serious observer of the field. Fair use is not an automatic shield, inventorship is not a formality, and disclosure obligations under the EU AI Act are not voluntary. Enterprises that treat AI legal risk as a line item in a vendor contract are the ones that will be surprised by litigation. Labs that treated training-data acquisition as a free resource are now paying nine and ten-figure settlements for the earlier mistake. The next phase of the market will reward teams that built documentation, licensing, and provenance workflows before the rulings forced them to adopt controls. The ethics argument and the economic argument have converged on a single conclusion that is hard to argue against.

Dimension	Traditional IP (pre-2023)	AI-era IP (2026)
Authorship	Human author required and assumed	Human authorship required but contested, machine-only output is public domain in the US
Inventorship	Named human inventors on each claim	Named human inventors with documented significant contribution for AI-assisted claims
Training Data	Not a legal category	Core risk vector, with disclosure obligations in the EU and settlements in the US
Trademark Scope	Logos, slogans, trade dress	Expanded to voices, likenesses, synthetic brands, and deepfake personas
Trade Secrets	Documents, formulas, source code	Model weights, fine-tuning datasets, prompt libraries, retrieval indexes
Enforcement Speed	Years from filing to judgment	Accelerating via summary judgment on training-data questions and preliminary injunctions on deepfakes
Jurisdiction Risk	Country-by-country, discoverable in advance	EU disclosure feeds US discovery, cross-border coordinated suits increasing
Licensing Markets	Mature in music, publishing, software	Emerging for training data, collective licensing bodies forming for creators

Real-World Companies Rebuilding IP Workflows Around AI

The three companies profiled below represent the clearest templates for how enterprises are rebuilding IP workflows around generative AI today. Each illustrates a distinct lever that practitioners can apply to their own operations, and each one has produced measurable results over the past two years.

Adobe's Firefly and Commercially Safe Training

Adobe launched Firefly in 2023 as a generative image product trained only on licensed Adobe Stock images and public-domain content. The outcome was a measurable 20 billion generations by late 2024, with sharp adoption increase in regulated industries. Adobe offered a commercial indemnity to enterprise customers that specifically covered training-data claims. Firefly adoption climbed quickly because risk officers could sign off on the commercially safe implementation within weeks. The limitation is that contributors to Adobe Stock did not initially contemplate AI training, which triggered contributor-compensation disputes handled through a bonus program paying millions annually. The still-ongoing contributor critique remains a trade-off of the licensed-only approach that other platforms will have to resolve. The lesson is that provenance-first training can be a feature when enterprise buyers assign real weight to indemnity scope.

Getty Images and the Dual-Track Strategy of Suing and Partnering

Getty Images filed suit against Stability AI in both US and UK courts in 2023 over 12 million allegedly ingested images. The outcome of the dual-track approach produced results on both fronts, with the Stability case producing early rulings on database rights. The licensed Getty generative product also generated recurring revenue in the single-digit millions per quarter by 2025, representing a meaningful lift. The Hollywood versus AI copyright showdown analysis places Getty's dual-track approach as the template that creative industries have started to follow. The limitation is that enforcement litigation is slow and expensive, with the trial lasting weeks and still-pending appeals taking years to resolve. The trade-off between enforcement reach and product breadth is the core critique of the dual-track strategy among skeptics. The lesson is that litigation and licensing can be complementary levers rather than mutually exclusive alternatives for creative content holders.

IBM's Patent Portfolio and AI-Assisted Invention Tracking

IBM has been the top US patent filer for 29 consecutive years and received 3,658 US patents in 2024 while deliberately winding down headline-count filings. The company adopted the AI-assisted invention tracking model required by the USPTO 2024 guidance within the first 60 days after publication, a sharp reduction in time to implementation. The outcome was a disciplined pipeline where IBM's invention disclosure form captures the AI tool, prompts, iteration, and inventive insight for each claim. IBM's patent counsel has described the shift as adding a tier of documentation that was not previously required, with the trade-off that filings now take weeks of extra lead time. The limitation is that IBM is a sophisticated filer with internal counsel capacity, so the model does not scale cleanly to smaller companies without comparable infrastructure. The AI governance trends and regulations tracker cites the IBM pattern as the benchmark for enterprise AI invention programs. The lesson is that documentation discipline, not raw volume, now wins the AI patent race in the current regulatory environment.

Case Studies in AI Copyright and Patent Litigation

The three case studies below show how major AI copyright and patent disputes have resolved over the past 18 months of litigation. Each one has rewritten the operational playbook for a different kind of AI product, and the implications continue to ripple through enterprise buying decisions.

Case Study: Bartz against Anthropic and the Training Settlement

The Bartz case began in August 2024 when three book authors filed a class action in the Northern District of California. The problem they identified was that Anthropic had trained Claude on roughly 500,000 pirated books sourced from Library Genesis and similar shadow archives. The authors could not consent to the training because Anthropic never asked for a license from the rights holders. The complaint identified the dataset through source metadata that Anthropic had retained inside the training pipeline. Judge William Alsup's early rulings signaled skepticism about Anthropic's fair-use defense for the pirated sources specifically. The court granted class certification in June 2025 and set a trial date for later in the year. The plaintiffs faced the challenge of proving the pirated nature of the ingested data at the scale required for class damages.

Anthropic deployed a settlement as its solution and introduced about 1.5 billion dollars in payouts in September 2025, which worked out to roughly 3,000 dollars per infringed book. The impact in dollar terms was a nine-figure revenue hit that produced a measurable reduction in cash reserves and a sharp industry-wide increase in licensing deals. The entire frontier-lab market shifted toward provenance-first offerings within weeks after the number went public, and competitors adopted similar approaches to training data. The limitation is that the settlement did not resolve whether training on lawfully acquired books is fair use, so the central doctrinal question remains contested and still open on appeal. The AI Business lawsuits tracker places Bartz as the single most consequential AI copyright event of 2025 among dozens of filings. The takeaway for enterprise buyers is that training-data provenance now has a market-tested price tag running into the nine figures.

Case Study: Thomson Reuters against Ross Intelligence and the First Fair-Use Rejection

Thomson Reuters sued Ross Intelligence in 2020 after Ross trained a legal-research AI on Westlaw's headnotes without a license. The problem for Thomson Reuters was that Ross had built a directly competing product using material that Thomson Reuters licenses rather than publishes freely. The plaintiff could not simply ignore the market harm, so it pursued the case through five years of discovery and motion practice. Thomson Reuters faced the challenge of proving that the Ross training was not transformative, with weeks of trial preparation on the record. Judge Stephanos Bibas granted Thomson Reuters summary judgment on direct copyright infringement in February 2025 in the District of Delaware. The court rejected Ross's fair-use defense specifically because the Ross product competed commercially with Westlaw's existing paid service. Ross had already shut down in 2021 citing legal costs, so the practical remedy was limited to damages against a defunct company.

Thomson Reuters introduced an aggressive litigation strategy as its solution and deployed experienced trial counsel over the five-year case. The impact of the ruling was a sharp reduction in defendant confidence across every pending AI copyright case, with a measurable increase in settlement activity afterward. Settlement activity rose by hundreds of millions of dollars across the AI copyright docket in the months after the ruling landed. The ruling gave plaintiffs a clear template for arguing that AI training on a commercial dataset to build a competing product is not fair use. The limitation is that the ruling is narrow on its facts, so it does not automatically cover every training scenario in every court. Judge Bibas was careful to distinguish non-competing uses, and the trade-off between narrow holdings and broad market impact remains contested across the appellate bar. The AI copyright lawsuits in the US explainer places Thomson Reuters as the first substantive doctrinal fair-use ruling.

Case Study: Getty Images against Stability AI and the UK Trial on Database Rights

Getty Images sued Stability AI in both the US District of Delaware and the UK High Court in early 2023 over training-data use. The problem was that Stable Diffusion was trained on approximately 12 million Getty images, with outputs sometimes carrying visible Getty watermarks as evidence. Getty argued that the watermarks proved both infringement and false association for consumer-facing uses of the model. Getty faced the challenge of coordinating two parallel proceedings in different common-law systems with different evidentiary rules and discovery regimes. The UK trial opened in June 2025 and became the first full trial on AI training-data liability in a common-law jurisdiction. The weeks of expert testimony produced detailed technical evidence about how Stable Diffusion was trained and how watermarked images appeared. Getty asserted copyright, database right, and trademark claims across the case while Stability defended on text and data mining exceptions.

Getty introduced a dual-track solution and deployed its own generative tool trained on licensed Getty content, launched in 2023 alongside the ongoing litigation. The impact was an immediate measurable spend reduction toward unlicensed providers in enterprise segments across the UK and EU markets. The licensed Getty generative product recorded recurring revenue in the single-digit millions per quarter by 2025, with other image agencies filing parallel suits. The approach built a defensible market position before any trial court ruling landed on the merits. The limitation is that trial-level rulings are not binding across jurisdictions, so even a complete Getty victory will not automatically bind US courts. Stability can still appeal through the UK Court of Appeal and Supreme Court over years of further proceedings, which is the central trade-off of the UK-first litigation strategy. The Disney lawsuit challenging AI copyright boundaries is tracking similar questions for cinematic content and screen libraries.

Frequently Asked Questions on AI IP Law