Anthropic Makes Huge and First-Ever Move to Protect its Users from Copyright Lawsuits

In what appears to be a first for large language models, Anthropic says its new Commercial Terms of Service will “enable our customers to retain ownership rights over any outputs they generate through their use of our services and protect them from copyright infringement claims. 

It is an important move to promote use of large language models without fear of such legal actions, given the nascent state of AI copyright law as it applies to use of LLMs. One might also note it creates a huge new level of uncertainty about the business risks faced by LLMs, as such litigation is virtually certain, over time. 

“Under the updated terms, we will defend our customers from any copyright infringement claim made against them for their authorized use of our services or their outputs, and we will pay for any approved settlements or judgments that result,” Anthropic says. “These new terms will be live on January 1, 2024 for Claude API customers and January 2, 2024 for those using Claude through Amazon Bedrock.”

There are other steps LLMs can take to limit the uncertainty associated with copyright risks, such as using robust copyright filters, which help identify and flag potentially infringing content before it's generated or shared with users.

Ensuring transparency and responsible sourcing of training data, with clear mechanisms for identifying and excluding copyrighted material, also can minimize the risk of incorporating infringing elements into LLM outputs.

Establishing partnerships and clear guidelines for collaboration with copyright holders can lead to mutually beneficial licensing agreements and promote fair use of copyrighted material within LLMs, and is an obvious avenue. 

Beyond all that, copyright related to LLMs will develop over time based at least in part on prior rulings related to use of content. 

Several existing legal precedents offer potential legal avenues for addressing large language model (LLM) copyright issues, one might suggest. 

Fair use is an obvious issue, as large language models are trained on huge amounts of existing content. After all, all human knowledge is built on prior work, with only the unique expression of facts being protected by copyright, not the facts themselves. 

Campbell v. Acuff-Rose Music, Inc. (1994) established a four-factor test for fair use, considering the purpose and character of the use, the nature of the copyrighted work, the amount and substantiality of the portion used, and the effect of the use upon the potential market for or value of the copyrighted work. 

Sony Music Entertainment v. Diamond Way Recordings, Inc. (2003) clarified the definition of a derivative work, stating that it must "recapture the essential elements of the original" and create a new work with a different purpose or character. 

Also, the case of Sony Corp. of America v. Universal City Studios, Inc. (Betamax VCR case) (1984) said the creation and use of a device solely for the purpose of making fair use copies of copyrighted works does not constitute copyright infringement.

Some might argue no LLM can create copyrighted material. Alfred A. Knopf, Inc. v. Colby (1992) held that an expert system's creative output lacked the requisite human authorship for copyright protection.

Is accumulated human knowledge similar to a database? If so, then some precedents related to databases could apply. Feist Publications, Inc. v. Rural Telephone Service Co., Inc. (1991) ruled on the scope of copyright protection for databases, stating that only the selection and arrangement of facts, not the underlying data itself, is protected. 

The European Union's "Text and Data Mining" exception allows certain research institutions to mine copyrighted works for non-commercial purposes without the copyright holder's consent. 

Also, open-source licenses like the GNU General Public License (GPL) could be relevant if LLMs are trained on datasets containing open-source materials.

New legal doctrines specific to AI and machine learning, such as the "scene a faire" doctrine and the "merger doctrine" limit copyright protection for elements that are dictated by the functionality or nature of a particular work.

Compulsory License and AI Copyright

Language models are creating new terrain for copyright law.

Personally, I favor as much freedom as we can possibly endure. On the other hand, practical economic realities are likely to dictate outcomes that balance payments to rights owners and encouragement of innovation by AI firms. 

History strongly suggests that pattern will emerge. 

In the past, rights holders usually wanted control (veto power over new uses). Courts and Congress usually compromise by substituting compensation (compulsory licenses, collective societies, levies) for control. 

Radio stations, cable companies, physical media distributors and streaming platforms captured enormous value from others' content, then eventually reached licensing accommodations. 

Player piano rolls in the late 1800s reproduced sheet music mechanically, and publishers argued this was infringement. 

The Supreme Court ruled in White-Smith Music v. Apollo (1908) that piano rolls were not copies because copyright covered notation readable by humans, not mechanical reproductions.

Congress then passed the Copyright Act of 1909, which created the  compulsory mechanical license, under which anyone could record a song already recorded by someone else, provided they paid a statutory royalty.

Radio created a new problem:

• Composers and publishers got performance royalties through ASCAP (founded 1914), which negotiated blanket licenses with broadcasters.

• Performers and record labels got nothing. The law treated a broadcast performance as fundamentally different from a mechanical reproduction, so the actual recording artists received no royalties when their records were played on air.

Film created other issues. Sound film after 1927 meant a single film now embodied multiple copyrights simultaneously:

• the screenplay

• the musical score

• the recorded performances

• the film itself. 

The question of who owned what in a collaborative industrial production led to the work-for-hire doctrine. Studios owned everything, employees and contractors owned nothing.

International distribution exposed the territorial nature of copyright immediately. The Berne Convention (1886, continuously expanded) became the framework for international harmonization.

Magnetic tape, then cassettes, then VHS created successive waves of the same basic problem: cheap, accessible reproduction technology. 

The 1971 Sound Recording Amendment was the first law to give sound recordings their own copyright protection (previously, only the underlying composition was protected).

Sony v. Universal City Studios (the "Betamax case," 1984) is one of the most consequential copyright decisions ever. Hollywood sued Sony for making VCRs, arguing Sony was liable for the infringement its customers committed. 

The Supreme Court ruled 5-4 that:

• Time-shifting (recording TV to watch later) was fair use

• Selling a device with substantial non-infringing uses didn't make the manufacturer liable.

The Betamax decision shaped technology law for decades. It's why VCRs, DVRs and the consumer electronics industry could exist without needing Hollywood's permission. 

The Audio Home Recording Act (1992) eventually addressed digital audio tape (DAT) by requiring copy-protection technology in DAT recorders and adding a levy on blank digital media, a compromise that established the principle that hardware makers could be taxed to compensate rights holders.

Cable TV initially retransmitted distant television broadcast signals without compensation. The Supreme Court ruled twice (1968, 1974) that cable retransmission wasn't infringement because cable companies weren't "performing" the works. 

But the Copyright Act of 1976 created a compulsory license for cable retransmission.

This compulsory-license-as-compromise model recurs throughout copyright's adaptation to new media.

Xerography created frictionless reproduction of printed text.

The 1976 Act addressed this partly, and the Copyright Clearance Center was founded in 1978 as a collective licensing organization, allowing libraries and businesses to pay blanket fees for photocopying rights.

CDs introduced a paradox: perfect digital copies. This era produced the first serious deployment of Digital Rights Management concepts. It also sharpened the debate about the first sale doctrine — you can resell a CD you bought, but can you resell a digital file? 

The DMCA (1998) was the legislative response. 

MP3 and the iPod era forced the unbundling of the album into individual songs, collapsing per-unit revenue and forcing the industry toward licensing models.

Streaming is philosophically interesting because it doesn't fit the reproduction model at all — you're not copying anything, you're accessing a performance. This is actually closer to the original concept of performance rights than anything since radio.

Spotify, Netflix, and their successors forced the creation of entirely new licensing frameworks:

mechanical licenses for on-demand streaming, negotiated rates between platforms and labels, windowing strategies for film. 

The Music Modernization Act (2018) was the most significant music copyright reform in decades, largely cleaning up the licensing mess streaming had created.

Crucially, streaming finally gave performers (not just composers) digital performance royalties. 

AI companies likely will eventually fit this compulsory license pattern almost perfectly.

The internet produced many precedents that might shape AI copyright:

• The Digital Millennium Copyright Act

• The notice-and-takedown system (rights holders can demand platforms remove infringing content, and platforms that comply get "safe harbor" protection from liability

• Anti-circumvention rules making it illegal to bypass digital rights management (DRM) technology

• ISP liability limits** — shielding internet service providers from liability for what users transmit, provided they act on takedown notices

• The No Electronic Theft Act (1997) closed a loophole where non-commercial infringement wasn't clearly criminal. Before it, you had to be *selling* pirated content to face criminal liability

• The Sonny Bono Copyright Term Extension Act (1998) extended copyright terms by 20 years

• Perfect 10 v. Amazon/Google (2007) established that search engine thumbnail images and inline linking could qualify as fair use

• Authors Guild v. Google validated Google's mass scanning of copyrighted books as fair use

• Viacom v. YouTube (2012) confirmed that platforms don't lose protection simply because infringement is general knowledge

• The Napster and Grokster cases (2001, 2005) established "contributory infringement" and "inducement" doctrines: you can be liable not just for infringing yourself, but for building a platform *designed* to encourage infringement. 

Beyond law, the internet forced copyright to adapt through raw economic pressure:

• Licensing at scale became essential

• Creative Commons (2001) created a voluntary licensing system letting creators specify in advance what reuse they permit

• Terms of service became a shadow copyright system

• Geoblocking and regional licensing became standard business practice as companies realized the internet didn't respect territorial copyright boundaries that the entire global licensing system was built around.

The historical pattern suggests the outcome will likely be some combination of new licensing frameworks, platform liability rules, and compensation mechanisms rather than either "AI training is fully free" or "AI training is categorically infringing." 

The internet precedents suggest the law will bend toward enabling the technology while extracting some structural protections for creators.

My own thinking leans towards permissiveness where it comes to the use of content. As for the argument that language models unfairly infringe copyrights because they “read” or “ingest” content, my argument would be that this non-infringing work is what humans do routinely, and it is not copyright infringement. 

When a person reads a novel, watches a film, or listens to music, they:

• Absorb patterns, styles, vocabulary, narrative structures

• Develop taste and skill influenced by what they've consumed

• Produce new works that are clearly downstream of that consumption

• Do all of this without paying royalties or seeking permission

Nobody considers this a copyright violation, even when the influence is obvious. A novelist can write "in the style of Hemingway" after reading all his books, or a musician whose sound is clearly shaped by artists they grew up listening to.

AI models do that, but with some important differences, some will argue:

• Scale and speed

• Reproduction during training, which involves making copies of copyrighted content and storing them (transiently) on servers. Courts have historically treated that mechanical act of copying as significant, regardless of what happens downstream

• Memorization and regurgitation at scale that arguably collapses the distinction between learning from something and copying it

• Market substitution when an AI trained on a photographer's portfolio can produce images that replace demand for that photographer's work.

The enduring issues suggesting preserving freedom are:

• knowledge and style aren't ownable

• culture builds on culture

• the internet was built on the assumption that content could be read and indexed.

The case for creator protection rests on: 

• copyright exists precisely to ensure creators can sustain the work of creation

• the AI industry is extracting enormous commercial value from creative labor without compensation

• "humans do it too" ignores that humans don't build billion-dollar products directly monetizing others' uncredited work.

The resolution will require development of compensation structures that make the ecosystem fair to creators while not strangling a genuinely transformative technology.

In my view, compulsory license in some form.

AI Clash Between Copyright and New Technology is an Old Tale

Every new technology brings with it new legal issues. Artificial intelligence, for example, raises copyright issues. 

It is not the first time new technology has clashed with established notions of copyright. 

When photocopying machines were commercialized, manufacturers tried to block the use of the machines for making copies of copyrighted work.

Sony tried to block the use of videocassette recorders to time shift video content for later viewing. Similar disputes erupted over the use of audiocassette tapes, music file sharing and video streaming as well. 

New Technology	Copyright Issues	Key Court Decisions
Photocopying Machines (1960s)	Mass reproduction of copyrighted materials without permission.	Fair Use Doctrine Established: Williams & Wilkins Co. v. United States (1964) established the four-factor fair use test: purpose and character of use, nature of copyrighted work, amount and substantiality of portion used, and effect of use upon the market. Copying for educational purposes could be fair use.
Audio Cassette Tapes (1970s)	Home recording of copyrighted music threatened record sales.	Audio Home Recording Act (1992): Established a royalty levy on blank audiotapes to compensate copyright holders for potential lost sales due to home recording.
MP3 Players and Napster (1990s)	Peer-to-peer file sharing enabled widespread music piracy.	A&M Records v. Napster (2001): Napster was found liable for contributory copyright infringement for failing to prevent users from sharing copyrighted music.
Streaming Services (2000s-Present)	Distribution model challenged traditional music licensing and revenue streams.	Negotiated Licensing Agreements: Streaming services like Spotify and Apple Music pay licensing fees to copyright holders based on user streams.
Digital Video Recorders (DVRs)	Shifting time viewing challenged broadcasters' control over programming.	Sony Corp. v. Universal City Studios (1984): Upheld the fair use of time-shifting for personal viewing using VCRs.

Similarly, conflicts have erupted over content, social media, search, open source software and user-generated content, for example. 

Content Issue	Copyright Issues	Key Court Decisions
Social Media Sharing	Sharing copyrighted content like photos, videos, and music raises questions of fair use and infringement.	Blurred Lines: Perfect 10 v. Amazon (2002) established thumbnails could be fair use for linking purposes. However, sharing entire works without permission is generally considered infringement. The specific context and amount used determine fair use.
User-Generated Content (UGC) Platforms	Platforms like YouTube or TikTok host user-uploaded content, potentially infringing on copyrights.	DMCA "Safe Harbor": The Digital Millennium Copyright Act (DMCA) provides a safe harbor for platforms if they remove infringing content upon notification from copyright holders. Platforms like YouTube have automated takedown systems based on copyright claims.
Software Sharing and Open-Source	Sharing copyrighted software raises concerns about piracy and unauthorized distribution.	Open-Source Licenses: Open-source licenses like GPL (General Public License) allow for modification and sharing of software code, as long as certain conditions are met. These licenses provide a framework for collaborative software development while protecting copyright.
Content Aggregation Services	News aggregators like Google News display headlines and snippets of copyrighted news articles.	Fair Use and Fair Reporting: Courts have generally allowed news aggregation under fair use for purposes of reporting and commentary. The amount and substantiality of content used are crucial factors.

Eventually we will figure out some balance between copyright and use of the new technology in non-infringing ways. But it may take a while.

AI Copyright Balance Will Emerge: All Prior Content Technology Innovations Have Done So

Copyright is a tricky business, and has been since the advent of digital media. But artificial intelligence likely will cause a rethinking or adaptation of copyright, in part because it is getting harder to distinguish between a human author’s particular formulation of an idea and an AI-generated alternative. 

Traditionally, copyright has been designed to encourage innovation by providing creators with limited monopolies over their work, protecting the particular expressions of ideas, but not the ideas themselves.

So copyright protects an author’s expression of an idea, not ideas, facts, or styles themselves. 

Abundance, instead of scarcity, partly explains why that is the case.

Copyright traditionally assumed scarcity and control over physical copies (books, records, films). All that became more challenging when digital distribution; the internet; user-generated content and easy remixing of content replaced scarcity and distribution cost:

• Digital formats (MP3s, JPEGs, PDFs) allowed lossless copying at near-zero cost, unlike analog media. Napster (late 1990s) and peer-to-peer sharing exemplified mass infringement

• Global, instantaneous sharing by websites, torrents, and streaming bypassed traditional gatekeepers

• Social media, YouTube and many other sites blurred lines between consumers and creators, increasing derivative works, remixes, and mashups

Era	Key Challenges	Fair Use Evolution	Legislative,  Policy Responses
Pre-Digital (pre-1990s)	Physical copying limits; analog scarcity	Narrow: criticism, parody, education	1976 Copyright Act  codifies fair use & rights
Digital Media and Internet (1990s-2010s)	Perfect digital copies, P2P sharing, online distribution	Expanded for search/indexing (Google Books), software interoperability (Google v. Oracle)	DMCA (1998): safe harbors, anti-circumvention; longer terms
AI/Generative (2020s-)	Mass scraping for training, output substitution, authorship	Split rulings: often transformative for training (Anthropic, Meta) but market harm weighs heavily; fact-specific	Ongoing reports (US Copyright Office, UK, EU); licensing debates; deepfake proposals

Generative AI arguably complicated matters further:

• Model training is based on use of massive datasets, often scraped from the internet, raising questions of reproduction rights. AI companies argue it's necessary for learning patterns/ideas (not protected expression); creators call it systemic infringement.

• AI can generate content mimicking styles, potentially causing "market dilution" as a new issue

• AI outputs generally lack human authorship for copyright, but human-prompted or edited works are gray areas

• Training data is often non-transparent and opt-out mechanisms are impractical at scale

• U.S. courts show diverging fair use rulings; reports from U.S. Copyright Office, UK, EU Parliament highlight needs for licensing, transparency, or new frameworks (e.g., compulsory licensing debates).

The main policy choices are whether to create new rules for training data, how to treat AI-generated outputs and whether to add special protections for likeness, voice, or style cloning.

The core tradeoffs, as always, are between innovation and creator compensation. Broader licensing can raise costs for AI developers and small firms, but it may also reduce litigation and give rights holders a clearer market. 

Looser rules can speed model development, but they may weaken incentives for human creators.

For now, the U.S. Copyright Office says existing copyright principles are flexible enough to handle AI. 

As with prior content industry technology changes, some balance will emerge that compensates content owners and also allows innovation.