Anthropic and music publishers argue fair use in AI copyright case

Requested: openai/gpt-6-astra · Returned: gpt-6-astra

### 1. How did the publishers characterize the market harm risk?

**Sources state:** The publishers argued that Anthropic uses their lyrics to generate competing content that harms human songwriters [L14]. Their attorney cited nearly 120,000 requests for “synthetic” lyrics over six months and warned that models like Claude “could overwhelm the market with AI-generated content very easily” [L15].

**Inference:** Their argument treats AI-generated lyrics as potential substitutes for human-written lyrics, with high-volume generation posing a competitive threat—not merely a risk of copying particular songs.

**Confidence: high.** The source directly reports both the publishers’ characterization of harm and their volume-based warning.

### 2. Which factual differences about market harm could matter most to Judge Lee?

**Sources state:** Judge Lee questioned whether Claude’s ability to compose lyrics threatens the publishers’ lyrics market [L12]. Anthropic’s attorney said demand had not declined and that the publishers could identify no specific instances of market harm [L16]. The source describes differing judicial views on AI-training fair use, but does not explain the market-harm evidence in those earlier cases [L20–L21].

**Inference:** The most relevant distinctions could be whether there is concrete evidence of displaced licensing transactions, whether users actually substitute Claude-generated lyrics for licensed lyrics, and whether any demand decline is attributable to Claude rather than other causes. Those facts would help distinguish demonstrated market effects from the publishers’ warning about potential market saturation. The pack does not establish how Judge Lee would weigh them or how they compare with the earlier cases.

**Confidence: medium.** The hearing’s reported focus supports these inferences, but the pack lacks detailed evidence and the judge’s conclusions.

### 3. What evidence would be needed to show whether Claude has reduced demand for licensed lyrics?

**Sources state:** The pack does not answer what evidence would be sufficient. It reports Anthropic’s assertion that demand had not fallen and that no specific instances of market harm had been identified [L16], alongside the publishers’ synthetic-lyrics request count [L15]. It supplies no licensing-demand data.

**Inference:** Useful evidence would include:
- Licensing volumes, revenues, and prices over time.
- Documented instances of customers choosing Claude-generated lyrics instead of purchasing licenses.
- Evidence about what users did with generated lyrics and whether they would otherwise have licensed lyrics.
- Comparisons that distinguish Claude’s effects from other influences on demand.

Request volume alone would not establish lost licensing demand.

**Confidence: medium.** These evidence categories address the reported dispute, but the source provides neither the underlying data nor a standard for sufficient proof.

### 4. What assumption connects the nearly 120,000 requests to a threat to licensed lyrics?

**Sources state:** The publishers’ attorney linked nearly 120,000 synthetic-lyrics requests over six months with the warning that AI could overwhelm the market [L15]. Anthropic disputed that there was reduced demand or identifiable market harm [L16].

**Inference:** The connecting assumption is that a meaningful share of those requests produces lyrics that replace—or could replace—lyrics users would otherwise license. That assumes substitutive demand, rather than merely experimentation or new uses that would never have generated licensing purchases. The request count itself does not establish that assumption.

**Confidence: high.** Substitution is the clear inferential link between generation volume and lost licensing demand, although the pack does not prove it occurs.

Requested: anthropic/claude-sonnet-5 · Returned: anthropic/claude-sonnet-5

1. **From the source:** The publishers' attorney Matt Oppenheim argued that Anthropic "misuses its lyrics to generate competing content that harms human songwriters" [L14]. He stated that Claude received nearly 120,000 requests to generate "synthetic" lyrics over a six-month period, and argued that large language models like Claude "could overwhelm the market with AI-generated content very easily" [L15]. Judge Lee questioned "the publishers' assertion that Claude's ability to compose lyrics threatens the market for the publishers' lyrics" [L12], indicating some skepticism about this characterization.

Confidence: high — the characterization is directly quoted and attributed in the source text.

2. **Inference:** The sources note that prior rulings split on fair use: Judge Alsup ruled for Anthropic on book training [L21], while Judge Chhabria ruled for Meta but warned AI training "in many circumstances" would not be fair use [L21]. Given this split, factual differences that could matter most to Judge Lee likely include: (a) whether there is concrete, documented evidence of market substitution (as opposed to speculative harm) — Mehta's argument that publishers "could not identify any specific instances of market harm" [L16] suggests this evidentiary gap could be decisive; (b) the volume and nature of generated content (e.g., the 120,000 synthetic lyric requests [L15]) as a proxy for scale of potential market displacement; and (c) industry-wide statements about AI's economic impact, such as UMG's own characterization of AI as "overwhelmingly a net positive" [L17], which could undercut the harm narrative.

Confidence: medium — this is a reasonable inference drawn from the specific arguments and quotes in the source, but the sources do not state what Judge Lee will actually weigh most heavily.

3. **Open question:** The sources do not provide specific data, methodology, or evidence that was presented to demonstrate actual demand reduction for licensed lyrics. Mehta asserted "AI has not reduced demand for the lyrics" [L16] and that publishers "could not identify any specific instances of market harm" [L16], but no empirical study, sales data, licensing revenue figures, or market analysis is described in the source pack. The sources do not answer what evidence would specifically be needed to prove or disprove demand reduction.

Confidence: high — the source pack plainly does not contain this information, making the gap clear.

4. **Inference:** The implicit assumption connecting the 120,000 synthetic lyric requests [L15] to market threat is that a request to generate synthetic lyrics is a reasonable proxy for either (a) a substitute transaction that would otherwise have involved licensed/human-written lyrics, or (b) evidence of scale that could, if extrapolated, "overwhelm the market with AI-generated content" [L15]. This assumes users treat Claude's output as a functional substitute for, rather than a complement to, publishers' lyrics — an assumption Mehta directly disputed by noting no specific harm instances were identified [L16] and citing UMG's own "net positive" characterization [L17].

Confidence: medium — this is my inference about the underlying logical connection; the source does not explicitly state this assumption but the structure of the argument implies it.

Requested: google/gemini-4-argon · Returned: google/gemini-4-argon

**1. From the source: How did the publishers characterize the market harm risk from AI-generated lyrics?**

* **What the sources state:** The publishers asserted that the ability of Anthropic's chatbot Claude to compose lyrics threatens the market for the publishers' own lyrics [L12]. Specifically, the publishers' attorney, Matt Oppenheim, argued that Anthropic misuses their lyrics to generate competing content that harms human songwriters [L14]. Oppenheim stated that Claude received nearly 120,000 requests to generate "synthetic" lyrics over a six-month period and contended that large language models like Claude "could overwhelm the market with AI-generated content very easily" [L15].
* **Inference:** The publishers frame the market harm primarily as a risk of market saturation and direct substitution, where the high volume and ease of generating synthetic lyrics will crowd out and replace the work of human songwriters. 
* **Confidence:** high — The source directly quotes and summarizes the publishers' attorney's arguments regarding market harm at the hearing.

***

**2. Inference: Given that prior rulings split on AI training fair use, which factual differences about market harm could matter most to Judge Lee?**

* **What the sources state:** Two other Northern California judges have split on whether generative AI training generally constitutes fair use: Judge William Alsup sided with Anthropic regarding books, while Judge Vince Chhabria ruled for Meta in a similar suit but warned that AI training "in many circumstances" would not be fair use [L20, L21]. In the current case, Judge Eumi Lee pressed both sides on fair use and specifically questioned the publishers' assertion that Claude's ability to compose lyrics threatens the market for their lyrics [L11, L12]. While the publishers pointed to nearly 120,000 requests for "synthetic" lyrics and the potential to overwhelm the market [L15], Anthropic's lawyer countered that AI has not reduced demand for lyrics, that the publishers could not identify any "specific instances of market harm" from Claude or other LLMs, and that UMG itself called AI "overwhelmingly a net positive" for the music industry [L16, L17]. Anthropic also argued its training is transformative, helping Claude understand language and enabling progress in science, business, and education [L13, L23].
* **Inference:** Because Judge Lee specifically questioned the publishers' market threat assertion [L12] and Judge Chhabria warned fair use depends on the "circumstances" [L21], the factual difference that could matter most to Judge Lee is whether the publishers can prove concrete, actualized market substitution (specific instances of reduced demand for their lyrics) rather than merely theoretical or potential market flooding (the 120,000 requests for synthetic lyrics), especially when weighed against UMG's own admission that AI is a "net positive" and Anthropic's claims of broad transformative purpose.
* **Confidence:** medium — While the source clearly identifies the exact points of contention regarding market harm and Judge Lee's skepticism, how a judge will weigh specific factual differences in a future ruling is inherently speculative.

***

**3. Open question: What evidence would be needed to show whether Claude has reduced demand for licensed lyrics?**

* **What the sources state:** The source pack does not state what specific evidence would be needed to prove a reduction in demand for licensed lyrics. It only notes that Judge Lee questioned the publishers' assertion that Claude threatens the market for their lyrics [L12], and that Anthropic's lawyer stated that AI has not reduced demand and that the publishers "could not identify any specific instances of market harm from Claude or other large language models" [L16]. 
* **Inference:** Based on Anthropic's critique that the publishers failed to identify "specific instances of market harm" [L16], the necessary evidence would likely need to go beyond usage volume (like the 120,000 requests for synthetic lyrics [L15]) and instead demonstrate concrete financial or licensing losses directly attributable to consumers or businesses using Claude's synthetic lyrics in place of the publishers' copyrighted lyrics.
* **Confidence:** high — It is certain that the source text does not outline evidentiary standards or requirements for proving reduced demand, beyond noting the absence of "specific instances."

***

**4. Inference: What assumption connects the nearly 120,000 requests for synthetic lyrics to the claim that this threatens the market for licensed lyrics?**

* **What the sources state:** Judge Lee questioned the publishers' claim that Claude's ability to compose lyrics threatens the market for their lyrics [L12]. The publishers' attorney argued that Anthropic misuses their lyrics to generate "competing content that harms human songwriters" [L14], citing that Claude received nearly 120,000 requests to generate "synthetic" lyrics in six months and arguing that LLMs "could overwhelm the market with AI-generated content very easily" [L15]. Anthropic countered that this has not reduced demand for the publishers' lyrics and no specific instances of market harm were identified [L16].
* **Inference:** The connecting assumption is that the "synthetic" lyrics requested by Claude users actually function as a commercial substitute ("competing content") for the publishers' existing lyrics or the services of human songwriters—meaning that users requesting AI lyrics would have otherwise paid for or licensed human-written lyrics, or that those 120,000 synthetic outputs are entering the commercial music market to compete for the same audience and revenue.
* **Confidence:** high — This inference directly bridges the explicit logical gap pointed out by Judge Lee and Anthropic's counsel between mere AI output volume and actual economic harm in the lyrics market.

Requested: xai/grok-4.7 · Returned: xai/grok-4.7

1. The publishers’ attorney, Matt Oppenheim, said Anthropic misuses their lyrics to generate competing content that harms human songwriters [L14]. He said Claude received nearly 120,000 requests to generate “synthetic” lyrics over a six-month period, and that large language models like Claude “could overwhelm the market with AI-generated content very easily” [L15]. The article also describes the publishers’ assertion, which Judge Lee questioned, that Claude’s ability to compose lyrics threatens the market for their lyrics [L12].

Inference: The publishers are characterizing harm mainly as present or near-term market substitution and displacement risk from competing AI lyrics, not as a measured drop in licensed-lyric demand already shown in the article.

Confidence: high — the characterization is stated directly in the hearing report.

2. The sources state that two Northern California judges split on whether generative AI training is generally fair use [L20]. Judge Alsup sided with Anthropic on training using books; two days later Judge Chhabria ruled for Meta in a similar case but warned that AI training “in many circumstances” would not be fair use [L21]. In this lyrics case, Judge Lee questioned whether Claude’s lyric-writing threatens the publishers’ market [L12]. Anthropic’s lawyer said AI has not reduced demand for the lyrics, that the publishers could not identify any specific instances of market harm from Claude or other models [L16], and that UMG has called AI “overwhelmingly a net positive” for the music industry [L17]. The publishers pointed to competing synthetic lyrics, nearly 120,000 such requests in six months, and a risk of overwhelming the market [L14–L15].

Inference: The market-harm differences most likely to matter to Judge Lee are whether harm is shown by specific instances or only by a risk that models “could” overwhelm the market; whether demand for licensed lyrics has actually fallen; whether synthetic lyrics compete with and substitute for the publishers’ lyrics rather than only supporting a claimed transformative use; and how to weigh UMG’s own “net positive” statement against the request volume.

Confidence: medium — those contrasts are in the pack, but the article does not say which facts Judge Lee will treat as decisive.

3. The sources do not say what evidence would be needed to show whether Claude has reduced demand for licensed lyrics. They report only that Anthropic argued AI has not reduced that demand and that the publishers could not identify any specific instances of market harm from Claude or other large language models [L16], while the publishers asserted competing synthetic lyrics, nearly 120,000 generation requests in six months, and a risk of overwhelming the market [L14–L15].

Inference: Evidence that would speak to the dispute as framed would need to connect Claude outputs or those requests to actual changes in licensing, sales, or other demand for the publishers’ lyrics, rather than request volume alone. The pack does not define the required proof.

Confidence: high — the gap in the sources is clear.

4. The sources state that Oppenheim cited nearly 120,000 requests for “synthetic” lyrics over six months and said models like Claude “could overwhelm the market with AI-generated content very easily” [L15], in support of the claim that Claude’s lyric composition threatens the market for the publishers’ lyrics [L12] and generates competing content that harms human songwriters [L14]. Anthropic disputed actual demand reduction and specific market-harm instances [L16].

Inference: The connecting assumption is that a large volume of requests for synthetic lyrics means those outputs compete with, and can displace demand for, licensed human lyrics in the same market—so request volume itself is evidence of market threat even without identified lost sales or licenses.

Confidence: medium — the link is a fair reading of the publishers’ argument, but the article does not spell out that assumption.

Requested: openrouter/meta/muse-spark-1.3 · Returned: meta/muse-spark-1.3

1. From the source: How did the publishers characterize the market harm risk from AI-generated lyrics?

What the sources state: The publishers' attorney Matt Oppenheim said Anthropic "misuses its lyrics to generate competing content that harms human songwriters" [L14]. He said Claude received "nearly 120,000 requests to generate 'synthetic' lyrics over a six-month period" and said large language models like Claude "could overwhelm the market with AI-generated content very easily" [L15]. Judge Lee is described as questioning "the publishers' assertion that Claude's ability to compose lyrics threatens the market for the publishers' lyrics" [L12].

Inference: The publishers frame the risk as future market flooding by competing AI output, not as a past loss already quantified in the article.

Confidence: high because the characterization is directly quoted in [L14]-[L15].

2. Inference: Given that prior rulings split on AI training fair use, which factual differences about market harm could matter most to Judge Lee?

What the sources state: Two other Northern California judges "have split on whether generative AI training generally makes fair use of copyrighted material" [L20]; now-retired Judge William Alsup "sided with Anthropic last year" on books, while two days later Judge Vince Chhabria "ruled for Meta in a similar lawsuit but warned that AI training 'in many circumstances' would not be fair use" [L21]. In this case, Judge Eumi Lee "did not rule at a hearing in San Jose but pressed both sides on the fair use question" [L11] and "questioned the publishers' assertion that Claude's ability to compose lyrics threatens the market for the publishers' lyrics" [L12]. Oppenheim pointed to 120,000 synthetic-lyrics requests and risk of overwhelming the market [L15], while Anthropic lawyer Sonal Mehta "countered that AI has not reduced demand for the lyrics, and that the publishers could not identify any specific instances of market harm" [L16] and said UMG itself has called AI "overwhelmingly a net positive" [L17].

Inference: The market-harm facts most likely to matter to Lee are whether publishers can show specific, realized harm/reduced demand and substitution by competing synthetic lyrics, versus only high request volume and potential future overwhelm with no identified instances.

Confidence: medium because Lee's questioning is reported but the sources do not state what test or facts she will find decisive.

3. Open question: What evidence would be needed to show whether Claude has reduced demand for licensed lyrics?

What the sources state: Mehta claimed "AI has not reduced demand for the lyrics, and that the publishers could not identify any specific instances of market harm from Claude or other large language models" [L16]. Oppenheim cited "nearly 120,000 requests to generate 'synthetic' lyrics over a six-month period" [L15]. The sources do not answer what evidence would be needed to prove reduced demand; no methodology, data source, or evidentiary standard is given.

Inference: Such evidence would likely need before-and-after licensing/sales data for the lyrics at issue and proof linking any decline to substitution by Claude output rather than other factors.

Confidence: low because the pack provides no evidentiary framework and any list is speculation beyond [L15]-[L16].

4. Inference: What assumption connects the nearly 120,000 requests for synthetic lyrics to the claim that this threatens the market for licensed lyrics?

What the sources state: Oppenheim said Claude received "nearly 120,000 requests to generate 'synthetic' lyrics over a six-month period" and that models like Claude "could overwhelm the market with AI-generated content very easily" [L15], after arguing Anthropic "misuses its lyrics to generate competing content that harms human songwriters" [L14]. The sources do not explicitly state the assumption linking request volume to market threat.

Inference: The link assumes synthetic-lyrics requests substitute for uses that would otherwise require licensed lyrics in the same market, so volume equals displacement/competition rather than separate non-competing use.

Confidence: medium because the competing-content and overwhelm language implies substitution but the sources never spell out that premise.

Requested: openrouter/qwen/qwen3.8-max-0902 · Returned: qwen/qwen3.8-max-0902

1. **From the source: How did the publishers characterize the market harm risk from AI-generated lyrics?**  
   Sources state: The publishers asserted that Claude’s ability to compose lyrics threatens the market for the publishers’ lyrics [L12]. Their attorney said Anthropic misuses lyrics to generate competing content that harms human songwriters [L14]. They also pointed to nearly 120,000 requests for “synthetic” lyrics over six months and said large language models like Claude “could overwhelm the market with AI-generated content very easily” [L15].  
   Inference: The publishers frame the risk as substitution and market flooding: AI-generated lyrics could compete with, displace, or overwhelm the market for publishers’ lyrics and harm human songwriters.  
   Confidence: high because the source directly states the publishers’ market-harm arguments, though the article is owner-pasted and not independently verified [L8].

2. **Inference: Given that prior rulings split on AI training fair use, which factual differences about market harm could matter most to Judge Lee?**  
   Sources state: Prior Northern California judges have split on whether generative AI training is fair use [L20]. One judge sided with Anthropic on fair use for books, while another ruled for Meta but warned that AI training “in many circumstances” would not be fair use [L21]. In the current case, the publishers claim Claude’s lyric-generation threatens the market for lyrics [L12], cite competing content and harm to human songwriters [L14], and cite nearly 120,000 synthetic-lyric requests as evidence that AI could overwhelm the market [L15]. Anthropic counters that AI has not reduced demand for lyrics and that the publishers cannot identify specific instances of market harm [L16], and adds that UMG has called AI “overwhelmingly a net positive” for the music industry [L17].  
   Inference: The factual differences that could matter most are whether the publishers can show specific, concrete market harm or reduced demand tied to Claude’s lyric generation, especially at scale, versus Anthropic’s position that no specific instances of harm have been identified and that AI may be net positive. The volume of synthetic-lyric requests, whether those outputs compete with publishers’ lyrics, and any evidence of actual substitution seem likely to be important.  
   Confidence: medium because the sources identify the disputed market-harm facts but do not state which facts Judge Lee will treat as decisive.

3. **Open question: What evidence would be needed to show whether Claude has reduced demand for licensed lyrics?**  
   Sources state: The sources do not directly answer what evidence would be needed. They state that Anthropic argued AI has not reduced demand for lyrics and that the publishers could not identify specific instances of market harm from Claude or other large language models [L16]. The publishers, however, assert that Claude’s ability to compose lyrics threatens the market for publishers’ lyrics [L12] and cite nearly 120,000 requests for synthetic lyrics [L15]. The source pack does not specifically mention “licensed” lyrics.  
   Inference: Evidence would likely need to connect Claude use to reduced demand for the publishers’ lyrics, such as specific examples where users or customers used Claude-generated lyrics instead of obtaining publishers’ lyrics, or data showing a decline in demand linked to Claude.  
   Confidence: low because the source pack does not specify the needed evidence or mention licensing evidence directly.

4. **Inference: What assumption connects the nearly 120,000 requests for synthetic lyrics to the claim that this threatens the market for licensed lyrics?**  
   Sources state: The publishers’ attorney said Claude received nearly 120,000 requests to generate “synthetic” lyrics over six months and argued that large language models like Claude could overwhelm the market with AI-generated content very easily [L15]. He also said Anthropic misuses lyrics to generate competing content that harms human songwriters [L14]. The publishers assert that Claude’s ability to compose lyrics threatens the market for publishers’ lyrics [L12].  
   Inference: The assumption is that requests for synthetic lyrics represent actual or potential substitutes for the publishers’ lyrics, so that enough AI-generated lyric use could reduce demand for, or overwhelm, the market for publishers’ lyrics.  
   Confidence: high because the source directly ties the large number of synthetic-lyric requests to the publishers’ claim of possible market overwhelm, though the article is owner-pasted [L8].