Discover the biggest AI patent lawsuits, key legal battles, and revealing stats on AI patent disputes. Stay ahead with the latest AI legal trends.

AI patent litigation has crossed an important line.

For years, most of the loudest “AI lawsuits” were not patent cases at all. They were copyright fights about training data, trade-secret disputes, privacy claims, or fights over who owned code.

That is changing.

Patent lawsuits are now reaching directly into the AI stack:

ChatGPT search, memory, and retrieval.

LLM-generated legal documents.

Voice cloning and text-to-speech.

Machine-learning property analysis.

Alexa and Siri.

AI servers, GPUs, and DPUs.

That creates a much harder problem for AI companies.

You do not need to copy somebody’s source code to infringe a patent.

You can independently build your own system and still run into somebody else’s patent claims.

And the patent threatening a 2026 AI product may have been drafted years before anybody used words such as “RAG,” “AI agent,” or “LLM.”

PatentPC analyzed ten important U.S. AI-related patent enforcement campaigns, the patents asserted in them, their filing dates, major rulings, PTAB activity, ITC proceedings, and one major public settlement.

The result gives founders a much clearer picture of where AI patent risk is actually moving.

The Fast Answer: The Biggest AI Patent Cases to Watch

As of September 14, 2026, these are among the most useful cases for understanding the direction of AI patent litigation.

Case	AI technology at issue	Plaintiff patents in campaign	Latest important development
Many Worlds v. OpenAI	ChatGPT search, retrieval, memory, recommendation	5	Filed Aug. 31, 2026
BTF IP v. ElevenLabs	AI voice generation and TTS	1	Filed Sept. 1, 2026
AI.Law v. Eve Legal	LLM legal-document generation	1	§101 dismissal fight underway
Xockets v. Nvidia/Microsoft/Amazon	DPU and AI-compute infrastructure	7	ITC investigation active
Cerence v. Amazon	Alexa, speech and conversational AI	12	District cases stayed for ITC
Cerence v. Apple	Siri, voice commands and text input	6	Ongoing patent campaign
Aon v. Zesty.ai	Machine vision and property-risk AI	4	Survived §101 dismissal
Recentive v. Fox	ML scheduling and network maps	4	Federal Circuit invalidated claims; Supreme Court review denied
Cerence v. Samsung	Voice and conversational technology	9	Resolved through cross-license and $49.5M payment
Cedar Lane v. ElevenLabs	Speech/media technology	1	Dismissed with prejudice

These cases do not form a complete census of every lawsuit that could be described as “AI related.”

That is intentional.

There is no clean federal court category called AI patent litigation. Calling a patent “AI related” can become subjective very quickly.

So instead of pretending there is a magic global lawsuit count, PatentPC built a transparent sample of litigation where AI, machine learning, conversational computing, GenAI, AI infrastructure, or an AI-native product is central enough to make the case useful to AI companies.

That produces better knowledge than a giant unsupported number.

PatentPC Original Research: 50 Patents Across 10 AI Patent Campaigns

Our ten-case U.S. sample contains:

50 plaintiff-asserted patents.

That is before counting defendant counterclaims.

But the distribution matters more than the raw number.

Seven of the ten campaigns asserted multiple patents

That is:

70%

Only three of the ten relied on a single asserted patent.

That tells AI companies something important.

Serious technology enforcement often looks more like a portfolio problem than a one-patent problem.

A company with five related patents may be able to attack the same product from several angles.

One patent might cover data processing.

Another might cover the user interaction.

Another might cover the network architecture.

Another might cover the inference path.

Another might cover an implementation detail.

PatentPC discusses this wider approach in its guides to building a patent portfolio for startups and building a lean patent portfolio without wasting money.

Another Finding: 68% of the Patents Sit Inside Just Four Portfolio Campaigns

The four largest plaintiff portfolios in our sample are:

Cerence v. Amazon: 12 patents

Cerence v. Samsung: 9 patents

Xockets campaign: 7 patents

Cerence v. Apple: 6 patents

Total:

34 patents

34 ÷ 50 =

68%

So more than two-thirds of the plaintiff-asserted patents in our sample sit inside four concentrated enforcement campaigns.

This is one reason a startup should not ask only:

“Do we infringe Patent X?”

A better question is:

“What patent families surround this technical layer, who owns them, and what other continuation claims could reach us?”

That is what a real freedom-to-operate analysis should help answer.

PatentPC also has a practical guide to conducting an FTO analysis without turning product development into legal paralysis.

The Most Interesting Number: The 90/90 Problem

Here is the finding AI founders should pay closest attention to.

In PatentPC’s ten-campaign sample:

9 of 10 enforcement campaigns began after ChatGPT’s November 2022 public launch.

But:

9 of 10 campaigns assert at least one patent family whose technical roots predate ChatGPT.

That gives us a strange 90/90 pattern.

90% of the litigation is from the post-ChatGPT era.

90% reaches back into pre-ChatGPT patent rights.

That means AI patent litigation is not simply a coming battle over patents being filed today.

The battle has already inherited years of older IP.

This may be the most important FTO lesson in the entire article.

Searching only for “LLM patents,” “RAG patents,” “AI agent patents,” or today’s product vocabulary can miss the patents that eventually matter.

Older patents may use language such as:

recommendation engine;
inferential system;
semantic relationship;
speech synthesis;
neural network;
intelligent agent;
contextual search;
automated classifier;
data processing;
distributed computing;
natural-language interface;
predictive model.

The product changed.

The vocabulary changed.

The claim may still read on the product.

PatentPC’s guide to patent clearance searches without slowing innovation becomes particularly important in AI for exactly this reason.

Case #1: Many Worlds v. OpenAI May Be the Case Every RAG Startup Should Read

On August 31, 2026, Many Worlds 2T Innovations sued OpenAI in the Eastern District of Texas.

It asserted five patents.

The complaint targets features associated with ChatGPT’s search, retrieval, memory, customization, and content generation.

Read the Many Worlds v. OpenAI docket.

The case is extremely new. No court has ruled that OpenAI infringes anything.

But the structure of the accusation matters.

According to the complaint, Many Worlds traces its relevant patent family back to 2011.

Yet those older concepts are now being mapped onto features of a modern LLM product.

That is the lesson.

Patent risk does not care what generation of AI marketing language you use

Suppose you build a modern RAG product.

You may describe it internally using:

embeddings → vector search → retrieval → ranking → context injection → LLM generation

An older patent may describe a similar functional path using:

semantic relationships → user profile → content analysis → weighted relationships → content selection → generated output

Those phrases are not identical.

That does not automatically mean the patent covers your system either.

Patent infringement turns on the claims.

But keyword mismatch is not freedom to operate.

That is why good AI patent searching needs to work backward through technical ancestry.

PatentPC explains how patent analytics can be used for this kind of competitive work in Using AI Patent Analytics for Competitive Advantage.

PatentPC Rule #1: Search the Ancestors of Your Architecture

If your AI product contains RAG, do not search only RAG.

Search:

information retrieval
contextual search
semantic retrieval
recommendation
document ranking
user profiling
adaptive content
query expansion
knowledge retrieval
similarity search
relationship scoring

Then search the relevant patent classifications. Do the same for agents.

Today’s AI agent may have technical ancestors called:

software agent
intelligent agent
automated task execution
workflow orchestration
rule-based execution
distributed process automation

The more valuable the feature, the further backward you should look.

PatentPC has a deeper guide to AI-powered patent searching in deep tech.

Case #2: BTF IP v. ElevenLabs Brings Patent Litigation Straight Into Generative Voice AI

One day after the OpenAI case was filed, another direct AI patent case appeared.

On September 1, 2026, BTF IP Holdings sued ElevenLabs in the Southern District of New York.

The asserted patent is U.S. Patent No. 11,521,593.

Its earliest listed priority date is September 18, 2019.

ElevenLabs was founded later.

The complaint maps the patent onto parts of ElevenLabs’ AI voice platform, including text-to-speech, voice selection, voice cloning, dubbing, and related functionality.

Read the BTF v. ElevenLabs docket and the BTF v. ElevenLabs claim analysis.

Again, there has been no finding that ElevenLabs infringed.

But look at the timing.

Patent priority

ElevenLabs founded

Lawsuit

That is exactly the kind of sequence that should change how founders think about FTO.

The question is not:

“Did anybody patent ElevenLabs?”

Nobody patents another company.

The question is:

“Which existing claims could cover technical acts our product performs?”

The Claim-Chart Lesson From ElevenLabs

The BTF complaint is useful because the asserted claim contains many concrete limitations.

The dispute may involve questions around things such as:

collecting online content;
selecting stored voices;
recording or cloning a subscriber’s voice;
classifying content;
converting text to speech;
language conversion;
voice marketplaces;
background sounds.

The complaint attempts to map those limitations onto ElevenLabs features.

Whether those mappings ultimately work is for the court.

But founders should learn the process.

Patent infringement is not:

“Their product feels like our patent.”

It is closer to:

Claim element 1 → where does the accused system allegedly do this?

Claim element 2 → where?

Claim element 3 → where?

And so on.

If an asserted claim requires A+B+C+D+E, proving A+B+C+D is generally not enough for literal infringement.

That is why FTO should eventually become claim-level work.

PatentPC goes deeper into infringement strategy in Handling AI Patent Infringement Cases.

Case #3: AI.Law v. Eve Legal Shows How Fast a New AI Patent Can Turn Into Litigation

The older-patent story is only half of the market.

AI.Law v. Eve Legal shows the opposite.

AI.Law’s U.S. Patent No. 12,461,932 was filed in October 2024 and covers technology for transforming unstructured material with AI into structured, long-form documents.

The patent issued in November 2025.

Then AI.Law sued Butler Labs, which operates Eve Legal, on June 17, 2026.

That is roughly seven months after issuance.

So while some 2026 AI lawsuits rely on decade-old patent families, others involve patents drafted squarely for the modern LLM era.

Read the AI.Law patent record, Reuters coverage of the filing, and the AI.Law v. Eve complaint analysis.

Eve’s Response Makes This Case Even More Important

Eve has moved to attack the patent, arguing in part that it covers little more than using ordinary AI tools to draft legal documents.

That puts the case near the center of one of AI patent law’s hardest questions:

When is an AI workflow a patentable technical system, and when is it merely an abstract task performed using AI?

That question leads directly to Recentive.

Case #4: Recentive v. Fox Is Now One of the Most Important AI Patent Eligibility Cases in America

Recentive owned four patents directed to machine learning used for generating broadcast network maps and live-event schedules.

Fox attacked them under 35 U.S.C. §101.

The district court dismissed the case.

The Federal Circuit affirmed in April 2025.

The Federal Circuit’s core rule was powerful:

Applying generic machine-learning technology to a new data environment, without claiming a technological improvement, can remain an abstract idea.

Read the Federal Circuit’s Recentive decision.

Recentive asked the U.S. Supreme Court to review the decision.

On December 8, 2025, the Supreme Court denied certiorari.

Read the Supreme Court docket.

That leaves the Federal Circuit ruling standing.

What Recentive Means in Plain English

Imagine a patent claim effectively says:

Take machine learning and use it to optimize restaurant staffing.

Or:

Use machine learning to select insurance leads.

Or:

Use machine learning to decide television schedules.

The fact that machine learning is useful does not automatically make the claim patent eligible.

The patent needs more.

For AI patent drafting, the better question is:

What did you improve technically?

Maybe:

the training method;
memory architecture;
data representation;
inference pipeline;
processor utilization;
retrieval mechanism;
model routing;
error correction;
computer security;
distributed execution;
latency;
bandwidth use;
storage;
continual learning.

PatentPC explains this issue in its guide to technical improvements in software patentability and its guide to the Alice test for software patents.

The USPTO’s own subject-matter eligibility guidance is also essential reading.

Case #5: Aon v. Zesty.ai Shows Recentive Does Not Mean “AI Patents Are Dead”

This is where the law gets more interesting.

Aon sued Zesty.ai in February 2025 over four patents involving machine-learning analysis of aerial images of properties.

The systems can identify things such as roof characteristics and conditions and use them in property-risk analysis.

Zesty attacked the patents under §101.

That sounds a lot like Recentive.

But the result was different.

The Delaware court denied the motion to dismiss.

Why?

The court found the claims sufficiently tied to a particular technical arrangement involving two independently trained classifiers—one identifying property characteristics and another evaluating condition.

Read the District of Delaware’s Aon v. Zesty opinion.

Recentive vs. Aon Is the Comparison Every AI Patent Drafter Should Study

Put them side by side.

Recentive	Aon
Apply ML to new environment	Specific classifier architecture
Generic ML techniques	Particular arrangement of ML components
Claims failed under §101	Claims survived dismissal
Federal Circuit affirmed	District litigation could continue

Do not overread that table.

Aon surviving a motion to dismiss does not mean the patents are ultimately valid or infringed.

Recentive does not mean every patent using known ML techniques is invalid.

But the contrast is incredibly useful.

AI patent strength often lives in the architecture—not in the sentence “we use AI.”

This is why PatentPC’s approach to strong software patents focuses on the technical mechanism.

Zesty Also Attacked the Patents at the PTAB

Zesty did not stop at §101.

It filed inter partes review petitions against all four asserted patents in July 2025.

The petitions were denied institution in November 2025.

That is another lesson.

A modern patent dispute may happen on several fronts:

District court → infringement and validity.

PTAB → patent validity challenges.

Federal Circuit → appeals.

ITC → import exclusion.

A startup facing litigation needs to think in forums, not just lawsuits.

PatentPC covers this more broadly in How to Handle Patent Litigation in the Tech Industry.

Case #6: Xockets v. Nvidia, Microsoft, and Amazon Shows AI Patent Risk Goes All the Way Down to the Hardware Stack

One of the biggest mistakes in AI IP strategy is thinking AI patents mean only models.

The AI stack starts much lower.

Xockets accused Nvidia and Microsoft of infringing patents covering data-processing-unit technology used to offload computing work and improve server and network performance. It also brought separate actions against Amazon and AWS around Nitro infrastructure.

One asserted family reaches back to at least 2013.

That is nearly a decade before the GenAI boom.

The campaign involves seven distinct patents across the district-court cases.

Then Xockets opened another front.

In May 2026, it filed an ITC complaint targeting GPU computing systems and DPU technologies.

The U.S. International Trade Commission instituted Investigation 337-TA-1505 in June 2026.

The ITC investigation names Nvidia, Microsoft, Amazon, AWS, and Annapurna Labs among the respondents.

Read the Xockets ITC notice.

Why the ITC Makes Xockets Especially Important

A normal patent damages case asks for money and perhaps an injunction.

An ITC Section 337 action can seek something commercially brutal:

an exclusion order preventing infringing imported products from entering the United States.

That can shift leverage dramatically for hardware-heavy technologies.

The Texas litigation was stayed while the parallel ITC dispute proceeds.

This is why AI infrastructure companies need an FTO map that extends beyond:

model → training → inference

into:

accelerators → interconnect → memory → networking → offload → orchestration → server architecture

PatentPC discusses this broader strategic view in Patent Litigation Trends Affecting Deep Tech Companies.

Case #7: Cerence v. Amazon Shows Conversational AI Is Becoming a Portfolio War

Cerence may be the best current example of an operating AI company using patents as a strategic weapon.

In May 2026, Cerence filed two patent cases against Amazon, Amazon Services, and AWS in the Eastern District of Texas.

One asserts five patents.

The other asserts seven.

Total:

12 patents.

The technologies span areas including:

conversational AI;
speech recognition;
voice command processing;
noise handling;
smart-device interaction.

Cerence also filed an ITC complaint.

The Commission instituted Investigation 337-TA-1504 in June 2026.

The products at issue include smart speakers, smart displays, televisions, tablets, and streaming devices.

Amazon’s Alexa ecosystem therefore sits at the center of both district-court litigation and an import-exclusion proceeding.

The district cases were later stayed pending the ITC process.

Read the first Cerence v. Amazon docket, the second Cerence v. Amazon docket, and the Cerence-Amazon ITC investigation announcement.

Cerence Shows What a Mature AI Patent Portfolio Can Look Like

Cerence’s enforcement activity does not stop with Amazon.

It has also pursued patent claims against:

Apple;
Samsung;
Sony;
TCL.

Its 2026 SEC filings identify multiple active patent-enforcement campaigns.

This is what founders should understand about patent portfolios.

The value of a patent portfolio is not necessarily:

“We have 100 patents.”

The better question is:

“Can our patents form useful claim coverage around commercially important products?”

PatentPC explores the business side of this in Leveraging AI Patents for Business Strategy and The Impact of AI on Patent Portfolio Management.

Case #8: Cerence v. Apple Shows That “Old” Interface Patents Can Still Matter to Modern AI Products

Cerence sued Apple in September 2025.

The case involves six patents directed to technologies including text input and voice-command monitoring used in Apple products.

Some of the technical roots go back more than two decades.

For example, patent-family information surrounding one asserted input patent traces priority through early-2000s and even late-1990s work.

Think about what that means.

Apple’s modern device stack now includes:

Siri;
Apple Intelligence;
speech input;
predictive interfaces;
multimodal features.

But the patent exposure may come from inventions created long before anybody spoke about “generative AI.”

That brings us back to the 90/90 finding.

New AI does not create a clean patent slate.

Read the Cerence v. Apple docket.

Case #9: Cerence v. Samsung Gives Us Something Rare—a Public Dollar Figure

Most patent settlements are confidential.

That makes reliable “AI patent settlement statistics” dangerous to invent.

Cerence gives us one unusually useful public data point.

Cerence filed its first patent infringement case against Samsung in October 2023, asserting five patents.

It filed a second in March 2024 asserting four more.

Samsung brought patent counterclaims of its own.

Then, in October 2025, the parties resolved the disputes through a cross-license.

Cerence disclosed in an SEC filing that Samsung agreed to make a $49.5 million lump-sum payment.

Read the Cerence SEC disclosure.

That is real money.

Not an estimated litigation statistic.

Not an anonymous survey.

A publicly disclosed transaction.

What the $49.5 Million Cerence-Samsung Deal Does—and Does Not—Tell Us

It does not mean:

AI patents are worth $49.5 million.

Patent value depends on:

claim scope;
validity;
market size;
infringement exposure;
remaining life;
licensing history;
portfolio depth;
design-around cost;
litigation leverage.

But the settlement proves something simpler:

AI-adjacent patent portfolios can become material commercial assets, not decorative certificates on an investor slide.

That matters in licensing.

It matters in M&A.

It matters in fundraising.

It matters when a competitor has patents of its own.

PatentPC explains how to connect patents to company value in Leveraging AI Patents for Business Strategy and How to Build a Patent Portfolio for Startups.

Case #10: Cedar Lane v. ElevenLabs Is a Reminder That Filing a Lawsuit Is Not Winning One

ElevenLabs has already faced an earlier patent case.

Cedar Lane Technologies sued ElevenLabs in September 2025 over U.S. Patent No. 8,438,485.

The case did not reach some dramatic AI precedent.

It was voluntarily dismissed with prejudice in January 2026.

Read the Cedar Lane v. ElevenLabs dismissal.

That is important too.

Patent complaints are allegations.

Some become massive trials.

Some settle.

Some die on eligibility.

Some fail on invalidity.

Some disappear.

A headline saying:

“AI company sued for patent infringement”

does not tell you whether the patent owner has a strong case.

The claims and evidence do.

The Global Case to Watch: Xiao-I v. Apple

The U.S. is not the only market where AI patent rights are being tested.

Shanghai Xiao-I has spent years fighting Apple over a Chinese patent covering a “chat robot system” and Apple’s Siri technology.

The procedural history gives one of the cleanest examples of the difference between patent validity and patent infringement.

In March 2026, China’s Supreme People’s Court rejected Apple’s effort to invalidate the relevant Xiao-I patents, according to Xiao-I’s SEC disclosure.

That was a major validity win for Xiao-I.

But it did not automatically mean Apple infringed.

On June 10, 2026, the Shanghai High People’s Court dismissed Xiao-I’s patent-infringement claims at first instance.

Xiao-I said it intended to appeal.

Read Xiao-I’s SEC disclosure on patent validity and SEC disclosure on the June 2026 infringement rulings.

That difference matters enormously.

A patent can be valid but not infringed.

An accused product can fall within a claim, while the claim itself is invalid.

Winning one question does not win the other.

Founders often collapse all of these questions into:

“Is this patent good?”

Patent litigation is more complicated.

The Five Questions Every AI Patent Lawsuit Eventually Forces

Most disputes become some combination of five questions.

1. What do the claims actually mean?

This is claim construction.

A single word can change the case.

Does “store” mean a literal marketplace?

Does “classifier” require one architecture or many?

Does “automatically” exclude manual steps?

Does “real time” impose a measurable timing requirement?

Words matter.

2. Does the accused product perform every required element?

This is infringement.

A competitor can build something that produces the same outcome using a different mechanism.

Patents protect claims.

Not vibes.

3. Was the claimed invention already known?

This is novelty.

Prior patents, publications, products, papers, standards, code, manuals, and other public disclosures may matter.

The USPTO’s Patent Public Search is one starting point.

4. Would the invention have been obvious?

An invention can be new but still obvious.

That often makes AI cases technically difficult because machine learning systems combine many known components.

The real question may become:

Was this particular architecture or combination an obvious way to solve the problem at the time?

5. Is the claim even patent eligible?

This is where Recentive, Alice, and modern AI eligibility fights enter.

A claim can be new and still fail §101.

PatentPC’s AI patent eligibility guide explores this in more detail.

Original Research Finding: “AI Patent Risk” Is Really Stack Risk

Look again at PatentPC’s ten-case sample.

The disputes do not cluster around one magic thing called “AI.”

They touch different layers.

Compute layer

Xockets.

Speech and interface layer

Cerence.

BTF.

Cedar Lane.

Machine-learning architecture

Aon.

Recentive.

GenAI application layer

AI.Law.

Retrieval, search, personalization, and memory

Many Worlds.

That means a company’s patent review should map the whole product stack.

PatentPC’s Seven-Layer AI FTO Map

Before an important AI launch, map patent risk across seven layers.

Layer	What to search
1. Compute	GPUs, accelerators, DPU, memory, networking, offload
2. Model	Architecture, training, fine-tuning, inference
3. Data	Processing, labeling, compression, transformation
4. Retrieval & memory	Search, ranking, embeddings, recommendation, context
5. Orchestration	Routing, agents, tools, workflows, fallback
6. Interface	Speech, vision, gesture, text input, multimodal interaction
7. Application	The vertical workflow your customer actually buys

The mistake is doing an FTO search only at Layer 7.

Suppose you sell an AI lawyer.

The application looks legal.

Your patent risk could still sit in:

document parsing
LLM continuation logic
retrieval
voice transcription
OCR
workflow orchestration
database synchronization
model routing

That is why PatentPC’s FTO guide for technology CEOs treats patent clearance as a product architecture problem.

Why AI Founders Should Stop Searching Patents by Product Category

Searching:

“patents for AI legal software”

is not enough.

Search the mechanism.

For an AI legal product, break the product into:

PDF ingestion

→ layout extraction

→ OCR

→ entity recognition

→ document classification

→ chunking

→ semantic indexing

→ retrieval

→ ranking

→ prompt assembly

→ model inference

→ long-document continuation

→ citation checking

→ formatted output

Now search those technical operations.

You may find relevant patents belonging to companies that have never sold legal software in their lives.

That is exactly the point.

PatentPC Original Research Finding: The Biggest Litigation Threat May Be a Patent Written for Another Generation of Technology

The Many Worlds/OpenAI case makes that visible.

So does Cerence/Apple.

So does Xockets.

This creates what we might call the technology ancestry problem.

Modern products often combine old technical ideas at a new level of scale.

An old patent might have been drafted around:

web content
Modern product:
LLM context
Old patent:
personalized recommendation
Modern product:
AI memory
Old patent:
speech synthesis
Modern product:
voice cloning
Old patent:
server computation offload
Modern product:
AI cluster acceleration

A serious patent search therefore needs to ask:

What older technical problem is our new product secretly solving?

That question can uncover prior art.

It can uncover blocking patents.

It can also uncover white space for your own patent strategy.

The Other Side of Litigation: Build Patents That Are Harder to Walk Around

These cases are not just warnings for defendants.

They are drafting lessons for patent owners.

A good AI patent portfolio should try to cover the technical idea at several useful levels.

For example:

System claim

Protect the architecture.

Method claim

Protect the sequence of technical operations.

Component claim

Protect the key subsystem.

Alternative embodiments

Protect different implementations.

Continuation strategy

Keep claim scope adaptable as the market develops.

The goal is not maximum vagueness.

Vague claims can create eligibility, written-description, enablement, and validity problems.

The goal is:

Broad enough to matter. Specific enough to survive.

PatentPC covers this balance in How to Get Strong Software Patents.

The Recentive Trap: Do Not Patent the Business Result

A bad AI invention story sounds like:

“We use AI to choose better insurance policies.”

Or:

“We use machine learning to schedule broadcasts.”

Or:

“We use an LLM to write legal documents.”

Those descriptions tell you the goal.

They barely tell you the technology.

A stronger disclosure asks:

What does the old architecture do?

Why does it fail?

What changed inside the machine?

What exact data enters each component?

What transformations occur?

What state is stored?

What decisions change the execution path?

What measurable technical effect results?

That is the raw material a strong AI patent attorney needs.

PatentPC’s original research on what actually makes an AI company defensible reaches the same conclusion from a different direction: simply having AI is becoming less of a moat. The proprietary systems around it matter more.

If You Receive a Patent Threat, Do Not Start by Asking “How Much Should We Pay?”

Start by building a claim map.

For every asserted independent claim:

Column 1

Claim limitation.

Column 2

Plaintiff’s alleged mapping.

Column 3

What your system actually does.

Column 4

Potential non-infringement gap.

Column 5

Potential prior art.

Column 6

Design-around option.

Column 7

Business cost of changing the system.

That table can change the negotiation.

A patent may look frightening until one required limitation fails.

Or it may look weak until the engineering team realizes the accused feature maps almost perfectly.

Do the technical work first.

PatentPC’s guide to preparing for patent litigation covers the wider process.

Do Not Forget Design-Around Economics

Patent analysis is not only legal.

It is economic.

Imagine a claim looks dangerous.

You have four options:

Fight.
License.
Settle.
Design around.

A design-around becomes attractive when:

engineering cost + migration cost + performance penalty

is much lower than:

expected litigation cost + expected damages + injunction risk + management distraction

That is why patent lawyers working with AI companies need to understand architecture.

A tiny claim change may require a massive engineering rewrite.

Or one simple pipeline change may remove a major infringement argument.

ITC Risk Makes Hardware and Smart-Device AI Different

Xockets and Cerence both show why AI companies with physical products need another layer of planning.

If your AI technology sits inside:

servers;
automobiles;
smart speakers;
televisions;
chips;
edge devices;
robots;
consumer electronics;

the ITC may matter.

An exclusion order can be more dangerous than a normal damages claim because it attacks the ability to bring accused products into the U.S.

In PatentPC’s ten-campaign sample, two campaigns—20%—used a parallel ITC track.

That percentage should not be generalized to all AI litigation.

Our sample deliberately focuses on major cases.

But it is enough to establish the strategic point:

For hardware-linked AI, forum risk can become product-supply risk.

Why “We Use OpenAI” Does Not Automatically Solve Your Patent Problem

Companies sometimes assume:

“We’re just using somebody else’s model. Patent risk is their problem.”

Not necessarily.

Suppose the patent claim is not about the model itself.

It may cover:

your retrieval architecture
your document workflow
your user interface
your voice layer
your orchestration logic
your combination of APIs
your edge processing

Whether your vendor’s indemnity helps depends on the contract and the actual allegation.

The more custom technology you put around the model, the less sensible it becomes to think of patent risk as purely the model vendor’s issue.

PatentPC’s guide to developing a robust IP strategy for AI innovations explains why IP strategy has to follow the product rather than the buzzword.

What AI Startups Should Do Before Their Next Major Launch

Use this checklist.

1. Draw the architecture

Not the pitch deck.

The actual system.

2. Mark what is proprietary

Which components did you invent?

Which came from vendors?

Which are open source?

3. Search each important layer

Do not search only the end product.

4. Search old terminology

Ask what this technology would have been called:

5 years ago.

10 years ago.

15 years ago.

The 90/90 finding makes this mandatory.

5. Identify dangerous claim families

One patent is not enough.

Look at:

parents;
continuations;
divisionals;
foreign family members;
pending applications.

6. Build claim charts for the highest-risk patents

Stop relying on titles and abstracts.

Claims control.

7. Look for design-around space

Patent analysis should inform engineering before launch.

Not after the complaint arrives.

8. Build your own defensive portfolio

Your own patents do not magically immunize you from infringement.

But a good portfolio can create:

bargaining power;
licensing leverage;
investor value;
acquisition value;
cross-license options.

PatentPC explains the strategy in The Impact of AI and Machine Learning on Patent Strategies.

One Patent Is Not a Moat. One Patent Can Still Matter.

The data creates an apparent contradiction.

Our sample shows large campaigns often rely on many patents.

Yet BTF v. ElevenLabs and AI.Law v. Eve each show a single patent can still create a real lawsuit.

The right lesson is:

Quality matters.
Portfolio depth matters.

They are not opposites.

One well-positioned patent can create leverage.

A strong family of related patents can create far more paths to enforcement.

PatentPC’s guide to AI patent portfolio strategy explains how to think beyond raw patent count.

What the Latest AI Patent Lawsuits Actually Tell Us

The evidence now points to six conclusions.

1. Direct GenAI patent litigation has arrived

OpenAI, ElevenLabs, and Eve Legal are no longer theoretical examples.

They are defendants in patent cases.

2. Old patents can reach new AI

Patent families from 2011, 2013, 2016, and 2019 are being asserted against modern AI technology.

3. Patent eligibility remains a major fault line

Recentive lost.

Aon survived.

AI.Law v. Eve is now testing related questions again.

4. AI infrastructure is part of the patent battlefield

Xockets shows that the patent problem extends into servers, networking, DPUs, and GPU systems.

5. Operating companies are enforcing AI portfolios

Cerence is not a passive entity that merely bought old patents to sue.

It is an operating conversational-AI company enforcing technology accumulated through years of development.

6. Patent strategy is becoming part of competitive AI strategy

As the technology becomes more valuable, companies have more reason to ask both:

What can we protect?

and:

What can stop us?

Those are opposite sides of the same patent strategy.

Why the Next Wave Could Be Much Larger

The stock of GenAI patents feeding future disputes is growing quickly.

PatentPC’s own analysis of WIPO’s GenAI patent data found that recent publication velocity has accelerated sharply. See PatentPC’s study on what actually makes an AI company defensible.

That does not mean lawsuits rise one-for-one with patents.

Most patents will never be litigated.

But more patents create:

more overlapping claim space;
more licensing opportunities;
more acquisition targets;
more FTO work;
more potential assertion.

At the same time, AI products increasingly converge on similar technical patterns:

RAG
agents
memory
model routing
voice
multimodal input
tool use
long-context processing
AI infrastructure

That convergence is exactly why patent strategy is becoming more important.

Why PatentPC Focuses So Heavily on Software and AI Architecture

The wrong way to patent an AI company is to send a pitch deck to a lawyer and ask:

“Can you patent this?”

The better process starts with engineering.

Patent counsel should understand:

the architecture;
data flow;
model interactions;
memory;
retrieval;
infrastructure;
latency problems;
security constraints;
fallback paths;
alternative implementations;
competitor design-arounds.

Only then can the patent strategy become useful.

PatentPC works deeply in software, AI, and advanced technology and builds its patent strategy around the technical mechanism rather than dressing a business idea in patent language.

Explore PatentPC’s AI patent resources, AI patent litigation guidance, and deep-tech patent litigation analysis.

For companies already facing a dispute, see Handling AI Patent Infringement Cases.

The Bottom Line

The biggest change in AI patent litigation is not that there are suddenly thousands of proven AI patent lawsuits.

We do not need an inflated statistic to make the case.

The real evidence is stronger.

In the last few years, patent litigation has reached:

OpenAI.
Nvidia.
Microsoft.
Amazon.
Apple.
Samsung.
ElevenLabs.
Zesty.ai.
Eve Legal.

And the patents being asserted span nearly every layer of the AI stack.

PatentPC’s case sample adds another warning:

90% of the enforcement campaigns we studied began after ChatGPT launched, yet 90% assert at least one patent family whose roots predate ChatGPT.

AI technology is moving fast.

Patent rights move on a much longer clock.

That mismatch is where much of the next decade’s litigation risk may come from.

So before asking:

“What AI patents should we file?”

ask one more question:

“What old patents might already cover what we’re building?”

A strong AI patent strategy needs both answers.

Talk to PatentPC about an AI or software patent strategy.

PatentPC Research Methodology

This article does not claim to count every lawsuit worldwide that might be described as AI related.

There is no standardized federal “AI patent litigation” case category.

PatentPC instead built a curated U.S. litigation sample based on ten campaigns in which AI, machine learning, conversational computing, GenAI, AI-native products, or AI infrastructure play a central role.

The ten campaigns analyzed were:

Many Worlds 2T Innovations v. OpenAI
BTF IP Holdings v. ElevenLabs
Corp. v. Butler Labs / Eve Legal
Cerence v. Amazon
Cerence v. Apple
Xockets v. Nvidia/Microsoft/Amazon
Aon Re v. Zesty.ai
Recentive Analytics v. Fox
Cerence v. Samsung
Cedar Lane v. ElevenLabs

Patent counts were taken from complaints, court dockets, government notices, and corporate SEC disclosures.

The 50-patent figure counts distinct patents asserted by the plaintiff within each selected enforcement campaign. It does not count counterclaims or double-count the same patent merely because it was asserted in both district court and the ITC.

Primary External Sources

Recommended PatentPC Deep Dives

General information only

This article is provided for general informational purposes and does not constitute legal advice. Reading it or using this website does not create an attorney-client relationship. Consult qualified counsel about the facts and law applicable to your situation.