Generative AI and Copyright Infringement: A Legal-Technical Analysis of AI Music Generation Systems Under 17 U.S.C. Title 17

Generative AI music tools (such as Google Gemini’s Lyria or other multi-stage pipelines) allow users to create songs by supplying prompts that include copyrighted lyrics and instructions to emulate a particular singer’s voice. This raises complex questions under U.S. copyright law [2]. This review finds that:

• •

  Lyrics and Composition Rights: If a user inputs copyrighted lyrics (or even large portions) into a GenAI system and generates a new song, the output likely infringes the musical composition. Lyrics are literary works protected as part of the composition (17 U.S.C. §102(a)(2)) [2], so copying them without permission or license violates the composer’s exclusive reproduction and derivative-work rights [3]. Even if the AI creates a new melody, unauthorized lyric reproduction is prima facie infringement absent fair use [5].

• •

  Voice Cloning vs. Sound Recordings: AI-generated "sound-alike" vocals typically do not infringe sound recordings unless actual fixed samples were copied. Section 114(b) [4] explicitly permits independent re-recordings or imitations of a vocal style. The recent decision in Richardson v. Kharbouch (7th Cir. 2025) held that sounding like French Montana without sampling his track was not infringement [13]. Likewise, Lehrman v. Lovo (S.D.N.Y. 2025) found no federal infringement when an AI cloned professional voices for training text-to-speech [12].

• •

  Publicity and State Law: Because federal copyright often fails to cover AI voice cloning, injured performers must rely on state rights of publicity or unfair competition [16, 17]. For example, Tennessee’s 2024 "ELVIS Act" and similar statutes recognize control over one’s voice and image when used without consent [20]. These state-law claims are independent of copyright [22].

• •

  Commercial Publication: Uploading AI-generated songs to streaming services or social media (with ads or subscriptions) creates direct evidence of exploitation. Owners of copyrighted works (lyrics, composition) can claim statutory damages (up to $150k/work) and profits under §504 [8] if registration is timely, plus injunctions [9]. Monetization increases remedies.

• •

  Case Law Trends: Recent cases illustrate this split: In Concord Music Group v. Anthropic, publishers sued over Claude outputting their lyrics [10]; Kadrey v. Meta addressed fair use in training [11]; Lehrman v. Lovo involved AI voice clones [12]; UMG v. Uncharted Labs challenges AI music streaming/ingestion [14]. Most outcomes favor copyright owners on lyric copying claims, and favor defendants on sound-alike voice claims [13, 15].

• •

  Regulatory Gap and Policy: The asymmetry between textual lyrics and vocal identity highlights a gap [19]. Current Title 17 protections robustly cover lyrics and compositions, but do not clearly cover persona or voice. We recommend policy reforms: extend right of publicity protections nationwide or introduce a federal performer right against unauthorized AI voice cloning, and clarify licensing/notice rules for AI music training and output [23, 18].

This report merges legal analysis with technical insights on GenAI music pipelines (transformers, diffusion models, vocoders, embeddings) and includes tables mapping key cases and technical elements to legal risks, as well as visual charts of case timelines and outcome distributions.

Modern generative AI music tools (e.g., Google’s Gemini Music, Meta’s MusicGen, MusicLM, Suno, Udio, Stability AI’s audio models, etc.) break song production into discrete technical modules [21]. A user might provide a text prompt containing the lyrics to a known song and stylistic instructions (e.g., "in the style of [famous artist]") and the AI system will process the prompt through multiple layers. First, natural language processing (NLP) encodes the lyric text and style instructions (often via a BERT-like transformer) into an internal embedding. Concurrently, AI modules generate melody and harmony (often using music-trained transformers or diffusion networks). Finally, a neural vocoder synthesizes the audio waveform, injecting a target vocalist’s characteristics via speaker embeddings or voice-conversion algorithms (e.g., VALL-E, RVC, or other zero-shot voice cloning tech). The result can be a fully produced song with instrumental tracks and a vocal track that sounds eerily like a well-known singer [19].

From the user’s perspective, the process is nearly frictionless: the user supplies lyrics (potentially copyrighted), chooses vocal timbre ("like [artist]"), perhaps adds percussion or genre tags, and outputs a song. If the user then uploads that song to platforms (YouTube, Spotify, TikTok, etc.) and monetizes it, multiple legal issues arise at different stages: Was copyrighted material used during model training? Did the user’s prompt or the output copy protected expression? Is the synthesized singer’s voice a protected persona or just a non-copyrightable sound imitation? This paper examines these issues under U.S. law.

We focus on 17 U.S.C. §106’s exclusive rights (reproduction, derivative works, distribution) for musical compositions and sound recordings [3]. We also consider §114(b) [4], which narrows sound recording rights, and note that voice-likeness claims often fall to state publicity laws (as in Tennessee’s ELVIS Act [20]). Our analysis is both legal (case law, statutes) and technical (identifying how AI components map to rights). We illustrate with recent litigation and propose remedies and policy reforms.

Early AI image and text cases have analogues in music. The literature distinguishes three layers: (1) Training-layer copying, where large datasets of copyrighted music (songs, lyrics, recordings) may be ingested into models; (2) Output-layer reproduction, where AI outputs may inadvertently or deliberately recreate protected works; and (3) Commercial dissemination, where outputs reach the market and may cause market harm [27, 28].

In music specifically, scholarly attention is nascent. Some authors highlight that musical compositions (melody, harmony, lyrics) are separate copyrights from sound recordings, complicating analysis [29]. AI outputs can blend these elements. The U.S. Copyright Office has noted that music AI raises novel questions about authorship and infringement [2].

Several law review and industry pieces have surfaced post-2022. For example, a Brooklyn Law School review discusses composers’ concerns about AI using sheet music and lyrics [11]. Another analysis notes that while AI art might not itself be copyrightable absent human authorship, the process still "materially affects" markets for original works [30]. On voice cloning, scholars note the "publicity gap": federal copyright (even with new owners like performing rights societies) hasn’t addressed the value in a performer’s unique voice, leading to calls for expanded rights [22, 23].

Prominent cases and industry actions have driven recent commentary. The RIAA publicly denounced unlicensed AI "cover" platforms like Suno and Udio, and UMG recently sued them for allegedly stream-ripping copyrighted recordings as training data [5, 14]. Law firms and news outlets have reported on publisher suits against AI companies for lyric output (Concord v. Anthropic [10], and a companion suit by Universal Music Group). Articles also discuss individual lawsuits: e.g., NPR’s former host David Greene sued Google (2024) for using his voice in the NotebookLM AI tool, and various voice actors have sued startups like Lovo for creating synthetic clones of their voices [12]. (See Section IV.C.)

Academic treatments include proposals to adapt Copyright Act factors to AI training data (fair use’s fourth factor of market effect looms large [11]). Others suggest legislative solutions: California and Tennessee have enacted laws protecting biometric data including voice (e.g. Tennessee S.B. 1921, the "ELVIS Act" effective July 2024, explicitly targeting AI voice cloning [7, 20]).

We analyze the hypothetical workflow under Title 17:

AI music pipelines include [19, 21]:

• •

  User Prompt: The text entry (lyrics, style). If this contains copyrighted lyrics, it triggers reproduction right violation. Users should avoid pasting protected text into prompts. (Mitigation: prompt only original or public-domain lyrics).

• •

  Embedding and Conditioning: The prompt is tokenized by an NLP model (e.g., BERT or GPT variants) into embeddings $c$ that condition downstream generation. Technically, $c$ enters the diffusion model as side information ($p_{\theta}(x_{t-1}|x_{t},c)$). From a legal view, the embedding itself is non-human-readable, but if $c$ contains unauthorized lyrics, the model’s output may reconstruct them. (Risk: reproduction of lyric content. Mitigation: filter training data/embeddings, flag protected lyrics.)

• •

  Melody/Harmony Generation: A music model (transformer or diffusion) generates new audio content. For example, a latent diffusion model iteratively denoises a random tensor $x_{T}\sim\mathcal{N}(0,I)$ to $x_{0}$ conditioned on $c$ (the prompt embedding). While this process is "generative," output can still infringe if it replicates an original composition’s melody too closely [25]. (Risk: derivative composition. Mitigation: adjust temperature or randomness, use licensing for any borrowed motifs.)

• •

  Neural Vocoder: Converts intermediate representations (e.g., spectrograms) to final waveform. If trained on artists’ voices, it synthesizes realistic vocals. Voice cloning often uses a speaker embedding vector $s$ extracted from a sample; then the vocoder generates a novel utterance with characteristics $s$. Because the waveform is original, copyright typically isn’t infringed [12, 13]. (Risk: appropriation of performer identity. Mitigation: require voice consent, watermark AI voices, treat voice as publicity issue.)

• •

  Output Distribution: Uploading the song to platforms. Technically, the audio file and metadata (title, lyrics, creator) are stored. This action triggers reproduction and distribution rights in §106(1) and (3) [3]. The platform may face secondary liability if it "knowingly" facilitates infringement (cf. §512 safe harbor rules). (Risk: infringement at scale. Mitigation: content ID filters on lyrics or audio; opt-in licensing agreements; takedown procedures.)

• •

  Commercialization: Monetization (ads, subscriptions, streaming revenue) creates evidence of profit from the infringing output. For damages (§504) [8], courts look at actual profits attributable to infringement and the infringer’s gross receipts. (Risk: increased statutory damages for commercial use. Mitigation: apply fair-use analysis or revenue sharing; secure licenses upfront.)

Table I summarizes key AI components, their legal risks, and possible mitigations.

The emerging AI music ecosystem shows a structural enforcement asymmetry. Traditional copyright doctrines align well with human-authored music: they protect lyrics and melody explicitly [2, 3]. AI complicates things by mechanizing composition. Yet output that literally reuses lyrics triggers the same legal analysis as any cover song done without permission [25]. Early cases suggest copyright owners will press these conventional claims vigorously [10, 11].

By contrast, AI-blurred sound recordings and performances escape Title 17 for now. Synthetic vocals are "new" sounds that mimic old ones. Section 114(b) [4] expressly allows imitation of the sound of another voice, as long as no actual audio bits were copied. While owners can (and have) argued that voice is something they should own, courts to date (like in Phillips v. SkyPortal and Lehrman v. Lovo) have refused to stretch 17 U.S.C. §114 that far [15, 12]. This creates a perverse incentive: a user may think "as long as I don’t sample, I’m safe," even if the output clearly trades on an artist’s persona [19].

The upshot is a legal gap. If XYZ’s AI sings a Frank Sinatra lyric convincingly, Sinatra’s heirs may have no federal remedy (since Sinatra’s estate likely owns the composition, not a particular recording). They could attempt a publicity suit (as Tony Bennett’s estate did against LMFAO for a nightclub ad that caricatured Bennett’s voice) [16, 17], but such claims are costly and uncertain [22]. Meanwhile, Swift’s publisher can sue for lyrics, giving one set of rights steam, and Sinatra’s estate another, but these are separate tracks.

From a policy standpoint, this imbalance is worrying. It signals that AI creators can exploit performer identity via synthetic clones with less legal risk than copying a lyric or melody [23]. Laws like Tennessee’s ELVIS Act [20] (now joined by similar legislation in California and Washington) partially address this by creating sui generis rights. But a patchwork of state laws is imperfect [26]. There have been proposals for a federal ‘Singing Voice Protection Act’ or expanding sound recording rights to include persona-based protections [18]. Another proposal is mandatory labeling of AI-generated music so consumers and artists know when a voice is synthetic (enabling transparency and opt-in arrangements) [19].

One concrete step is to apply existing tools more aggressively: platforms could scan uploads for known lyrics (text fingerprinting) and for voiceprints (audio fingerprinting against sound recording databases) and flag potential infringement. Licensing regimes could be created for AI models (analogous to mechanical licenses for covers) whereby using a particular artist’s voice or lyrics would require a fee [25, 21].

To harmonize protection of lyrics and vocal identity in the AI era, we suggest:

• •

  Extend Federal Rights or Harmonize States: Consider a federal right of public performance or identity for vocal likeness [22, 23]. If infeasible, push for uniform state publicity laws (like an updated right-of-publicity statute that expressly covers generative AI) [20, 26].

• •

  Training Data Licensing: Require AI developers to use licensed datasets or clearly fall under safe harbor (compulsory licensing for training data could be explored) [11, 18].

• •

  Content Identification: Encourage the development of AI-detection tools and content-ID systems for music that flag copyrighted inputs or outputs [19].

• •

  Transparency Labels: Mandate that AI-generated music include metadata disclosing AI involvement, aiding rights holders in enforcement and informing consumers [21].

• •

  Copyright Office Guidelines: The USCO should clarify registration practices for AI-generated music (as it did in 2022). Current guidance requires "meaningful human authorship," so users need to explicitly assert which parts (e.g. lyrics) were human-authored [2, 25].

• •

  Legislative Clarification: Amend 17 U.S.C. to address generative AI specifically (e.g., safe harbor for licensed training, rights in "non-fixation" of voice patterns, etc.) [18, 23].

As GenAI music tools like Google Gemini become mainstream, both creators and consumers must navigate a complex legal landscape. Our analysis finds that entering copyrighted lyrics into an AI prompt, then publishing and monetizing the AI-produced song, likely infringes the song’s composition rights under Title 17 [3, 10, 25]. By contrast, synthesizing a singer’s voice without copying their recording tends not to trigger federal copyright claims, shifting the dispute to state publicity law [12, 13, 16, 17]. This split means lyric owners (composers/publishers) have stronger remedies than performers when their vocal likeness is AI-cloned [22].

The contrast underscores a regulatory gap: the law covers words and notes, but treats the "sound of a voice" as unprotected sound [19, 23]. Unless addressed by courts or legislatures, AI music generation will continue to exploit this asymmetry. In the meantime, users should err on the side of caution: assume that copying lyrics or melodies without permission is infringement, and that generating a famous voice without authorization, while not a copyright crime, may still be illegal under other laws or contracts [20, 26]. Platforms and policymakers should update practices and laws to ensure creators’ rights are balanced with innovation in this new AI-driven age.

The following tables summarize key cases and the mapping of technical components to legal issues. Figures include charts for timeline and outcome distributions.