The Attribution Revolution: Artists, AI, and the Future of Getting Paid

The Attribution Revolution: Artists, AI, and the Future of Getting Paid — Ai4

Ai4

AI-assisted summary of The Attribution Revolution: Artists, AI, and the Future of Getting Paid — original source. Check the original for full context.

Source published:

Detailed notes

## Background Published by the Ai4 channel on YouTube on August 22, 2026, this panel features Matt Adell (Co-Founder and COO of Musical AI) alongside Kristen Agee (Founder/CEO) and Xena Plat (COO) of 411 Music Group. The presentation explores how generative AI is reshaping the economics of music, detailing rights-holder consent strategies, attribution technology that functions as a modern performing rights organization, and the future of personalized sync licensing across media.

## TL;DR Opening with legal precedents like GEMA's copyright victory over Suno, the panel dispels the misconception that commercial AI licenses protect enterprise users from underlying training data liability. 411 Music Group and Musical AI lay out the CAM framework—Consent, Attribution, and Monetization—highlighting how granular contributor tracking and influence-percentage algorithms route upstream subscription pools and downstream performance royalties back to creators. Supported by real-world implementations like their SOCAN partnership, the speakers argue that AI will not displace human artistry, but will instead unlock dynamic, real-time personalized audio across gaming and digital media.

## Atmosphere & style The session strikes a pragmatic, authoritative, and clear-eyed tone, cutting through speculative AI hype and apocalyptic industry anxieties with concrete legal and financial architectures. The rapport among Agee, Plat, and Adell is grounded and collaborative, blending the operational frontline experience of catalog administrators with the technical vision of rights-tech builders. The dialogue moves purposefully from the sharp financial risks of statutory infringement toward an energizing, practically optimistic blueprint for the future of creator compensation.

## Key takeaways - [00:04](https://www.youtube.com/watch?v=We9POs5uaho&t=4s) A German court ruling against Suno established that AI platforms face copyright infringement liability and statutory damages when training on protected works without authorization. - [00:04](https://www.youtube.com/watch?v=We9POs5uaho&t=4s) Commercial terms of service granted by AI vendors do not protect downstream enterprise users from third-party copyright claims regarding underlying training data. - [00:04](https://www.youtube.com/watch?v=We9POs5uaho&t=4s) Responsible generative AI integration across all creative industries depends on three foundational pillars: Consent, Attribution, and Remuneration (CAM). - [02:07](https://www.youtube.com/watch?v=We9POs5uaho&t=127s) Informed AI data licensing requires unbundling master and composition rights, demanding granular consent tracking per contributor, right, and use-case modality. - [02:07](https://www.youtube.com/watch?v=We9POs5uaho&t=127s) Flat per-asset ingestion fees disproportionately benefit deep catalog owners, whereas independent creators need recurring downstream attribution upside to justify opting in. - [05:49](https://www.youtube.com/watch?v=We9POs5uaho&t=349s) Attribution technology can calculate the exact percentage influence of specific training tracks on a generated output, providing an objective basis for royalty distribution. - [09:05](https://www.youtube.com/watch?v=We9POs5uaho&t=545s) Mass-scraping copyrighted catalogs exposes AI developers to statutory damages up to $150,000 per work, posing systemic multi-billion-dollar liabilities for tech companies. - [09:05](https://www.youtube.com/watch?v=We9POs5uaho&t=545s) The tech industry must shift from valuing training data based on raw volume to measuring the proprietary value and stylistic influence assets contribute to outputs. - [13:08](https://www.youtube.com/watch?v=We9POs5uaho&t=788s) AI creator compensation relies on two distinct streams: an upstream SaaS subscription pool (~70% payout model) and downstream commercial sync collections. - [15:06](https://www.youtube.com/watch?v=We9POs5uaho&t=906s) Performing rights organizations like SOCAN are integrating AI attribution engines directly into legacy collection rails to track and remit royalties at global scale. - [16:10](https://www.youtube.com/watch?v=We9POs5uaho&t=970s) Generative music's immediate commercial opportunity centers on real-time dynamic personalization for gaming and digital ads to boost retention, rather than displacing human composers. - [17:57](https://www.youtube.com/watch?v=We9POs5uaho&t=1077s) The distinction between AI and traditional media will soon disappear, turning generative technology from an existential threat into an economic engine via structured attribution rails.

## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=We9POs5uaho&t=4s) Copyright Lawsuits and the Commercial AI Illusion

Kristen Agee opens the discussion by highlighting a recent legal milestone: a German court ruled against the generative AI music platform Suno in an infringement lawsuit brought by the German performance rights organization GEMA. The court found that Suno had utilized protected musical works without authorization, ordering the platform to disclose its related profits, pay damages, and immediately cease certain unauthorized uses. Agee addresses why professionals across the broader artificial intelligence sector must pay close attention to music litigation, warning against a dangerous industry-wide misconception. Many enterprise customers and developers mistakenly assume that if an AI platform contractually grants permission for commercial use of its outputs, those assets are automatically legally cleared to broadcast, monetize, or integrate into commercial software.

In reality, a vendor's commercial terms of service do not resolve whether the underlying foundational model was trained on properly licensed data, whether generated outputs reproduce protected elements, or whether third-party rights holders will pursue copyright claims. Music is simply the earliest battleground for an issue that directly impacts visual art, video, journalism, publishing, film, proprietary code, and enterprise business data. The central dilemma facing every generative AI company is whether they can build and monetize products using proprietary human work without obtaining explicit prior permission, attributing original creators, or sharing the derived revenue. Agee frames the panel around three interconnected foundational pillars: consent, attribution, and remuneration.

- **Legal Precedent**: GEMA successfully secured a ruling against Suno establishing copyright infringement, profit disclosure mandates, and statutory damage liabilities. - **Enterprise Risk**: Contractual commercial licenses granted by AI platforms do not indemnify or protect downstream users against third-party training data infringement claims. - **The Core Framework**: Responsible and legal AI integration hinges on three pillars: Consent (input governance), Attribution (influence mapping), and Remuneration (value distribution).

[02:07](https://www.youtube.com/watch?v=We9POs5uaho&t=127s) Securing Informed Artist Consent and the Catalog Divide

Xena Plat details how 411 Music Group approached data licensing across its roster of over 5,000 artists and 120 global record labels. Obtaining valid consent is not a blanket exercise; music copyright requires unbundling master recording and composition rights, which almost always involve separate co-writers, publishers, producers, and labels. Because no single entity can authorize use on behalf of all contributors, consent decisions must be tracked per contributor, per right, and per specific use case. Before gathering permissions, 411 organized dedicated town halls—deliberately separating record labels and artists into different rooms—to educate rights holders on the mechanics of AI data monetization without pushing a sales pitch.

```mermaid flowchart TD accTitle: Music AI Licensing and Consent Workflow accDescription: Flowchart illustrating the rights holder consent and licensing process at 411 Music Group.

subgraph RightsHolders["Rights Holders Roster"] L["120+ Record Labels<br/>(Deep Catalogs: 100+ tracks)"] A["5,000+ Indie Artists<br/>(Small Catalogs: 5–10 tracks)"] end

subgraph Governance["Consent & Education Layer"] TH["Segregated Town Halls<br/>(Labels vs. Artists)"] Split["Per-Contributor Tracking<br/>• Master Rights<br/>• Composition Rights"] end

subgraph DealTerms["Licensing Parameters"] Scope["Use Case Scope<br/>• Generative creation<br/>• Assistive DAW production"] Terms["Commercial Terms<br/>• Flat upfront fee per asset<br/>• Subscription backend share<br/>• Term: 1-yr / 5-yr / Perpetual"] end

L --> TH A --> TH TH --> Split Split --> Scope Scope --> Terms ```

This education process revealed a sharp split across the creator community. Labels and legacy artists holding deep catalogs of 100 to several thousand tracks were eager to opt in immediately, as existing AI licensing deals compensate rights holders with flat fees per ingested asset, creating substantial upfront windfalls. Conversely, smaller independent creators—including Plat herself, who has a boutique catalog of a few five-track albums—found little financial incentive in flat per-track data fees. However, Plat notes that evolving attribution frameworks may soon shift the math for indie artists by offering recurring downstream upside. For creators to provide genuine informed consent, AI platforms must disclose granular deal parameters: - **Model Modality**: Clarifying whether training data powers generative replacement engines or non-generative, assistive tools that support composer production workflows. - **Remuneration Structure**: Defining whether compensation is strictly an upfront flat ingestion fee or includes recurring equity/backend shares of platform subscriptions. - **Licensing Duration**: Establishing strict term limits (e.g., one-year or five-year licenses) versus permanent, irrevocable perpetual training rights.

[05:49](https://www.youtube.com/watch?v=We9POs5uaho&t=349s) AI Attribution Technology and Revenue Distribution

Matt Adell explains that while consent governs what enters an AI model, attribution determines how value is measured once millions of data points are blended to generate new outputs. Music has historically served as the vanguard for establishing digital copyright frameworks, and the economic mechanisms forged in music will dictate how all copyrighted media is treated across the web. AI companies desperately require not just vast volumes of data, but high-quality curated data to make their models commercially viable. Once models are operational, they produce two distinct revenue streams: recurring subscription fees paid by end users, and downstream commercial revenue generated when outputs are distributed onto streaming platforms like Spotify or licensed into television commercials.

To align these revenue streams with rights holders, Musical AI developed technology that calculates the exact percentage influence of every piece of training data on any generated output. Adell illustrates this granularity: for any given generated song, the system can determine that the output was "10% influenced by this Drake song, 3% influenced by this Beatles song, [and] 0.001% influenced by that punk rock band from Peoria that you've never heard of." This attribution report establishes the objective basis for dividing downstream income, mirroring the performance rights organization (PRO) model where higher stream counts correlate to higher royalties. Rather than treating training data as an undifferentiated monolith, attribution ensures that the specific works driving generative quality receive an equitable share of every dollar earned.

[09:05](https://www.youtube.com/watch?v=We9POs5uaho&t=545s) Statutory Infringement Risks and Cross-Industry Value Attribution

Addressing the legal exposure facing generative AI companies, Adell warns against common shortcuts like mass-scraping copyrighted music from YouTube. He notes that unauthorized ingestion triggers severe liability: under United States copyright law, statutory damages reach up to $150,000 for every individual copyrighted work copied without authorization. If an AI enterprise duplicates and stores multiple copies of entire global music libraries across training clusters, potential statutory liabilities quickly escalate into tens of billions of dollars, introducing catastrophic systemic risk to the tech sector's valuation. Adell dismisses the common developer defense that clearing global music catalogs is impossible, noting that streaming platforms like Spotify have successfully cleared worldwide rights for decades: "It is totally possible to license all the music in the world."

```mermaid flowchart LR accTitle: Generative AI Attribution and Revenue Routing Flow accDescription: Flowchart demonstrating how Musical AI calculates influence percentages and routes subscription and commercial revenues back to rights holders.

subgraph Ingestion["Input & Training"] CD["Consented Data Catalog<br/>(411 / Sony / UMG)"] Model["Generative AI Model"] end

subgraph Measurement["Attribution Engine"] Attr["Musical AI Attribution<br/>• 10% Drake track<br/>• 3% Beatles track<br/>• 0.001% Indie punk band"] end

subgraph Revenue["Monetization & Payout"] SubPool["Platform Subscriptions"] SyncPool["Commercial & DSP Royalties"] Payout["Dynamic Payout to Rights Holders"] end

CD --> Model Model --> Attr Attr --> SubPool Attr --> SyncPool SubPool --> Payout SyncPool --> Payout ```

Adell emphasizes that mathematical attribution applies far beyond audio. While it struggles with widely repeated factual statements (such as historical news events cited across thousands of journalistic articles), it maps cleanly to any proprietary or creative asset where stylistic or compositional influence can be calculated—including film, television, poetry (such as Shel Silverstein's style), visual character design (such as a *Star Wars* stormtrooper), and complex pharmaceutical research data. He argues that the tech industry must abandon brute-force volume metrics in favor of value-driven attribution, delivering a blunt critique of standard data ingestion metrics:

> "We measure manure in volume. We shouldn't measure other human output besides manure in volume. We should measure it based on the value it creates, and that can be done on any copyrighted or proprietary data."

[13:08](https://www.youtube.com/watch?v=We9POs5uaho&t=788s) The Two Revenue Pools: Subscriptions and Downstream Outputs

The financial architecture for compensating human creators in the generative AI era splits into two distinct revenue pools: upstream subscription revenue and downstream commercial exploitation. The first pool encompasses software-as-a-service (SaaS) subscription and usage fees. When a subscriber pays an AI platform to generate music, video, or specialized data, that AI company allocates a percentage of its top-line revenue into a central rights pool, which is then divvied up based on the proportional value of the human assets used. This framework mirrors the existing streaming economy, where platforms like Spotify pay roughly 70% of gross revenues into a collective pool distributed to rights holders based on stream counts.

```mermaid flowchart TD accTitle: Generative AI Revenue and Attribution Flow subgraph Upstream["Upstream: Generation Phase"] A["User Subscriptions & SaaS Fees"] --> B["AI Subscription Revenue Pool\n(~70% Model)"] B --> C["Attribution & Value Measurement Engine"] end

subgraph Downstream["Downstream: Commercial Exploitation"] D["Commercial Output\n(Films, TV, Ads, Games, Streaming)"] --> E["Collecting Societies & PROs\n(e.g., SOCAN)"] E --> C end

subgraph Payout["Distribution"] C --> F["Royalties to Songwriters, Performers & Rights Holders"] end ```

The second, more complex revenue stream governs downstream assets once AI-generated outputs enter the wider commercial ecosystem. When a generated work is monetized on external platforms—such as a track streaming on Spotify, a video on YouTube, or a sync placement in a major feature film—revenue must be captured at the source ("first dollar"). Rather than inventing an entirely new licensing framework, this relies on established collecting societies and performing rights organizations (PROs). In the same way radio broadcast revenues have long been pooled and remitted back to songwriters and performers, AI attribution engines supply the granular tracking data necessary for societies to calculate and distribute payouts to millions of human creators at scale.

- **Upstream SaaS Pool:** AI platforms pay a predetermined percentage of recurring subscription/generation revenue into an initial rights holder pool, echoing Spotify's ~70% payout structure. - **Downstream Sync & Performance Pool:** Generated assets generating money across third-party platforms (streaming services, broadcast television, cinema) are tracked and collected by existing rights societies using attribution reporting.

---

[15:06](https://www.youtube.com/watch?v=We9POs5uaho&t=906s) Scalability, SOCAN Partnership, and Historical Precedent

To demonstrate that attribution technology can operate across global music catalogs, Musical AI announced a formal partnership with SOCAN (the Society of Composers, Authors and Music Publishers of Canada), one of the world's largest music rights management organizations. SOCAN has adopted Musical AI's attribution technology to process reporting data across AI-generated content, allowing the society to determine the exact distribution shares owed to individual songwriters and performers whenever monetized AI assets are identified.

This level of tracking represents an evolution of established industry precedent rather than an untested experiment. Scalable collective licensing existed long before digital computers: during the big-band era, collective systems tracked live sheet-music performances so that when an entertainer like Jack Benny performed a Henry Mancini composition in a Las Vegas nightclub, the original composer received their rightful performance royalties.

- **SOCAN Integration:** Canada's primary PRO has integrated Musical AI's attribution reporting directly into its royalty collection and distribution pipeline. - **Historical Precedent:** The model mirrors legacy PRO performance tracking that dates back to the sheet-music era, ensuring scalability across complex licensing networks.

---

[16:10](https://www.youtube.com/watch?v=We9POs5uaho&t=970s) The Near-Future Horizon: Real-Time Dynamic Music Personalization

Looking at the immediate horizon for generative music, the core industry priorities center on three pillars: establishing explicit consent as the universal standard, securing fair compensation for human creators, and unlocking real-time music personalization. The potential for AI lies not in replacing human composers, but in dynamically adapting their works to individual listener preferences across gaming, social platforms, and digital advertising.

In interactive gaming, a game's score could adjust in real time based on a player's aesthetic tastes and emotional state. Similarly, digital advertisements on platforms like Meta and TikTok could serve dynamic soundtracks customized to individual viewers. Because media platforms prioritize viewer retention, implementing personalized audio that keeps a gamer playing for 30 additional minutes or an ad viewer engaged for 30 extra seconds will drive widespread adoption, opening up brand-new sync licensing revenue streams for music creators.

- **Three Core Pillars:** Industry-wide consent standards, fair remuneration models, and dynamic audio personalization. - **Engagement-Driven Sync:** Platforms and advertisers will monetize adaptive audio to lift user retention in video games and social media feeds without displacing human-composed stems.

---

[17:57](https://www.youtube.com/watch?v=We9POs5uaho&t=1077s) Normalization of AI Media and the "CAM" Framework

Within a short period, the distinction between "AI-generated content" and traditional media will dissolve entirely. In the same way that the phrase "digital music market" was eventually shortened to just "the music market," AI-assisted audio will become ubiquitous across everyday listening environments—from grocery store public address systems and hotel hallways to Alexa smart speakers, television programming, and streaming platforms. Ultimately, consumers do not focus on the technical tools used in production; they seek an emotional or informational connection with the content.

To sustain this transition, the commercial relationship between AI developers and music rights holders must adhere to the foundational framework of **CAM**: 1. **Consent:** Securing explicit authorization from rights holders before training or generating derivative works. 2. **Attribution:** Tracking the precise human inputs and stylistic ancestry behind every generated asset. 3. **Monetization:** Routing recurring royalty streams back to the original creators through existing distribution rails.

---

[19:21](https://www.youtube.com/watch?v=We9POs5uaho&t=1161s) Addressing the "Terminator" Narrative with Practical Optimism

Addressing the anxieties of artists who have opted out of AI data training over fears of an apocalyptic, "Terminator, Judgment Day" scenario for their careers, the panel expresses firm optimism for the next five years. With attribution frameworks on track to be implemented at scale within the next year, the industry is positioned for a structured, equitable transition by 2027.

The session closes with a call for pragmatism over panic. As Matt Adell summarizes: "I choose optimism. I don't know any other way to live my life." By pairing technological attribution with established rights infrastructure, the music industry can transform generative tools from an existential threat into a sustainable economic engine for human artistry.

## Notable quotes > [05:49](https://www.youtube.com/watch?v=We9POs5uaho&t=349s) 10% influenced by this Drake song, 3% influenced by this Beatles song, [and] 0.001% influenced by that punk rock band from Peoria that you've never heard of. — Matt Adell

> [09:05](https://www.youtube.com/watch?v=We9POs5uaho&t=545s) It is totally possible to license all the music in the world. — Matt Adell

> [09:05](https://www.youtube.com/watch?v=We9POs5uaho&t=545s) We measure manure in volume. We shouldn't measure other human output besides manure in volume. We should measure it based on the value it creates, and that can be done on any copyrighted or proprietary data. — Matt Adell

> [13:08](https://www.youtube.com/watch?v=We9POs5uaho&t=788s) first dollar

> [17:57](https://www.youtube.com/watch?v=We9POs5uaho&t=1077s) digital music market

> [17:57](https://www.youtube.com/watch?v=We9POs5uaho&t=1077s) the music market

> [19:21](https://www.youtube.com/watch?v=We9POs5uaho&t=1161s) Terminator, Judgment Day

> [19:21](https://www.youtube.com/watch?v=We9POs5uaho&t=1161s) I choose optimism. I don't know any other way to live my life. — Matt Adell

## Mentioned Matt Adell, Kristen Agee, Xena Plat, Musical AI, 411 Music Group, Ai4, Suno, GEMA, Spotify, YouTube, Drake, The Beatles, Shel Silverstein, Star Wars, SOCAN, Jack Benny, Henry Mancini, Meta, TikTok, Alexa

## Video description Matt Adell - Co-Founder and COO, Musical AI Kristen Agee - Founder/CEO, 411 Music Group Xena Plat - COO, 411 Music Group

Generative AI is rewriting the economics of music faster than the industry can agree on the rules - and the real question isn't whether it reshapes the business, but who gets paid when it does. 411 Music Group sits at the center of that question, representing 5,000+ artists and 120 labels worldwide, with active AI-licensing deals and a roster split between artists who've opted in and those who've opted out. Expect a candid look at the technology functioning as a "new PRO" for the AI era - detecting when generated audio draws on your work and routing payment back to rights holders - and at a future where AI doesn't replace human-made music and sync, but personalizes it: real-time, listener-specific soundtracks across film, TV, and gaming.

[Source](https://www.youtube.com/watch?v=We9POs5uaho)

中文简报

署名革命：艺术家、AI与未来获酬机制 专家组以GEMA对阵Suno的版权胜诉等法律判例为切入点，破除了“商业AI许可可使企业用户免受底层训练数据责任影响”的误解。411 Music Group与Musical AI提出了CAM框架（知情同意、署名归因与商业化），重点阐释了精细化的贡献者追踪及影响力百分比算法如何将上游订阅资金池和下游表演版税回流至创作者。在与SOCAN合作等实际应用案例的支持下，发言嘉宾指出，AI不会取代人类艺术创作，反而将在游戏和数字媒体领域释放动态、实时的个性化音频新可能。 德国法院对 Suno 的侵权判决为行业敲响警钟，凸显出 AI 厂商的商业服务条款并不能免除下游企业因底层训练数据侵权而面临的第三方追责风险。 (0:04) 音乐版权许可需对录音与词曲版权进行细致拆分；传统的固定入库费主要利好海量曲库所有者，而独立艺术家则亟需下游长期归属收益的激励机制。 (2:07) Musical AI 研发的归属技术能够精确计算每首训练歌曲对生成结果的具体影响比例，为客观分配订阅与商业版税提供了量化依据，有效应对无授权抓取带来的巨额法定赔偿风险。 (5:49) 加拿大版权组织 SOCAN 已正式引入归属报告技术处理版税分配，依托许可、归属与变现（CAM）框架，推动生成式 AI 转型为创作者的可持续经济引擎。 (15:06)

Reading summary

Industry executives argue that generative AI enterprises face catastrophic copyright liabilities unless they adopt a structured framework of consent, attribution, and remuneration, using the music industry's collective licensing history as a model for all proprietary data. Legal Vulnerabilities and the Commercial Licensing Illusion (0:04) Generative AI developers face severe legal exposure if they rely on mass-scraped training data without proper authorization. Kristen Agee highlights a German court ruling against AI music platform Suno, brought by the performance rights organization GEMA, which ordered the vendor to cease unauthorized uses, disclose profits, and pay damages. Agee warns that commercial terms granted by AI platforms do not indemnify or protect downstream enterprise customers from third-party copyright claims. The misconception that contractual vendor terms equate to legally cleared outputs exposes enterprise users across video, visual art, journalism, and enterprise software to substantial legal liability. Under U.S. copyright law, Matt Adell notes, statutory damages can reach up to $150,000 for each individual copyrighted work duplicated without authorization. If an enterprise scrapes and stores entire catalogs across training clusters, potential liabilities can rapidly reach tens of billions of dollars. Adell rejects the defense that licensing vast catalogs is logistically impossible, pointing out that streaming platforms such as Spotify have successfully cleared worldwide rights for decades. Informed Consent and Catalog Size Disparities (2:07) Operationalizing consent requires unbundling master recording and composition rights across co-writers, publishers, producers, and record labels. Drawing on 411 Music Group's licensing process across 5,000 artists and 120 labels, Xena Plat explains that authorization cannot be granted en masse; it must be tracked per contributor, per right, and per specific application. This process exposed a clear divide between creator tiers. Labels and legacy artists holding deep catalogs favored immediate opt-ins to secure flat upfront ingestion fees per asset. Conversely, independent creators with smaller catalogs saw little economic benefit in one-time flat fees. Plat argues that genuine informed consent requires AI platforms to offer granular parameters, including whether data is used for assistive tools versus generative replacement, licensing durations with explicit term limits, and recurring backend equity or subscription shares. Influence Attribution and Dual Revenue Architecture (5:49) Attribution technology is designed to measure the precise mathematical influence of training data on generated outputs. Adell explains that Musical AI's system calculates fractional contributions—such as determining an output is 10% influenced by one specific artist and a fraction of a percent by another—enabling downstream income to be divided proportionately. Adell argues that the tech industry must abandon raw data volume metrics and instead measure compensation based on the specific value a work creates across audio, visual design, and proprietary enterprise data. Financial distribution is structured around two distinct revenue pools. Upstream, AI platforms allocate a percentage of user subscription fees into a central pool, echoing Spotify's ~70% payout model. Downstream, when generated content is commercially exploited in advertising, television, or streaming, existing collecting societies capture revenue at the source. Demonstrating scalability, Canadian rights organization SOCAN integrated Musical AI's attribution reporting directly into its royalty collection pipeline, establishing a modern equivalent to historic sheet-music performance tracking. Real-Time Personalization and Market Normalization (16:10) Looking ahead, the panel forecasts that the near-term economic engine for generative audio lies in real-time personalization rather than human replacement. Generative engines can dynamically adapt human-composed music to user preferences and emotional states in interactive video games or optimize digital ad soundtracks on TikTok and Meta. Because dynamic audio improves user retention, it opens new synchronization licensing revenue for rights holders without displacing original composers. As AI-assisted audio becomes ubiquitous across retail spaces, broadcasts, and streaming platforms, the panel projects that the distinction between artificial and traditional media will dissolve. Sustaining this ecosystem requires universal commitment to the CAM framework (Consent, Attribution, and Monetization). Addressing anxieties about career displacement, the panel asserts that scaling attribution over the next five years will convert generative AI into an equitable revenue system for human creators. The viability of this compensation architecture hinges on whether the broader artificial intelligence industry voluntarily adopts transparent attribution rails or must be forced into compliance through escalating copyright litigation and statutory damage awards.

精炼阅读摘要

Kristen Agee、Xena Plat 与 Matt Adell 指出，生成式人工智能企业普遍误判了商业授权的法律合规性，若不建立“授权同意、价值归因与合理报酬”（CAM）治理闭环，将面临数十亿美元法定赔偿的系统性法律风险；依托数学归因技术与传统集体版权管理组织，行业有望将侵权危机转化为可持续的创作者分润机制。 商业许可假象与高额法定侵权风险 (0:04) Kristen Agee 以德国法院判定音乐生成平台 Suno 侵犯德国表演版权组织 GEMA 权益的诉讼为例指出，生成式 AI 平台向用户单方承诺的“商用许可”并不等同于底层技术合规。企业客户常误以为平台合同保障了输出内容的合法性，但实际上此类条款无法豁免第三方版权方针对未经许可的底层训练数据所提出的侵权诉求。音乐产业诉讼仅是视觉艺术、影视、出版以及企业专有数据面临版权风暴的前哨战。 Matt Adell 进一步警示，无授权从 YouTube 等平台批量抓取版权数据的做法存在严重法律风险。依据美国版权法，未经授权复制每一件受保护作品所面临的法定损害赔偿金最高可达 150,000 美元。若 AI 企业在计算集群中复制并存储整个全球音乐库，潜在赔偿总额将迅速攀升至数百亿美元，对科技行业构成毁灭性估值风险。Adell 驳斥了全球版权不可能完全清理的行业辩解，并以 Spotify 长期成功清理全球版权为例，主张全面合法许可在商业操作上完全可行。 精细化授权机制与曲库规模引发的利益断层 (2:07) 411 Music Group 的 Xena Plat 结合管理 5,000 多名独立艺术家与 120 多家唱片厂牌的经验指出，AI 训练授权不能采取一揽子全权许可，而必须将录音版权与词曲版权解绑，细化追踪至每一位合作创作者、出版商与制作人。为避免利益冲突，411 Music Group 采取了厂牌与艺术家分场交流的形式开展版权教育。 实际调研显示创作者群体存在显著诉求分化：拥有数百至数千首曲目深度曲库的传统唱片厂牌倾向于立即签约，因为按摄入曲目支付单笔固定费用的交易模式能为其带来可观的一次性现金回报；然而，仅拥有少量专辑（如数张五轨专辑）的独立创作者却无法从微薄的单曲固定费用中获利。Plat 强调，为保障创作者知情同意，AI 平台必须明确披露模型是属于生成式替代引擎还是辅助制作工具、报酬是单笔固定费用还是平台订阅分成，以及授权是严格限期（如 1 年或 5 年）还是永久不可撤销。 基于影响比例的数学归因技术与价值衡量 (5:49) Matt Adell 提出，在海量数据混合生成输出的阶段，必须依靠归因技术来衡量人类资产的实际价值贡献。他直言科技界应摒弃仅按原始数据吞吐量计算价值的粗放模式，强调除了肥料以外，人类智力产出不应单纯以体量来衡量。Musical AI 开发的技术能够精确测算出单次生成内容中每项训练数据的具体影响占比，例如精细至某个输出受 Drake 作品影响 10%、受 The Beatles 影响 3%，或受不知名独立乐队影响 0.001%。 Adell 说明，该归因报告构成了分配后端收益的客观依据，使驱动模型生成品质的核心作品能获得公平成果。这种数学归因逻辑在影视、诗歌风格、角色外观设计乃至药物研究专有数据中均可广泛适用；但他同时承认，该技术目前尚难以精确拆解在数千篇新闻报道中高度重复的历史事实性表述。 双重收益池架构与集体版权组织的落地整合 (13:08) 生成式 AI 时代的人类创作报酬体系被划分为两个独立收益池：上游的软件即服务（SaaS）订阅资金池，以及下游的商业化使用资金池。在上游阶段，AI 平台需将订阅与调用收入的一定比例拨入版权池，效仿 Spotify 将约 70% 的总收入按流媒体播放份额分给版权方的传统机制；在下游阶段，当 AI 生成的音乐流向 Spotify、YouTube 或影视广告配乐时，则由现有集体版权管理组织在源头直接截留分成。 为验证归因结算的可扩展性，加拿大作曲家、作家和音乐出版社协会（SOCAN）已正式接入 Musical AI 的归因报告技术，直接用于计算并分配 AI 资产对应的创作者版税。该模式借鉴了大乐队时代以来的传统版权基础设施，正如早年拉斯维加斯夜总会演出 Henry Mancini 的乐谱时，原有集体许可体系便能精准将版税回流给词曲作者，无需重构底层版权运作机制。 实时动态个性化应用与CAM治理框架前景 (16:10) 展望应用前景，生成式音频的核心突破在于实时动态音乐个性化，而非简单替代人类作曲家。在互动游戏与 Meta、TikTok 等社交广告场景中，AI 可根据玩家情绪状态或受众偏好实时微调配乐，通过延长 30 分钟游戏时长或 30 秒广告留存率，为音乐人开辟全新的同步授权收益渠道。 讨论总结指出，AI 生成媒介终将像数字音乐取代传统载体一样融入日常背景，产业平稳过渡必须依托“同意（Consent）、归因（Attribution）、商业变现（Monetization）”三位一体的 CAM 框架。针对部分创作者对技术替代产生的前途毁灭性恐慌，与会专家预测归因系统有望在一年内规模化部署，并在 2027 年前推动音乐与科技生态实现良性对接。 该讨论确立了通过数学归因技术与既有集体版权管理网络化解 AI 侵权危机的实施路径，但其体系能否顺利推行，仍取决于全行业能否普遍接受类似流媒体约 70% 的高比例收入分成模式，以及该归因算法在音频之外更为分散的文本与事实性内容领域中能否实现同样精细的权利结算。

Detailed notes — zh

## Background 该视频由频道 Ai4 于 2026 年 8 月 22 日发布在 YouTube 平台上。内容由 411 Music Group 创始人兼 CEO Kristen Agee、COO Xena Plat 与 Musical AI 联合创始人兼 COO Matt Adell 共同对谈，深入探讨生成式 AI 如何重构音乐产业经济学、基于归属技术的版权分账新机制，以及实时个性化音乐对创作者变现的深远影响。 ## TL;DR 对谈围绕生成式 AI 时代版权落地的三大核心支柱——授权同意（Consent）、影响力归属（Attribution）与报酬结算（Remuneration）展开。嘉宾们从德国法院裁决 Suno 侵权案切入，指出科技平台商业许可无法免除底层版权风险，并驳斥了无法实现全量授权的论调。Musical AI 介绍了能够精准测算单条训练数据贡献权重的逆向归属算法，并联合加拿大版权管理组织 SOCAN 跑通了订阅与下游衍生双资金池的分账闭环。展望未来，超个性化音乐将在游戏与广告等场景创造增量价值，推动行业在 2027 年走向人机共生的良性商业生态。 ## Atmosphere & style 整场座谈节奏紧凑且充满产业一线的务实洞察，三位嘉宾以直言不讳的态度迅速打破科技界与版权界的话语壁垒。对话中既有针对数十亿美元潜在侵权风险的严峻法律警示，也有将粗放计费比作“给农家肥按体积称重”的辛辣幽默，在法理严谨性与生动比喻间取得了极佳平衡。现场情绪从对创作者被替代恐慌的细致剖析，逐步走向对历史版权追踪经验的笃定与对 AI 赋能创作者未来的开朗乐观。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=We9POs5uaho&t=4s) AI 平台用户协议中的“商用许可”无法免除底层训练数据未获授权的侵权法律责任，德国法院裁决 Suno 侵权为行业敲响警钟。 - [00:04](https://www.youtube.com/watch?v=We9POs5uaho&t=4s) AI 时代版权合规与健康发展的底层逻辑由三大核心支柱构成：知情同意（Consent）、影响力归属（Attribution）与合理报酬（Remuneration）。 - [02:08](https://www.youtube.com/watch?v=We9POs5uaho&t=128s) 音乐版权涉及录音母带与词曲著作权等多方权益，必须针对每位创作者与使用场景落实精细化的主动选择加入（Opt-in）授权机制。 - [02:08](https://www.youtube.com/watch?v=We9POs5uaho&t=128s) 曲库持有者与小型独立音乐人对 AI 授权态度分化：大型厂牌青睐单资产计费变现，小独立音乐人因微薄收益与作品被利用顾虑更倾向观望。 - [05:51](https://www.youtube.com/watch?v=We9POs5uaho&t=351s) Musical AI 开发的归属技术能够逆向测算训练数据对生成结果的影响力权重，解决 AI 时代价值贡献精准度量与按比例分账的痛点。 - [09:09](https://www.youtube.com/watch?v=We9POs5uaho&t=549s) 未经授权抓取版权数据训练模型存在数十亿美元的高额法定赔偿风险，而全面正规授权在实操上完全可行，是 AI 规模化合规运营的必由之路。 - [11:41](https://www.youtube.com/watch?v=We9POs5uaho&t=701s) 内容授权度量标准应打破“按体积称重”的粗放模式，归属算法可无缝拓展至影视 IP、文学风格及生物医药专利等高价值专有数据领域。 - [13:08](https://www.youtube.com/watch?v=We9POs5uaho&t=788s) 生成式 AI 的商业报酬分为上游 SaaS 软件订阅资金池与下游流媒体及商业发行衍生收益池，延续了传统广播版税的结算逻辑。 - [15:18](https://www.youtube.com/watch?v=We9POs5uaho&t=918s) 加拿大版权组织 SOCAN 采纳 Musical AI 归属判定报告进行版税分配，证明历史悠久的版税集体管理网络能成功适配 AI 资产追踪。 - [16:10](https://www.youtube.com/watch?v=We9POs5uaho&t=970s) 生成式 AI 将推动游戏动态配乐与社交广告等“超个性化音乐”革命，通过提升受众留存为传统作曲家开辟增量变现场景而非单纯替代人类。 - [17:47](https://www.youtube.com/watch?v=We9POs5uaho&t=1067s) 未来生成式内容将自然融入并统一被称为“媒介”，随着归属技术成熟部署，产业生态有望在 2027 年迎来多方共赢的规范化格局。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=We9POs5uaho&t=4s) 德国法院裁决引爆的行业警钟：AI 商用许可不等于法律豁免 座谈会一开始，411 Music Group 创始人兼 CEO Kristen Agee 直奔主题，指出在短短 20 分钟内要探讨的话题极其庞大。她首先引用了几天前刚发生的一起标志性司法判决：德国表演权组织 GEMA 针对知名生成式 AI 音乐平台 Suno 提起版权诉讼，德国法院最终判决 Suno 侵权，勒令其披露相关利润并支付赔偿金，同时禁止其进行未经授权的特定使用。这一裁决在整个内容与科技界激起千层浪。 Kristen 进一步向在场非音乐领域的 AI 从业者抛出核心警示：当下许多科技公司存在一种危险的误区，以为只要 AI 平台在用户协议中写明生成结果“可用于商业用途”，企业就能合法安全地将其变现、发布、广播或嵌入产品。然而事实并非如此——平台给予用户的合同许可，无法自动免除底层模型训练数据是否获得正规授权的法律责任，也无法保证输出结果不侵犯原权利人的受保护元素。音乐领域经历的版权风暴只是前沿缩影，同样的问题正迅速蔓延至图像、视频、新闻出版、电影、代码软件以及企业专有数据领域。整场讨论将围绕三大核心支柱层层展开：同意（Consent）、归属（Attribution）与报酬（Remuneration）。 ```mermaid flowchart TD accTitle: AI版权合规与价值分配闭环 subgraph S1["第一阶段：合法准入"] A["权利人许可 / Opt-in 同意"] --> B["合规训练数据集"] end subgraph S2["第二阶段：模型与生成"] B --> C["生成式 AI 模型训练"] C --> D["生成内容与商业输出"] end subgraph S3["第三阶段：归属与分成"] D --> E["Musical AI 影响力归属分析"] E --> F["双资金池结算：平台订阅费 + 下游生态变现"] F --> G["收益精准返还原始版权方"] end ``` ### [02:08](https://www.youtube.com/watch?v=We9POs5uaho&t=128s) 411 Music Group 的授权实践：同意权机制与版权方分歧 411 Music Group 首席运营官兼音乐人 Xena Plat 结合公司代理超过 5,000 位艺术家和 120 家独立厂牌的实操经验，详细拆解了版权准入流程。她强调，“获得权利人的真正同意是绝对的基石”。在音乐工业中，录音母带版权（Master）与词曲著作权（Composition）往往归属不同的创作者或机构，因此没有任何单一个人能够全权代表一首歌做出决定。企业必须按每一项权利、每一位创作者、每一种使用场景逐一获取并记录真实的 Opt-in（选择加入）授权。 为了推动知情同意，411 组织了多次市政厅式（Town Hall）闭门沟通会，将艺术家与厂牌分开安排，不搞商业推销，而是客观普及 AI 授权作为新型变现渠道的运作机制。这一过程呈现出两极分化的态度： - **深厚曲库拥有者（厂牌与资深音乐人）**：拥有 100 首至数千首曲目的权利人态度极度积极，因为 AI 训练数据交易通常采用“单资产计费”（Per-asset fee），上千首歌曲能直接兑现可观的现金收益。 - **小型独立创作者**：像 Xena 本人及合作创作者这种仅拥有两三张专辑、十几首作品的音乐人则极为谨慎，因为微薄的一次性收益难以打消对作品被利用的顾虑，很多人选择暂时观望。 Xena 补充道，创作者在签署授权前必须明确核心交易条款：AI 是用于生成式模型还是仅用于辅助制作（Non-generative/Assistive tool）；报酬模式是一次性买断费（Flat fee）还是包含后端订阅分成（Subscription share）；以及授权年限究竟是一年、五年还是永久。 ### [05:51](https://www.youtube.com/watch?v=We9POs5uaho&t=351s) Musical AI 的归属技术：从训练数据影响度到精细化分账 针对海量数据涌入模型后的“黑盒”难题，Musical AI 联合创始人兼首席运营官 Matt Adell 指出，音乐产业历来是数字版权经济生态的“探路先锋”，当前音乐领域的机制必将复刻到整个互联网内容行业。AI 研发不仅需要庞大的数据规模，更需要高质量的合法数据。一旦合规进入训练阶段，商业化变现将形成两大资金池： 1. **工具订阅资金池**：终端用户向 AI 平台支付的软件订阅费用。 2. **下游生态资金池**：生成的音频资产流向 Spotify、流媒体平台或电视广告等场景产生的实际商业变现。 Matt 阐述了 Musical AI 所构建的归属技术（Attribution）：系统能够对每一次生成输出进行逆向影响力测算，精确量化各条训练数据的贡献权重。“我们可以明确告诉你，这段生成音频有 10% 受到 Drake 某首歌曲的影响，3% 受到披头士的影响，还有 0.001% 受到某支来自 Peoria、你从未听过的小众朋克乐队的影响。” 这一技术生成的影响力权重报告，解决了“如何公平分账一美元”的行业痛点。正如流媒体时代歌曲根据播放量结算版权费一样，AI 时代版权方将根据自身作品对生成结果的实际价值贡献度，按比例从数百亿美元的订阅与下游生态池中获得分账。 ### [09:09](https://www.youtube.com/watch?v=We9POs5uaho&t=549s) 侵权法律风险与十亿级风险敞口：为什么全面授权并非不可能 在探讨合规发展路径时，Matt 直率地指出当前 AI 公司存在的严重侵权行为。许多平台在未经许可的情况下擅自从 YouTube 抓取全网音频作为训练集。他打趣道，自己 1984 年组建的朋克乐队作品就被上传在 YouTube 上，如今毫无疑问已被卷入多家 AI 公司的模型之中。 Matt 严肃警告了潜在的系统性法律风险： - **高额法定赔偿**：根据美国版权法，未经授权复制版权作品的法定赔偿金最高可达每份拷贝 150,000 美元。 - **数十亿美元的风险敞口**：若一家 AI 公司私下保存了全球数千万首版权歌曲的多个副本，其潜在法律赔偿责任高达数十亿美元，这种隐藏风险甚至可能直接戳破当前的 AI 产业估值泡沫。 面对部分 AI 企业声称“给全世界所有音乐做授权在实操上根本不可能”的说法，Matt 给予了严厉驳斥。他指出自己从事版权授权工作已有 30 年，Spotify 早就实现了全球音乐的全面授权，“全面正规授权不仅完全可行，而且是 AI 企业在全球合法规模化运营的唯一路径”。企业应当主动接入白标同意管理系统（Consent Management System），与环球（Universal）、索尼（Sony）及 411 等版权机构达成深度分账合作。 ### [11:41](https://www.youtube.com/watch?v=We9POs5uaho&t=701s) 归属技术的跨行业延伸：告别“按体积称重”的内容评估 Matt 进一步明确，归属与溯源算法不仅适用于音乐，还可以无缝拓展至全行业的专有与受保护数据。尽管通用事实（例如“某位美国总统在哪一年遇刺”）因出现在海量公共新闻中而难以溯源，但任何具有高度创意性或专有价值的资产都能精准追踪其生成影响力： | 行业领域 | 典型适用资产示例 | 归属与价值转化方式 | |:--- |:--- |:--- | | **影视动漫** | 《星球大战》风暴兵（Stormtrooper）视觉造型 | 追踪提示词生成图像对特定 IP 视觉资产的依赖度 | | **文学出版** | 知名诗人谢尔·希尔弗斯坦（Shel Silverstein）的诗歌风格与文本 | 测算生成文本在押韵、句式与修辞上对专有文学作品的继承比例 | | **生物医药** | 制药研发专利数据、分子结构模型与专有实验数据 | 追踪生成式分子设计中对特定专有数据集的参考与贡献权重 | Matt 用一个辛辣幽默的比喻总结了行业度量标准的变革：“世界上唯独在衡量农家肥（Manure）的时候我们才看体积。除了肥料，人类的一切智力产出都不应当按体积来称重。”AI 时代的内容授权必须打破粗放的“按量计费”模式，转向基于实际价值创造的归属计量。随着同意与归属两大技术基准的确立，座谈随后转向了整个闭环的落脚点——多方报酬机制（Remuneration）的落地设计。 ### [13:08](https://www.youtube.com/watch?v=We9POs5uaho&t=788s) 生成式 AI 的双重收益池与归属结算机制 在生成式 AI 生态中，创作者与版权方获取报酬的核心机制可以划分为两个主要收益池。第一类是**订阅或使用费收益池**（Subscriber/Usage Revenue Pool）。当用户为生成音乐、影视或药物研发方案的 SaaS 工具付费时，这些 AI 平台会将部分收入划入资金池，并根据各训练素材的价值权重进行分配。这与流媒体巨头的分配逻辑如出一辙——以 Spotify 为例，平台会将大约 70% 的总收入注入统一资金池，再按照每首歌曲的实际播放次数分配给音乐创作者；在当前的 AI 订阅服务中，这种模式已经开始落地。第二类则是更为复杂的**下游衍生收益池**（Downstream Output Revenue Pool）。当 AI 生成的电影、电视剧、书籍或音乐流入市场并在 YouTube、Spotify 或院线等渠道产生盈利时，版权方需要从资产变现的源头进行第一笔资金归集。 这一过程并非凭空捏造的新逻辑，而是对传统广播电台版税结算机制的现代延续。历史上，每当歌曲在电台上播放，版税都会汇入版权集体管理组织（Collecting Society），由其评估各单曲的价值并通过报告将资金转交给出资人、词曲作者及演奏者。这证明在大规模生态中为数百万人类创作者按价值贡献发放报酬在技术与机制上完全可行。 ```mermaid flowchart TD accTitle: AI 音乐产业双重收益流与归属结算架构 subgraph S1["第一收益池：上游订阅与使用费"] A["用户/企业付费订阅<br/>(AI SaaS 平台)"] --> B["订阅收益池<br/>(按比例切分收入)"] B --> C["基于训练数据资产价值分配"] end subgraph S2["第二收益池：下游内容衍生与流媒体变现"] D["AI 生成内容分发<br/>(Spotify / YouTube / 影视同步)"] --> E["下游资产产生商业收益"] E --> F["版权集体管理组织 (PROs) 归集资金"] F --> G["Musical AI 归属报告精确分账"] end C --> H["人类词曲作者、演奏者与版权方"] G --> H ``` - **流媒体版税先例**：Spotify 将其总收入的约 70% 投入收益池，依据播放量全量分配给内容生产者。 - **广播与演出逻辑**：无需重新发明轮子，将 AI 归属模型接入既有的表演权组织（PROs）即可实现规模化运作。 --- ### [15:18](https://www.youtube.com/watch?v=We9POs5uaho&t=918s) 接入传统版权组织：与 SOCAN 的合作及历史先例 为了让 AI 归属模型在现实世界中运转，必须将其与既有的音乐版权基础设施无缝对接。Musical AI 宣布与加拿大最大的音乐版权管理组织之一 **SOCAN**（加拿大作曲家、作家和音乐出版社协会）达成战略合作。SOCAN 正式选用 Musical AI 的归属判定技术，利用其生成的精确数据报告作为版税结算依据；每当 SOCAN 归集到涉及 AI 资产的版权费用时，便能准确识别出具体由哪位词曲作者或演奏家参与贡献，并按比例完成精准分账。 这种版税分配机制并非计算机时代的产物，而是拥有深厚的历史根基。早在计算机尚未普及、大乐队（Big Band）依然依赖实体乐谱演出的年代，行业就已经建立了成熟的追踪系统。正如 Matt Adell 所举的生动例证：当年杰克·本尼（Jack Benny）在拉斯维加斯的夜总会演奏亨利·曼西尼（Henry Mancini）的作品时，原始曲作者就能通过既有的版税追踪网络获得应得的报酬。历史经验表明，版权追踪与利益分配完全能够随着技术的革新快速实现规模化落地。 - **SOCAN 合作落地**：加拿大顶级版权管理组织采纳归属报告，用于指导实际版税分配。 - **历史验证**：从拉斯维加斯夜总会的实体乐谱时代到现代数字流媒体，版权追踪体系历经多轮技术变迁均能成功演进。 --- ### [16:10](https://www.youtube.com/watch?v=We9POs5uaho&t=970s) 音乐个性化革命与创作者的全新变现机遇 谈及生成式 AI 在音乐领域的未来图景与核心用例，411 Music Group 团队提出了三大核心愿景：首先是**“获取授权（Consent）”成为全行业严格通行的标准**；其次是**保障音乐创作者获得公平合理的报酬**；而最具变革性的应用方向则是**超个性化音乐（Music Personalization）**。在未来的娱乐交互中，背景音乐将完全根据用户的实时偏好动态生成。例如，在沉浸式游戏中，玩家耳畔的配乐会根据个人喜好进行实时微调；在 Meta 或 TikTok 等平台上浏览视频广告时，背景音轨也会智能切换为受众最想听的旋律风格。 从商业转化来看，个性化背景音轨如果能有效提高用户留存——哪怕仅仅让玩家多玩 30 分钟游戏，或者让用户在广告上多停留 30 秒——各大平台与品牌方就会毫不犹豫地大规模部署该技术。与外界普遍担心的“AI 替代论”不同，这种变革并不会消灭作曲家与人类创作者，反而会为创作者的作品开拓前所未有的全新使用场景与商业变现管道。 - **游戏音轨动态定制**：根据玩家实时行为与审美偏好调整游戏配乐。 - **社交广告精准留存**：在 Meta 与 TikTok 等平台利用定制音乐提升广告观看时长与转化率。 - **赋能而非替代**：为传统作曲家开拓增量应用场景，释放存量曲库的新商业价值。 --- ### [17:47](https://www.youtube.com/watch?v=We9POs5uaho&t=1067s) 告别“审判日”恐慌：走向共生共赢的 2027 回溯科技传播史，Matt Adell 直言预测未来虽难，但有一点毋庸置疑：无论是在杂货店、会议走廊、智能音箱 Alexa，还是在电视、流媒体平台 YouTube 和智能手机中，AI 生成的内容很快将无处不在。不久之后，行业将不再特意用“AI 内容”或“生成式内容”来加以区分，它将直接被统称为“媒介（Media）”。这与历史上的“数字音乐（Digital Music）”演变如出一辙——人们曾将部分市场划归为数字音乐市场，而如今它已自然成为音乐市场的全部；消费者本质上不在乎创作工具是什么，他们只在乎与内容之间是否产生了情感共鸣与信息连接。 在 AI 科技公司、音乐机构与艺术家之间，健康合作的基石依然是三个不可动摇的支柱：**授权（Consent）、归属（Attribution）与变现（Monetization）**。尽管 411 Music Group 旗下以及行业内有部分艺术家因担心“终结者审判日（Terminator / Judgment Day）”般的毁灭性打击而选择退出（Opt-out）AI 训练数据池，但行业领袖普遍对未来保持高度乐观。随着归属技术预计在短期内完成规模化部署，整个生态有望在 2027 年迎来成熟、规范且互利共赢的全新格局。 > “消费者不在乎工具。他们在乎的是与所消费的媒介产生情感或信息上的连接。” > “我选择保持乐观。在生活中我不知道还有其他什么过法。” ## Notable quotes > [02:08](https://www.youtube.com/watch?v=We9POs5uaho&t=128s) 获得权利人的真正同意是绝对的基石 — Xena Plat > [05:51](https://www.youtube.com/watch?v=We9POs5uaho&t=351s) 我们可以明确告诉你，这段生成音频有 10% 受到 Drake 某首歌曲的影响，3% 受到披头士的影响，还有 0.001% 受到某支来自 Peoria、你从未听过的小众朋克乐队的影响。 — Matt Adell > [09:09](https://www.youtube.com/watch?v=We9POs5uaho&t=549s) 给全世界所有音乐做授权在实操上根本不可能 — 部分 AI 企业观点 > [09:09](https://www.youtube.com/watch?v=We9POs5uaho&t=549s) 全面正规授权不仅完全可行，而且是 AI 企业在全球合法规模化运营的唯一路径 — Matt Adell > [11:41](https://www.youtube.com/watch?v=We9POs5uaho&t=701s) 世界上唯独在衡量农家肥（Manure）的时候我们才看体积。除了肥料，人类的一切智力产出都不应当按体积来称重。 — Matt Adell > [17:47](https://www.youtube.com/watch?v=We9POs5uaho&t=1067s) 消费者不在乎工具。他们在乎的是与所消费的媒介产生情感或信息上的连接。 — Matt Adell > [17:47](https://www.youtube.com/watch?v=We9POs5uaho&t=1067s) 我选择保持乐观。在生活中我不知道还有其他什么过法。 — Matt Adell > [17:47](https://www.youtube.com/watch?v=We9POs5uaho&t=1067s) 终结者审判日（Terminator / Judgment Day） — Matt Adell ## Mentioned Matt Adell, Kristen Agee, Xena Plat, Musical AI, 411 Music Group, GEMA, Suno, Spotify, Drake, 披头士 (The Beatles), YouTube, 环球音乐 (Universal), 索尼音乐 (Sony), 《星球大战》风暴兵 (Stormtrooper), 谢尔·希尔弗斯坦 (Shel Silverstein), SOCAN (加拿大作曲家、作家和音乐出版社协会), 杰克·本尼 (Jack Benny), 亨利·曼西尼 (Henry Mancini), Meta, TikTok, Alexa ## Video description Matt Adell - Co-Founder and COO, Musical AI Kristen Agee - Founder/CEO, 411 Music Group Xena Plat - COO, 411 Music Group Generative AI is rewriting the economics of music faster than the industry can agree on the rules - and the real question isn't whether it reshapes the business, but who gets paid when it does. 411 Music Group sits at the center of that question, representing 5,000+ artists and 120 labels worldwide, with active AI-licensing deals and a roster split between artists who've opted in and those who've opted out. Expect a candid look at the technology functioning as a "new PRO" for the AI era - detecting when generated audio draws on your work and routing payment back to rights holders - and at a future where AI doesn't replace human-made music and sync, but personalizes it: real-time, listener-specific soundtracks across film, TV, and gaming. [Source](https://www.youtube.com/watch?v=We9POs5uaho)

Source quotes

[2:08] Xena Plat: 获得权利人的真正同意是绝对的基石 [5:51] Matt Adell: 我们可以明确告诉你，这段生成音频有 10% 受到 Drake 某首歌曲的影响，3% 受到披头士的影响，还有 0.001% 受到某支来自 Peoria、你从未听过的小众朋克乐队的影响。 [9:09] 部分 AI 企业观点: 给全世界所有音乐做授权在实操上根本不可能 [9:09] Matt Adell: 全面正规授权不仅完全可行，而且是 AI 企业在全球合法规模化运营的唯一路径 [11:41] Matt Adell: 世界上唯独在衡量农家肥（Manure）的时候我们才看体积。除了肥料，人类的一切智力产出都不应当按体积来称重。 [17:47] Matt Adell: 消费者不在乎工具。他们在乎的是与所消费的媒介产生情感或信息上的连接。 [17:47] Matt Adell: 我选择保持乐观。在生活中我不知道还有其他什么过法。 [17:47] Matt Adell: 终结者审判日（Terminator / Judgment Day）

Source quotes

[5:49] Matt Adell: 10% influenced by this Drake song, 3% influenced by this Beatles song, [and] 0.001% influenced by that punk rock band from Peoria that you've never heard of. [9:05] Matt Adell: It is totally possible to license all the music in the world. [9:05] Matt Adell: We measure manure in volume. We shouldn't measure other human output besides manure in volume. We should measure it based on the value it creates, and that can be done on any copyrighted or proprietary data. [13:08] first dollar [17:57] digital music market [17:57] the music market [19:21] Terminator, Judgment Day [19:21] Matt Adell: I choose optimism. I don't know any other way to live my life.

Original transcript

[0:04] Um, I'm Kristen, as he said, CEO CEO and founder  of Foro Music Group. Um, we have a very ambitious [0:14] panel today and a lot to cover in only 20  minutes. So, we'll see how far we can get. Um, [0:21] I'm going to start with just a headline from a few  days ago. Um, a German court ruled against an AI [0:28] music company, SUNO, in a copyright case brought  by a performing rights organization, GIMA. The [0:35] case involved protecting songs that GAA argued  had been used without authorization. The court [0:40] found infringement, ordered Sunno to disclose  related profits, pay damages, and prohibited [0:46] certain unauthorized uses. So, why should a room  full, semif of people working across the broader [0:55] AI industry care about a music case? [snorts]  Um, because many companies currently assume that [1:01] if an AI platform says its outputs can be used  commercially, those outputs are automatically [1:08] legally safe to monetize, post, broadcast, or  incorporate into another product. And this is [1:14] not necessarily true. Um so a platform may give  its cons customer contractual permission to use [1:21] an output but that does not automatically resolve  whether the material used to train the model was [1:27] properly licensed whether the output reproduces  protected elements or whether another rights [1:33] holder could bring a claim. So music is simply one  early example of a much broader question affecting [1:40] rights uh images, video, publishing, journalism,  film, software, and proprietary business data. [1:48] So can a company build and monetize an AI product  using other people's work without first obtaining [1:55] permission, identifying the cont contributors  or sharing the resulting value. So this is what [2:01] we're going to talk about today. um through these  three connected principles, consent, attribution, [2:07] and renumeration. Big big topic. Uh so let's start  with Zena, our COO of 411 Music Group and musical [2:16] artist herself. Um so at our company at 411,  we work with over 5,000 artists, 120 labels, [2:25] and we went through the process of asking them to  opt in or out of AI training and data licensing. [2:31] So what does this process look like for a company  and what needs to happen before music is used in [2:37] training? Yeah, thank you. I'll first and foremost  want to um point out that getting consent uh from [2:45] the right holders is fundamental. You need to  get a real optin from every contributor. Um as [2:52] you probably some of you know the master and the  composition usually have different contributors [2:58] um and owners. So it's never the one just  one person saying yes. Um that decision has [3:05] to be captured and recorded per rights per  contributor per per use and um before you [3:14] collect the consent. Uh we need to educate because  it's a brand new use case. Um, in our in our case, [3:22] um, we put together town halls and put  artists in one room, labels in another room, [3:29] and we didn't pitch them anything. We, um, told  them what AI licensing actually is, and that's [3:37] the new way for them to monetize their music. Um  uh and got to be honest like the opinion split the [3:45] labels and the artists with deep cataloges call it  like a 100 tracks and up. Uh they were super eager [3:53] um because as you probably know um in AI deals  you get paid per asset. So if you have a thousand [4:02] tracks 2,000 tracks you're looking at real value  real ma money. Whereas uh smaller creators like [4:10] myself and my coowriters um we really had to think  about it because if we only have like two albums, [4:18] five tracks on each album, it's not much that  we're getting out of those deals. But maybe, [4:24] you know, now that attribution is rolling out,  things will change and we'll change our opinion [4:31] on AI licensing. um as me as an a artist and  creator um and will opt in eventually. Um yeah, [4:42] but are there certain things that you as a rights  holder or artist need to know before you can give [4:48] like genuinely give informed consent? Yeah. When  you get paid maybe uh when what what this company [4:57] is, what is the use case, is it generative,  non-generative? because there are plenty of [5:02] platforms out there uh that train AI models um  not for music generation but maybe assisting [5:11] composers, producers in their music production  process and that's a different use case. So, [5:17] um I would be as as a creator interested in  learning about that aspect and what does this deal [5:23] um entail. Um and yeah uh and whether there is  a any backend um like a um subscription share [5:36] or am I looking only at a flat fee? Um what's the  term? Is it perpetual? Is it one year or fiveyear [5:43] term? Yeah. So standard deal points that we would  look at uh in sync licensing which is our core [5:49] business anyway. Okay. So I'm going to bring the  conversation to Matt. Um so consent establishes [5:55] whether the material should have have entered the  system in the first place. Um but once millions of [6:01] pieces of content and data enter a model then  another question arises can we determine what [6:08] contributed to the result? So can you first Matt  explain what is attribution and when we talk about [6:16] attribution in AI what exactly are we trying to  attribute? Sure. Uh uh thanks for having me. As [6:23] you said consent comes first and I think you're  really right to point out that this is not only [6:30] about music. uh music has always been the first uh  enterprise to figure out the economic ecosystem of [6:39] the digitization of copywritten content. So what's  happening in music now is going to happen across [6:44] media across all copywritten material uh out  there on the internet. And consent is uh the key [6:53] component. The goal of a consent management system  is to get as many people hopefully to consent to [7:00] their material, their copywritten stuff being  used because AI companies need a large amount of [7:05] data. They don't only need a large amount of data,  they need great data. And that's really important [7:10] because the next step is once there has been  consent and an AI company is training on material [7:16] that they have permission to train on then they're  m then they're building a business based on the [7:23] value of that material they've trained on and  there is subscription revenue consumers who are [7:28] using the tool and then there's a second pool of  revenue which is once something is generated once [7:34] an asset is generated in the case of music it  might go to Spotify it might being a television [7:39] commercial and that generative asset now is  out making money in the ecosystem. And what [7:45] attribution does is measures the value of each  and every piece of training data related to every [7:54] single output. So what we do at musical AI is for  every single output from a generative AI system, [8:03] we can tell you this output was 10% influenced  by this Drake song, 3% influenced by this Beatles [8:10] song, 0.001% influenced by that punk rock band  from Peoria that you've never heard of. And this [8:18] is significant because then what you have is a  report that can helps that figure out how to share [8:24] a dollar. And in a world where AI companies  are making hundreds of millions of dollars [8:28] in subscription revenue and then the outputs  are going out in the ecosystem and generating [8:34] revenue on their own. Uh there has to be an agreed  upon system for sharing that revenue back to the [8:42] original copyright owners. That's a requirement of  the permissions and consents that the AI companies [8:47] are getting. Just like when you listen to Spotify,  the more a song is played on Spotify, the more it [8:53] gets paid. We've built a system that the more  influential a piece of content is in the value [9:00] creation of the generative model, that piece of  content is paid a greater share of a dollar and we [9:05] generate those reports. Right? Sounds complicated.  Um, what should AI companies that want to operate [9:13] legally be doing now um rather than waiting for  lawsuits to force a standard? Well, first off, I [9:22] think there's some things they shouldn't be doing.  They shouldn't be scraping all the music off of [9:26] YouTube. Uh, I've never consented. My punk rock  band from 1984 is on YouTube and is I know for a [9:33] fact now in a bunch of those models. Now, I'm not  going to uh go and sue somebody over that. Uh, [9:40] but as you mentioned, all the big rights holders  across the globe are suing the AI companies. [9:47] And in the United States, there's a statutory  penalty for every single copy of a copywritten [9:54] piece of content you make a copy of without author  authorization. It's $150,000 per copy. So imagine [10:01] a world where an AI company might have 10 copies  of every single song ever made lying around and [10:09] they owe $150,000 for every single copy. That's  a that's a financial risk of billions of dollars. [10:16] And that should concern all of us because that  really ties back to uh whether we're experiencing [10:21] a bubble or not. And what the AI companies need  to do is they need to work with folks like you who [10:27] use a consent management system, either one that's  homebrewed or one like we have a white label [10:32] consent system uh that rights holders can use to  get in touch with their constituents and let them [10:38] consent. And then the AI company should train on  that consented material. And if and when they do [10:45] a licensing deal with a Universal Music, a Sony,  a 411, uh those rights holders will insist that [10:52] they are paid a per a percentage of subscription  revenue just like in the world of Spotify and [10:58] a percentage of the revenue the outputs go and  generate when they're out there in the ecosystem [11:04] being listened to, used on YouTube, used in  television commercials. And our reports basically [11:09] provide the report of how to share every one of  those dollars. And so, uh, AI companies that want [11:16] to operate, uh, at scale globally and without risk  of lawsuits should be taking those steps. And, and [11:26] we know that it's it's really easy for companies  to say, well, it's impossible to license all the [11:31] music in the world. Well, gee, I've been doing  that for 30 years. It's not impossible. Spotify [11:36] has done it. It is totally possible to license all  the music in the world. Yeah, it's an interesting [11:41] point. Um, can this technology that you've built  and attribution be applied to other industries? It [11:50] can. We can measure the influence on a percentage  basis of any piece of training data to an output. [11:58] Now, if you're talking about facts, like Ronald  Reagan was shot in whatever year he was shot in, [12:05] that's very difficult because that factoid appears  in tens of millions of journalistic data. But if [12:11] you're talking about creative output, like what  does a stormtrooper in Star Wars look like? Uh [12:17] uh uh what does a Shell Silverstein  poem say? Um uh television, film, music, [12:25] and even pharmaceutical data. And any highly  [snorts] valuable proprietary or copywritten data [12:31] can be its influence can be traced and monetized  not just based on volume because you know we we [12:38] measure manure in volume. We shouldn't measure  other human output besides manure and volume. [12:44] Uh we should measure it based on the value it  creates and that can be done on any copywritten [12:49] or proprietary data. Yeah. Okay. Makes sense.  Um so this sort of brings us to renumeration. [12:56] Did you have something? No. Uh, so once consent  has been obtained and the contribution can be [13:03] identified through attribution, how should the  people and companies supplying the underlying [13:08] material be compensated? You can also the the uh  there's two pools of revenue. Uh the first pool [13:15] of revenue is the subscriber or usage revenue.  uh somebody who pays an AI company who wants to [13:21] generate music or generate film or generate a  pharmaceutical idea uh those SAS services will [13:28] pay a portion of their revenue into a pool uh  and that pool will be divvied up based on the [13:35] value of each asset. So Spotify, for those of  you who don't know, pays about 70ome% of all the [13:41] dollars they collect into a pool and then based  on how many times every single song is played, [13:46] that pool is broken up and paid out to all the  people that made that music. The same will be [13:51] true and is already true in some instances  in the world of AI subscription companies. [13:57] The the next piece is a little more complicated  which is once the generative outputs the films [14:02] the television shows the books the music are out  in the ecosystem generating revenue then first [14:10] dollar needs to be collected from where that  asset is generating revenue. An asset might be [14:14] on Spotify making money. It might be on YouTube  making money. It might be uh a song appearing in [14:20] a film you all have seen recently. uh that money  gets collected and just like already happens when [14:27] songs are played on the radio, this is not a  new invention, there's a collecting society [14:32] that collects that revenue, measures the value of  each individual piece of content and uses a report [14:39] like ours to send money out. A lot of you may not  know that every time a song is played on radio, [14:44] money goes into a pool and eventually makes its  way back to the people who performed and wrote [14:49] that song. So again, this is not a new idea that  it is totally possible and feasible at scale to [14:57] pay millions of human creators based on the value  they create in an ecosystem. So it's scalable if [15:06] you connect it to the existing systems already out  there like all the performing rights organizations [15:12] and like how does attribute attribution and your  model fit into the ecosystem like when a song is [15:18] is performed. Sure. So we just announced a  deal with Soan, one of the largest managers [15:24] of music rights in the world based in Canada  and uh they've chosen us as their attribution [15:31] technology. They will use our reports. So when  they collect money on these assets, they will [15:38] use our reports to determine which songwriter,  which musical performer receives what portion of [15:44] those payments. Okay. So it is scalable it sounds  like and it's happening fast. It it it's scalable [15:50] and it has happened again and again repeatedly  even before there were computers when when there [15:56] were just big bands using sheet music there were  systems in place so that when Jack Benny played [16:03] a Henry Mancini song in a nightclub here in Vegas  the original composer was paid. Okay. So we only [16:10] have 3 minutes and 42 seconds and counting. Uh so  we're going to just get some final takeaways. Um, [16:17] and I'll I'll pose the same question to both of  you. Where is this all headed? Um, what is the use [16:23] case for AI in the music space? Let's start with  that one. Yeah, just uh three final thoughts. Um, [16:32] what I'm excited for. Um, first of all, getting  consent being the standard in the industry. Um, [16:40] second of all, um, music creators getting paid  fairly. And, uh, I think where it's headed, [16:51] uh, is music personalization. Imagine the world  where you play a game and the score that you hear [16:58] within the game is adjusted to your preference.  or if you watch an ad on Meta or Tik Tok and the [17:06] song that's playing is something that you would  like to hear and I think um companies sooner or [17:14] later will recognize that and if music um plays  a part in retaining that attention and making the [17:25] gamer to play 30 more minutes or watch the ad for  30 more seconds. Um, that's will be implemented [17:34] um at scale and I think it's not going  to replace composers and music creators, [17:40] but it will definitely open up um a new case use  case uh for them and new ways to monetize their [17:47] music that they haven't seen before. Yes, Matt.  Well, I'm old enough to know that guessing the [17:57] future is a a losing proposition, but I do believe  that everywhere there is music, the grocery store, [18:06] the hallway out here, uh your Alexa speaker,  everywhere there is media on the television and [18:12] your streaming service on YouTube, on your phone,  there will be AI generated content. Uh there [18:18] is zero doubt we will get to a place very soon  where we stop talking about it in terms of AI or [18:24] generated content. It is just simply media. Do you  remember when we used to all call a sector of the [18:29] music market the digital music market and now it's  just the music market and the the same will happen [18:37] very rapidly here. Consumers don't care about  tools. They care about a connection emotionally [18:44] orformationally with the media they're consuming.  But what what does permission and partnership look [18:51] like between AI and tech companies and music  companies and artists? The consent, attribution [18:58] and monetization is exactly what has happened in  every other digitization revolution of media and [19:06] it will happen here as well. that this is not new  in that respect. This is simply another technology [19:14] that rests upon human created media to create  new value for consumers. And a lot of artists [19:21] in our last 30 seconds um that have opted out  of AI data training um within 411 and generally [19:29] speaking have a very like specific thing thinking  that it's going to be like it's Terminator, [19:35] it's judgment day, it's over for us. Is are  do you both feel this way? Are you optimistic [19:40] about the next five years or is it doomsday? I'm  very optimistic, especially knowing that Matt uh [19:49] just gave me like a a timeline on when attribution  can be implemented at scale that it's going to be [19:56] like next year. So, I'm very optimistic to see how  it's going to unfold in 2027. I choose optimism. I [20:04] don't know any other way to live my life. Agreed.  Thank you so much for having us, AI4. Thank you, [20:10] Zena. Thank you, Matt. If you have any questions,  uh, come see us after. And good luck, everyone.