As of Sept. 16, 2025 there were 48 open lawsuits against nine AI companies, as depicted in Figure 1 below. The Bartz v Anthropic class action lawsuit is the focus of this blog. Facts associated with this historic case came from a variety of sources, including an excellent article by Authors Guild.

A preliminary settlement of historic proportion ($1.5B) has been reached on the piracy aspect between the plaintifs and the defendant. And authors whose books may have been involved have the option to provide information for the claims process.

Figure 1: Open Lawsuits against AI companies.
Image AI-generated, modified in Photoshop.

In June 2025, US District Court Judge William Alsup ruled that Anthropic’s training of books without permission or compensation was “exceedingly transformative” and therefore “fair use” if the books were acquired legally. However, he also ruled that training on books that were acquired illegally – 7,000,000+ books in this case – are not “fair use” but are piracy. Anthropic not only trained using these books, it retained copies of them in a central library.

Piracy Trial

Judge Alsup scheduled a class-wide piracy trail for December 2025. That was a first in the current AI lawsuits.

On August 26, 2025 the plaintifs and Anthropic reached a settlement that calls for Anthropic to pay approximately $3,000 for each of the 500,000 books estimated to qualify with the appropriate US Copyright Office registration in the timeframes required for award of statutory damages. That brings the total to $1,500,000,000 ($1.5 billion) if there are 500,000 books that meet the copyright registration criteria. If there are more qualifying books than that, the additional books will receive the same compensation.

Preliminary Settlement

On September 25, 2025 Judge Alsup granted a preliminary approval of the settlement terms based on additional information submitted by both the plaintifs and the defendant. He set Oct. 2, 2025 as the deadline for the creation of a searchable Works list with qualifying titles, a long-form notice, and Claim Form at www.anthropiccopyrightsettlement.com.

On Oct. 2, 2025 the search mechanism was activated at https://www.anthropiccopyrightsettlement.com/ to enable authors to determine if any of their books are listed among those pirated by Anthropic. The labeling of how to search is a bit confusing. Click the phrase “Works List” to access the lookup on the website homepage. Searches may be done by ISBN/ASIN, title, author, or publisher. Only pirated works that were registered appropriately with the US Copyright Office and satisfied the date criteria in the class action definition are listed. Works registered more than five years after publication or after Anthropic downloaded them are not eligible to file a claim.

This is the same location to file a claim if your book is on the list. It will be helpful if you find the “Unique ID” for your book provided in the “Works List” citation for you book.

And for more details about the case see the excellent article on Authors Guild.

The post Update on Anthropic Piracy Settlement appeared first on .

Jonathan’s blog calls attention to three aspects of the settlement that are important considerations for all content creators going forward. He makes the critical point that this settlement is about the piracy of over 7,000,000+ books, not about how AI used the books to train its platform using them. It’s about how they acquired the books. He also clarifies  which authors stand to benefit from the settlement and why the vast majority of the authors of the 7,000,000+ will not receive anything. Lastly, he explains why the “eye-watering” settlement of $1.5B will likely have very little impact on Anthropic and other large AI platforms.

NEWS FLASH: It was widely reported the day after Jonathan posted this blog that Judge Alsup in this case postponed approval of the settlement. He indicated he is skeptical on numerous issues in the settlement and has given the parties in the settlement a September 15 deadline in which to provide additional details of the settlement. The next hearing is set for September 25. Jonathan posted a follow up blog that describes the judge’s reaction.

The post 3 Problems with Anthropic Settlement appeared first on .

OpenAI warns us that the outputs generated by their models may not be accurate in their “Terms and Conditions.” Their “Terms and Conditions” also indicate that we should not rely on their AI outputs as a sole source of truth or factual information. That is excellent advice!

https://openai.com/policies/terms-of-use
 

Hallucination Evaluation by OpenAI

 
On April 16, 2025, OpenAI published a report on its own internal evaluation of hallucinations by OpenAI’s o3 and o4-mini models in comparison to its earlier o1 model. In addition to some hard data the report also provides a brief overview of the very complex nature of AI hallucinations.

https://cdn.openai.com/pdf/2221c875-02dc-4789-800b-e7758f3722c1/o3-and-o4-mini-system-card.pdf
 

SimpleQA

 
The first of the two datasets used in the evaluation, named “SimpleQA,” included 4,326 concise questions across many areas. Each question had only one correct answer. See Table 1 below.

Table 1: Hallucination Evaluation by OpenAI – SimpleQA

In the SimpleQA test the older o1 model was accurate 47% of the time and hallucinated 44% of the time. The o1 model declined to respond for 9% of the questions. In comparison the o3 model was accurate 49% of the time and hallucinated 51% of the time. This “new and improved” o3 model responded more often but hallucinated more often in its output. The o4-mini model, on the other hand, was accurate only 20% of the time and hallucinated an incredible 79% of the time.

So, OpenAI was spot on – We absolutely should not rely on their AI outputs as a sole source of truth or factual information!
 

PersonQA

 
The other dataset, named “PersonQA,” asked well-publicized factual questions about public figures. See Table 2 below.

Table 2: Hallucination Evaluation by OpenAI – PersonQA

The o1 model was accurate 47% of the time and hallucinated only 16% of the time. But that left 37% of questions that it declined to answer. In comparison the o3 model was accurate 59% of the time and hallucinated 33% of the time. The o4-mini model was accurate 36% of the time and hallucinated 48% of the time.

The parting thought: “Don’t trust and always verify the AI output.”

The post Investigating AI: How accurate are AI outputs? appeared first on .

In every draft, let honesty preside,
Mark where the silicon muse has supplied.
A ledger kept of prompts and code,
Reveals the hands that helped the words unfold.
Transparency’s torch in the author’s night,
Guides readers true to borrowed light.
For credit given where credit’s due,
Honors the old and welcomes the new.
Let records show what minds entwined—
So trust and truth in books are signed.

AI-generated Badge

The post Investigating AI: Poem on Tracking AI Usage appeared first on .

AI-Generated by ChapGPT

AI Disclosure Requirements

The US Copyright Office requires the disclosure of AI-generated content when applying for a copyright registration. Some AI-created content may be copyrightable, while other uses may not. AI assistance that results in only  a “de minimis” change, such as spell and grammar checking, does not need to be disclosed. However, an increasing number of publishers, academic and scientific journals,  and distribution platforms like Amazon KDP now require disclosure of AI-generated content.

Tracking System

Anyone who has ever written a lengthy work knows the value of employing a robust system to maintain and track revisions, sources, resources, changes, permissions, and related metadata throughout the lifecycle of the work. A comprehensive tracking system should include:

• Revision Tracking
• Source and Resource Management
• Change Management
• Permissions Tracking
• Centralized Notes and Revision Lists
• Collaboration and Access Control

Each of these features may involve the use of artificial intelligence. Records should clearly distinguish between AI-generated and AI-assisted content. For instance, an AI platform could be used to refine the language in a book chapter to better suit a particular audience.

When authors or editors make changes, tools like Track Changes can document the “before” and “after” versions. Tracking changes made by humans is already complex: introducing AI further complicates this process, especially since disclosure of AI-generated content is often required. For disclosure to occur, AI involvement must first be identified.

Until mainstream editors like Microsoft Word and Google Docs can natively distinguish between human and AI edits, a practical workaround could be used. Figure 1 demonstrates how an AI platform can be identified in the “label” field when comparing two documents.

Fig. 1: MS Word Track Changes Label Field

De Minimis

There are two considerations that determine if an AI use needs to be disclosed when required by most entities:

1. What is the origin of the output?

• Did the output start with human-authored content that was then refined by AI?

The copyrightability of content is most likely when is was solely created by a human. The use of traditional tools for spelling or grammar checking do not have to be disclosed.

• Was the output the result of a prompt input into a generative AI system.

The US Copyright Office (USCO) has taken a hard position that content that was produced on a generative AI system with the only human involvement being the input of prompts is not copyrightable. Recent court cases challenging this position have failed, regardless of the number of prompts that were input.

2. How much of the output was created by the generative AI system?

The USCO does not require the disclosure of a “de minimis” amount of content that was generated by a generative AI system. By definition, “de minimis” means a small, insignificant amount of something. For copyrightability purposes, that means a small amount of change to something that was first created by a human.

How can you make the judgement of how much AI content is “de minimis” or not. If the AI-generated content was created by a human, would it be copyrightable? If it would be copyrightable, it is not “de minimis” and would have to be disclosed. And content and context are important. The number of words is not necessarily a good indicator. A short poem that was generated from prompts into an AI system is not copyrightable for reason #1 above.

While Amazon KDP takes this same position of no need to disclose a “de minimis” amount of AI, other entities may require disclosure any use of AI content generation.

Recommendation

When you use AI for a work you wish to publish, make a record of each AI transaction and categorize the output as either AI-generated to be disclosed or “de minimis” that may not need to be disclosed.

If you are not yet convinced, have a look at A Poem on “The Virtues of Keeping Track of the Uses of AI in Writing a Book”.

The post Investigating AI: Track AI Usage appeared first on .

Preface by webmaster: This is a repost of a blog by Jonathan Bailey at Plagiarism Today titled “The Gradient of AI Usage.” He identifies ten excellent examples of content generation that range from entirely human authorship to entirely AI-Generated content using only prompts. This blog is reposted with permission.

It should be noted that Jonathan’s post was discovered after we had created the “Copyrightability Gradient” graphic that was included in a blog we initially posted on April 16, 2025. The blog has subsequently been updated to include new reports out of the US Copyright Office.

The post The Gradient of AI Usage appeared first on .

This blog focuses on the copyrightability of AI-created content and requirements to disclose and differentiate between AI-Generated, AI-Assisted and Human Authored content. Why is some AI content not copyrightable and other AI content is copyrightable?

Our understanding of the nuances of copyright law comes from having researched and obtained the rights to use thousands of copyrighted sources. We have been making presentations on “copyright basics” since 2010. Our award-winning guide “Copyright Clearance for Creatives” gets into the details of copyright law.

The US Copyright Office has published several major publications such as annual reports, policy studies, topical reports, and entries into the Federal Register that identify the rules, regulations, policies, and procedures related to the copyrightability of AI-created content.

We have heavily been using AI assistance to research the copyrightability of AI-created content. The US Copyright Office has disclosure requirements when applying for copyright registration. It has been a challenge to distill this into something that is easy to get your head around.

Relevant Publications

The following publications identify and clarify the US Copyright Office’s positions regarding artificial intelligence:

Compendium of U.S. Copyright Office Practices, Third Edition. January 28, 2021. This 1,396-page publication is the governing administrative manual for copyright registration and recordation on and after 1/28/2021.

Copyright Registration Guidance: Works Containing Material Generated by Artificial Intelligence. March 16, 2023.
This is a 6-page statement of policy “to clarify its practices for examining and registering works that contain material generated by the use of artificial intelligence technology.”

Application Process for Registration of Works with Artificial Intelligence-Generated Content. Transcription from Online Webinar on June 28, 2023.
This 11-page webinar transcript discusses how the US Copyright Office examines a work that contains AI-generated material. The USCO has historically had four categories of unclaimable material that cannot be copyrighted. They are now treating AI-generated content as the fifth category of unclaimable material.

Copyright and Artificial Intelligence, Part 1: Digital Replicas. July 31, 2024. This 57-page publication addresses “the use of digital technology to realistically replicate an individual’s voice or appearance.”

Copyright and Artificial Intelligence, Part 2: Copyrightability.  January 29, 2025. We have cited this 41-page publication extensively in our presentation “Copyright in the Age of Artificial Intelligence.”

Copyright and Artificial Intelligence, Part 3: Generative AI Training. Pre-publication Version. May 9, 2025. This 108-page report contains 588 footnotes. The section on “Fair Use” is particularily noteworth, given the number of lawsuits against AI platforms for copyright infringement. The courts will determine what is fair use and what is not.

“When you come to a fork in the road, take it.”

Yogi Berra famously said, “When you come to a fork in the road, take it.” This saying resonates with the choices we faced during the copyright registration process, as described in the first three blogs of this series.

In our experience, the “fork in the road” analogy can be expanded to represent three distinct paths in the realm of copyright and artificial intelligence:

• Human Authorship: This is the traditional and well-understood route, where works are created entirely by humans. These works are clearly eligible for copyright protection.
• AI-Generated Materials: This path involves works created solely by artificial intelligence, with minimal human input—typically just the submission of prompts. According to current U.S. Copyright Office policy, such works are not eligible for copyright protection. This policy is currently being challenged in court.
• AI-Assisted Materials: This is the third path, representing a grey area where humans and AI collaborate on a work. The eligibility of these works for copyright is evaluated on a case-by-case basis, depending on the extent and nature of human creative involvement.

Just as Yogi Berra’s advice suggests making a choice when faced with multiple options, those navigating copyright registration must carefully consider which path their work follows—human authorship, AI-generated, or AI-assisted. Each option has different implications for copyright eligibility.

Illustration of Copyrightability

We have identified a simple and hopefully easily understood way to graphically explain these options. Figure 1 captures our understanding of the copyrightability of AI-created content, and the standalone illustration is copyrightable in our view. We filed an application to register the copyright for the illustration as a standalone 2-dimensional artwork to legally confirm this is sufficiently creative to warrant copyright protection. We shall see if the US Copyright Office concurs.

Look for additional blogs that discuss AI-Generated and AI-Assisted materials.

@TheCopyrightDetective is the handle used to identify Dr. C. Daniel Miller on social media, and the name is federally registered as a trademark.

#TheCopyrightDetective

The post Registration Process – Part 4: Copyrightability appeared first on .

Copyright and Artificial Intelligence, Part 3: Generative AI Training

The report is 113 pages with 588 footnotes gives a lot of feedback from comments for and against AI platforms copying everything that has been digitized in the world. It provides a good explanation of how AI platforms are trained by making copies of copyrighted content, which is one of the exclusive legal rights of the copyright holder.

While the report identifies this copying as “prima facie infringement” of copyrighted content, it leaves the door wide open for a primary “fair use” defense by AI developers. And it will likely be the main defense that is being used in the court cases that are already in progress. On Feb. 11, 2025 the first of these cases was decided. The court ruled against the “fair use” defense used by Ross Intelligence in the Thomson Reuters v Ross Intelligence infringement case.

A large portion of this report discusses how and why the  “fair use” doctrine will be used to defend the actions of AI developers. It is a good explanation of the concept and how the courts may weigh the four factors that are used to determine if a use was fair or not. The US Copyright Office did not, however, make any blanket statements or take a stance on AI developers’ liability. The report basically says infringement liabilities will be decided in the courts on a case-by-case basis.

The most significant factor is the “Purpose and character of the use” that considers the “transformative” nature of the outputs of the AI platforms. This factor was the key to the court ruling in favor of Google in 2015 in Authors Guild v Google, declaring that Google’s digitally coping millions of books without permission or payment was “transformative” and therefore “fair use.”

The bottom line is that the US Copyright Office expects some of the courts to rule against the AI developers, as they just did in Reuters v Ross, and rule in favor of others, as they did in Authors Guild v Google.

The legal landscape is very unsettled, with numerous lawsuits and legislative efforts underway. The US Copyright Office’s role is to implement laws passed by Congress and court decisions. The Office creates the rules, regulations and policies for that implementation, and this report is not that. It does provide a framework to analyze copyright implications of generative AI training.

The post Part 3: Generative AI Training appeared first on .

Additionally, the series will identify some best practices for using AI ethically, legally, and with full disclosure and transparency in mind. The AI Transparency Badges that were introduced in a previous blog will be used to indicate if the entire blog is either AI-Generated, AI-Assisted, or Human Authored.

AI Transparency Badges

Future Blog Topics

Topics likely to be addressed in future blogs include:

Copyright Basics

• Legal Rights
• Infringement
• Exercising Legal Rights
• Fair Use
• Public Domain

AI Basics

• Generative AI
• AI uses in Publishing
• How does AI Work?
• AI Training
• Copyright Infringement
• AI prompts
• AI Hallucinations
• AI Biases

Intersection of copyright and AI

• Claimable and Unclaimable Materials
• Part 1: Digital Replicas
• Part 2: Copyrightability
• Part 3: Training AI Models

AI’s impacts on Creative Industries

• Job Displacement
• Devaluation of Skills
• Discovery Bias
• Copyright & Remuneration
• Quality & Originality
• Over-reliance on AI

The blog series will be coordinated with a series of short podcasts on the same topic.These podcasts can be found on The Copyright Detective’s YouTube channel.

@TheCopyrightDetective is the handle used to identify Dr. C. Daniel Miller on social media, and the name is federally registered as a trademark.

#TheCopyrightDetective

The post Investigating AI: Introduction appeared first on .

AI Transparency Badges

Human Authored Badge

AI transparency is a rapidly growing issue. AI transparency badges could be one way to quickly identify how artificial intelligence was used in content creation. This blog was Human Authored without any use of artificial intelligence, for example.

The US Copyright Office requires writers to disclose how artificial intelligence was used in crafting a work submitted for copyright registration. The word crafting is defined as planning, making and executing activities, all of which could involve AI. I clearly identify AI-generated and AI-assisted content on the PowerPoint slides in my “Copyright in the Age of Artificial Intelligence” presentation, for example.

A recording of a Zoom presentation was submitted with an application to register the copyright for the PowerPoint slides and and my voice narration. Another recent blog describes this disclosure process. Success! The US Copyright Office issued a Certificate of Registration on April 9, 2025.

Basis of Claim: text, photograph(s), compilation of artwork generated by artificial intelligence.

I recently joined The Authors Guild because I like and want to support an organization that has advocated on behalf of professional writers for over a century. They have also started a “Human Authored Certification” program to distinguish human creativity from materials that have been generated by artificial intelligence. They don’t offer badges to identify when a work is not exclusively human authored. That is what motivated me to create my own badges.

Copyrightability Illustration

Copyrightability of AI-Generated and AI-Assisted Materials

These three types of content creation/generation are represented graphically in a recent blog. That illustration, depicted to the left, has been submitted for copyright registration. The application is for a two-dimensional artwork. No artificial intelligence was used in crafting the illustration. It has only Human Authorship, which is  “Copyrightable” if sufficient creativity and originality is demonstrated.

The Three Badges

Human Authored Badge

The first badge we created indicates that the work was Human Authored: no AI was used in it’s creation. As previously stated, works that are solely human created can be registered for copyright protection if the work shows “some minimal degree of creativity” and originality. I will provide an update when I find out if the “Copyrightability” illustration shown above passes those tests.

AI-assisted Badge

The second badge indicates that the work started with content that was human authored. This content was then refined, enhanced, clarified, and otherwise modified a “de minimis” amount using artificial intelligence flatworm and applications. A de minimis amount of AI-assisted content does not have to be disclosed on an application for copyright registration or in a submission of a work on Amazon KDP.

AI-generated Badge

The third badge is for content that was AI-generated only using prompts to generated the content. A common example of this kind of use is in the production of images using the ChatGPT AI platform. The US Copyright Office has consistently ruled that content created by only using prompts, no matter how many prompts, is not claimable. In other words, it may not be registered for copyright protection. There are cases working their way through the courts challenging these rulings, but none of them have made to the US Supreme court to date.

@TheCopyrightDetective is the handle used to identify Dr. C. Daniel Miller on social media, and the name is federally registered as a trademark.

#TheCopyrightDetective

The post AI Transparency Badges appeared first on .