Using AI To Detect AI Music (and other music industry data-porn)

Using AI To Detect AI Music (and other music industry data-porn) — Benn Jordan

Benn Jordan

AI-assisted summary of Using AI To Detect AI Music (and other music industry data-porn) — original source. Check the original for full context.

Source published: 2025-01-05

Key ideas

Scraping 560 top-ranked Suno tracks revealed that 549 (98%) were actively monetized on digital streaming platforms under disguised artist profiles, with a single track pulling nearly 700,000 streams from the shared royalty pool. (1:51)
Benn Jordan developed an AI detection pipeline reaching 100% accuracy on Suno audio by identifying residual compression artifacts left over from the lossy, non-consensual web audio used to train the models. (5:05)
An audit of royalty-free loop services with Venus Theory showed that assembling tracks from Splice, Loopcloud, and vintage sample CDs triggered high rates of automated Content ID collisions, offering no safe option among tested platforms. (10:00)
Streaming payout analysis reveals major DSPs pay near the bottom—Spotify at $0.0029 and YouTube Music at $0.0027 per stream—while Qobuz reaches $0.0136 and curated fitness provider Peloton tops the list at roughly $0.03. (13:19)

Detailed notes

## Background Published by Benn Jordan on YouTube on January 5, 2025, the video investigates key data points across the music industry. According to the description, it covers which streaming platforms pay the highest royalty rates, audits sample libraries for copyright collision issues, and demonstrates how to determine whether music is human-made.

## TL;DR Benn Jordan presents an empirical deep dive into modern music industry dilemmas, beginning with an AI detection pipeline that leverages psychoacoustic compression artifacts to identify scraped Suno tracks monetized under fake artist identities on streaming platforms. Collaborating with Venus Theory, he audits prominent royalty-free sample libraries like Splice, Loopcloud, and vintage sample CDs, showing that automated Content ID systems generate alarming rates of false copyright matches across all services. Finally, drawing on independent distributor data from his own fully owned catalog, Jordan breaks down real-world pay-per-stream royalties, revealing that platforms like Spotify sit near the bottom while Qobuz and curated fitness services like Peloton offer substantially higher payouts.

## Atmosphere & style The presentation blends dry, sarcastic industry weariness with the rigor of a technical investigative report. Jordan speaks with calm, analytical authority, balancing dense spectrogram and audio codec physics with darkly humorous commentary on corporate venture capital and algorithmic streaming dilution. It feels like an intimate, behind-the-scenes laboratory dispatch, transitioning from technical defiance in unmasking bad actors to an impassioned, principled critique of modern streaming economics.

## Bulletin - [01:51](https://www.youtube.com/watch?v=QVXfcIb3OKo&t=111s) Scraping 560 top-ranked Suno tracks revealed that 549 (98%) were actively monetized on digital streaming platforms under disguised artist profiles, with a single track pulling nearly 700,000 streams from the shared royalty pool. - [05:05](https://www.youtube.com/watch?v=QVXfcIb3OKo&t=305s) Benn Jordan developed an AI detection pipeline reaching 100% accuracy on Suno audio by identifying residual compression artifacts left over from the lossy, non-consensual web audio used to train the models. - [10:00](https://www.youtube.com/watch?v=QVXfcIb3OKo&t=600s) An audit of royalty-free loop services with Venus Theory showed that assembling tracks from Splice, Loopcloud, and vintage sample CDs triggered high rates of automated Content ID collisions, offering no safe option among tested platforms. - [13:19](https://www.youtube.com/watch?v=QVXfcIb3OKo&t=799s) Streaming payout analysis reveals major DSPs pay near the bottom—Spotify at $0.0029 and YouTube Music at $0.0027 per stream—while Qobuz reaches $0.0136 and curated fitness provider Peloton tops the list at roughly $0.03.

## Key takeaways - [01:51](https://www.youtube.com/watch?v=QVXfcIb3OKo&t=111s) A lucrative gray market exists where users scrape generative AI outputs from tools like Suno and distribute them to DSPs under verified artist personas to siphon streaming royalties. - [01:51](https://www.youtube.com/watch?v=QVXfcIb3OKo&t=111s) Over 50 million tracks were registered via ISRC within a 12-month span, marking a sharp surge in unvetted independent uploads while major label registrations shrank. - [05:05](https://www.youtube.com/watch?v=QVXfcIb3OKo&t=305s) Psychoacoustic compression discards 75% to 90% of audio information that human brains filter out, making lossy 128-256 kbps files virtually indistinguishable from lossless masters to most listeners. - [05:05](https://www.youtube.com/watch?v=QVXfcIb3OKo&t=305s) Generative neural networks ingested lossy web audio as a black box, permanently baking the codecs' compression anomalies into synthesized outputs as an acoustic delta fingerprint. - [08:08](https://www.youtube.com/watch?v=QVXfcIb3OKo&t=488s) Commercial Content ID architectures are engineered for wholesale audio piracy and cannot differentiate between legitimate shared royalty-free licenses and copyright infringement. - [08:08](https://www.youtube.com/watch?v=QVXfcIb3OKo&t=488s) Sample distributors lack the economic incentive to resolve Content ID clashes because subscription and distribution fees are collected upfront before downstream collisions occur. - [10:00](https://www.youtube.com/watch?v=QVXfcIb3OKo&t=600s) Empirical testing showed commercial loop libraries frequently triggered false copyright matches, with vintage Akai sample CDs producing the highest collision risk. - [10:00](https://www.youtube.com/watch?v=QVXfcIb3OKo&t=600s) Splice's premium discovery features, such as 'Rare Finds' tags and 'Splice Stacks', failed to prevent automated copyright strikes on short test compositions. - [13:19](https://www.youtube.com/watch?v=QVXfcIb3OKo&t=799s) Social media platforms provide negligible streaming revenue, with VK paying $0.00005 per stream and TikTok generating just $0.0021. - [13:19](https://www.youtube.com/watch?v=QVXfcIb3OKo&t=799s) Spotify and YouTube Music rank near the bottom of traditional music DSP payouts at under $0.003 per stream, whereas Apple Music pays double and Amazon Unlimited roughly 3.5 times as much. - [15:00](https://www.youtube.com/watch?v=QVXfcIb3OKo&t=900s) Independent Spotify payouts have declined by 10.9% over the past year and dropped 68% over the last decade when adjusted for inflation. - [15:00](https://www.youtube.com/watch?v=QVXfcIb3OKo&t=900s) Peloton stands as the highest-paying digital streaming platform at roughly $0.03 per stream due to its closed ecosystem and hardware-tied direct licensing model. - [17:37](https://www.youtube.com/watch?v=QVXfcIb3OKo&t=1057s) The channel relies on non-profit status, media liability insurance, and direct Patreon backing to publish investigative exposés without fear of corporate legal intimidation or brand sponsor loss.

## Detailed walkthrough ### [00:00](https://www.youtube.com/watch?v=QVXfcIb3OKo&t=0s) Introduction and the Music Data Experiments

Benn Jordan introduces the video by acknowledging that the presentation might feel somewhat disjointed or all over the place, yet it is thoroughly packed with vital data for musicians and music enthusiasts alike. Throughout 2024, the single most frequent question he received was about how to support favorite artists through streaming—specifically, which streaming service actually pays artists the highest rate per stream. Benn notes that he has gathered extensive hard data on the topic, and the ultimate answer is completely unexpected; he bets that almost none of his viewers would be able to guess the top-paying service correctly.

Beyond streaming payouts, the video sets out to tackle two other major industry issues. First, Benn has developed a functional method to expose individuals who scrape tracks from AI music generators, pretend they wrote them, and distribute them to DSPs (digital streaming platforms) to hijack royalty pools meant for human artists. He reveals that he has developed an AI detection pipeline capable of identifying music generated via Suno with 100% accuracy. Second, he teamed up with fellow creator Venus Theory to audit sample and loop services. The two creators subscribed to every prominent loop platform, downloaded random samples, assembled nearly 100 songs, and routed them through an automated copyright fingerprinting API to determine which platform is most—and least—likely to cause false copyright claims. Benn humorously remarks: “Some guys like football. Some guys like motorcycles. Some guys like anime. I like to generate strange sounds on AI music sites by the thousands, then upload them with an API that compares them to previously copyrighted music, and then try and isolate reliable methodology that would stand up in court should, I don't know, an artist or record label say, 'Hey, assholes, you stole my work without asking and used it to raise hundreds of millions of venture capital dollars. How about you pay me now?'”

- Streaming payout analysis answering the most common 2024 audience inquiry. - Creation of an AI detection tool targeting platforms like Suno and Udio. - Empirical testing of royalty-free sample libraries alongside Venus Theory to audit automated Content ID collision rates.

---

[01:51](https://www.youtube.com/watch?v=QVXfcIb3OKo&t=111s) Identifying Fake Music on Spotify and the Gray Market

Benn explains that his investigation into generative music began years prior, dating back to experiments around the release of Google's Magenta project. More recently, he began scraping songs created by users on platforms like Suno and Udio and cross-referencing them against major streaming services. His findings confirm a bustling and highly profitable gray market. While streaming services like Spotify operate under vague policies regarding artificial intelligence, they have little desire to distribute their shared royalty pool to works that lack legal copyright protection.

To gauge the scale of the issue, Benn scraped 560 of Suno's staff picks and top-ranked generated tracks. Astonishingly, out of those 560 songs, only 11 were *not* currently monetized on commercial streaming platforms under the guise of an authentic human artist. Verifying these matches was grueling because he had to listen and compare them by ear to avoid falsely identifying legitimate human musicians. During this process, he found that bad actors deliberately obfuscated song and artist names to evade detection. Many of these accounts were fully "Verified Artists" on Spotify, possessing extensive discographies and algorithmic placement like "Artist's Pick." Inspecting one such account, Benn observes a single AI-generated song with 688,946 streams—translating to roughly $2,000 extracted from the streaming royalty pool.

Analyzing ISRC (International Standard Recording Code) data reveals the sheer scale of the dilution: daily new song registrations showed a 15% bump by June compared to preceding years, signaling that over 50 million songs were uploaded to platforms like Spotify within the prior 12 months. Concurrently, label registrations are shrinking while independent uploads are skyrocketing. Benn dryly laments: “'Independent artists representing a larger portion of released music is really good news,' said the version of me in a parallel universe where AI music isn't flooding into streaming services.”

- 549 out of 560 top Suno tracks (98%) were found actively monetized on digital streaming platforms under fake artist personas. - A single identified track logged 688,946 streams, siphoning approximately $2,000 from legitimate creators. - Industry ISRC data indicates an upload volume surpassing 50 million tracks annually, driven by an artificial explosion of unvetted "independent" releases.

---

[05:05](https://www.youtube.com/watch?v=QVXfcIb3OKo&t=305s) Psychoacoustic Compression and Building an AI Music Detector

```mermaid flowchart TD accTitle: AI Music Fingerprinting Flow accDescr: Diagram showing how lossy artifacts from web scraping create detectable acoustic fingerprints in AI models.

subgraph Ingestion ["1. Training Ingestion"] A["Scraped Audio<br/>(YouTube, Spotify, SoundCloud)"] B["Lossy Compression<br/>(7:1 to 10:1 Ratio / 128-256 kbps)"] A --> B end

subgraph Generation ["2. Black Box Synthesis"] C["Neural Network Model<br/>(e.g., Suno)"] D["Baked-in Artifacts<br/>(Unnoticed by Human Hearing)"] B --> C C --> D end

subgraph Detection ["3. Fingerprint Extraction"] E["Delta Identification Algorithm"] F["100% Detection Accuracy<br/>(Model / Source Verification)"] D --> E E --> F end ```

Benn shifts into the physics of sound and the history of audio distribution, recounting how MP3 encoding revolutionized audio sharing in his late teens through psychoacoustic compression. In modern environments, humans struggle to perceive what has been removed from compressed files. Benn admits that even on his highest-grade studio monitors, dedicated headphone setups, and amplifiers, he cannot reliably differentiate a 256 kbps MP3 from an uncompressed 24-bit WAV file. He points to an NPR double-blind test where 80% of participants could not tell a 128 kbps MP3 apart from a CD-quality lossless file, with most listeners ironically choosing the lowest-bitrate file as the best sounding. Streaming services rely on this psychoacoustic exploit, maintaining compression ratios between 7:1 and 10:1 (turning a 50 MB master into a 5 MB to 8 MB file) by discarding 75% to 90% of audio information that human brains filter out via latent inhibition.

This psychoacoustic blind spot forms the foundation of Benn’s AI detection breakthrough. Generative platforms like Suno (valued at $500 million) scraped enormous quantities of training audio non-consensually from websites like YouTube, SoundCloud, and Spotify. Because these repositories consist almost exclusively of lossy, compressed audio files, the neural networks ingested the specific artifacts and limitations of those codecs:

> “My point here is that 75 to 90 percent of information that humans don't realize is missing when listening to compressed audio could potentially be used as a fingerprint when identifying if something was created with generative AI and potentially even identifying the content that was scraped in the dataset.”

Because neural networks treat their input data as a black box, they bake these compression anomalies directly into their generative output. Benn successfully isolated this residual "delta" fingerprint in music generated by Suno, achieving a 100% detection rate in his initial tests. To prevent generative companies from updating their models to circumvent detection, Benn intentionally keeps the precise algorithm confidential, sharing the research only with trusted peers to validate and refine the model.

- Double-blind testing shows 4 out of 5 individuals cannot separate 128 kbps compressed files from lossless source material. - Streaming compression strips out roughly 75% to 90% of raw audio information. - Non-consensual web scrapers ingested lossy, compressed audio, leaving distinct mathematical artifacts inside the models' generative output. - Benn’s custom tool leverages these structural artifacts to identify Suno-generated tracks with 100% reliability.

---

[08:08](https://www.youtube.com/watch?v=QVXfcIb3OKo&t=488s) Copyright Auditing Splice and Sample Libraries

Turning to music production workflows, Benn explores the systemic flaws within digital sample libraries, directing viewers to Venus Theory's prior video addressing the "Splice problem." When artists assemble tracks out of commercially available, royalty-free sample packs, Content ID systems on YouTube and Spotify misinterpret the shared elements. Because automated copyright monitors are designed to catch wholesale piracy—such as an unauthorized re-upload of Pink Floyd's *The Dark Side of the Moon*—they cannot distinguish between legitimate, shared royalty-free licenses and blatant copyright infringement. When two independent producers release songs built around the same underlying sample, the Content ID system flags one creator as a fraud, frequently blocking their music or redirecting their hard-earned monetization straight to the other artist.

Benn emphasizes that this widespread industry friction is completely preventable through common-sense technical systems. Streaming and hosting platforms could easily establish an open whitelist for registered sample identifiers, or sample providers like Splice could register ownership of their catalog with Content ID prior to public distribution while leaving monetization claims turned off. However, because sample distributors and digital aggregators collect their subscription and distribution fees up front, they lack the economic incentive to resolve Content ID clashes that occur months downstream:

> “Unfortunately, if you run into this problem with loop or sample services, it will be long after you've paid the loop library and the distributor, so there's not much incentive for anybody to do much about it.”

To determine which sample platform causes the most severe copyright issues, Benn and Venus Theory put the major players to the test. Together, they subscribed to every leading loop library, constructed nearly 100 original tracks using random loops, and submitted them through a commercial Content ID API to measure how often false claims occur across different catalogs.

[10:00](https://www.youtube.com/watch?v=QVXfcIb3OKo&t=600s) Auditing Sample Libraries: Splice, Loopcloud, and Content ID Traps

To test how frequently commercial sample libraries trigger Content ID claims, Benn Jordan and Venus Theory generated roughly one-minute test tracks using loops that were pitched and time-stretched to aesthetically fit standard production workflows. Loopcloud proved technically frustrating during testing, repeatedly crashing in both its standalone app and DAW bridge. Furthermore, sample services often monetize safety features: accessing Splice’s filtering tools for undiscovered or Content ID-safe sounds requires paying extra for a Creator or Creator+ tier rather than base subscriptions like Splice Sounds+ or Splice x Studio One Pro.

To compare services, Benn calculated a relative flag rate score where a higher number denotes a higher risk of triggering a copyright match against existing published music:

| Sample Library / Source | Songs Created | Copyright Flags Received | Relative Risk Score | |:--- |:--- |:--- |:--- | | **FL Studio Cloud** | 14 | 18 | 128 | | **Loopcloud** | 18 | 28 | 155 | | **Splice** (standard browse) | 16 | (high rate) | (high score) | | **Vintage Akai Sample CDs** (Zero-G, Ueberschall) | 9 | 29 | 322 |

Reflecting on the findings, Benn concludes: “In my opinion, there is no real winner here. These are all really troubling numbers that should probably make you reconsider how you use loops.” He suggests using a loop strictly as initial inspiration, or composing around it and deleting it before final export.

Testing Splice’s specialized workflows yielded similarly poor outcomes: - **Rare Finds tag**: Designed to surface sounds undiscovered by the broader user base, four test songs built exclusively from these loops still racked up 6 copyright flags. - **Splice Stacks**: A feature that uses internal BPM and harmonic tagging to randomly blend loops produced 8 copyright pings across 10 generated tracks (most under 20 seconds long). Benn notes the slightly lower hit rate was likely due purely to the brevity of the clips. - **Tracklib**: Excluded from the test because it operates as a rights-holding entity backed by investment groups, leasing catalog masters and combining publishing holdings with work-for-hire music. Users must explicitly register sample use and share backend streaming royalties, meaning flags on Tracklib tracks generally reflect paperwork failures rather than algorithmic detection collisions.

---

[13:19](https://www.youtube.com/watch?v=QVXfcIb3OKo&t=799s) Streaming Royalty Pay-Per-Stream Breakdown

Assuming an artist navigates copyright hurdles and lands a hit, the question turns to streaming remuneration. Because Benn owns 100% of his discography—having bought back his vintage catalog masters and controlling all mechanical and publishing rights—his distributor data reflects direct per-stream royalties for independent musicians, verified alongside peers with similar complete catalog ownership.

Social media platforms pay fractions of a cent per stream: - **VK (VKontakte)**: $0.00005 per stream (requiring 200 streams to earn a single penny) - **Snapchat**: $0.0001 per stream - **TikTok**: $0.0021 per stream - **Meta (Facebook/Instagram)**: $0.003 per stream

Traditional streaming platforms fare marginally better, though major services sit near the bottom: - **YouTube Music**: $0.0027 per stream (tied with Pandora) - **Pandora**: $0.0027 per stream - **Spotify**: $0.0029 per stream - **Apple Music**: $0.006 per stream (more than double Spotify's payout) - **Deezer**: $0.0078 per stream - **Amazon Music Unlimited**: $0.0096 per stream (roughly 3.5 times Spotify's rate) - **Qobuz**: $0.0136 per stream

```mermaid flowchart LR accTitle: Pay Per Stream Hierarchy accDescr: Streaming platforms ordered from lowest payout per stream to highest payout per stream.

subgraph Bottom["Bottom Tier (< $0.003)"] VK["VK: $0.00005"] --> Meta["Meta: $0.003"] Meta --> YT["YouTube / Pandora: $0.0027"] YT --> Spot["Spotify: $0.0029"] end

subgraph Mid["Mid Tier ($0.006 - $0.0096)"] Apple["Apple Music: $0.006"] --> Deezer["Deezer: $0.0078"] Deezer --> Amazon["Amazon Unlimited: $0.0096"] end

subgraph Top["Top Tier (> $0.01)"] Qobuz["Qobuz: $0.0136"] --> Peloton["Peloton: ~$0.03"] end

Spot --> Apple Amazon --> Qobuz ```

---

[15:00](https://www.youtube.com/watch?v=QVXfcIb3OKo&t=900s) The Peloton Anomaly and the Long-Term Decline of Streaming

Qobuz represents the rare mainstream threshold where royalties cross one cent per stream, giving independent creators a semblance of fair compensation. However, Benn remains skeptical about its sustainability: “I don't think that optimism has ever rewarded me in this industry, and I'm sure like everybody else, as soon as they get a decent market share, they'll start wringing the towels and we'll see those streaming rates go down just like they did with Spotify and everybody else.” Historical tracking of his own Spotify payouts illustrates this erosion: payouts dropped 10.9% in the past year, declined over 50% across the last decade, and represent an inflation-adjusted 68% drop—meaning artists earn less than one-third of what they accepted ten years ago.

The single highest-paying digital streaming platform in Benn's reports is **Peloton**, paying approximately $0.03 per stream. Peloton operates a closed, curated, and directly licensed catalog for its fitness hardware. If Benn’s entire listener base hypothetically purchased a $1,500 exercise bike and paid the $40 monthly Peloton subscription, his music would generate around $6 million annually. Benn argues that Spotify and dominant DSP models remain toxic to musical sustainability, urging listeners to cancel predatory subscriptions, transition to higher-paying alternatives, or back artists directly via platforms like Patreon.

---

[17:37](https://www.youtube.com/watch?v=QVXfcIb3OKo&t=1057s) Channel Independence, Investigative Journalism, and Patreon

Investigative reporting on this channel—tackling issues like Spotify bot farms, predatory ADHD tele-health pharmaceutical firms, and automotive engineering practices—demands resources and significant legal risk that traditional ad-supported media cannot sustain. Major corporate brands like BetterHelp, MyHeritage, or NordVPN do not sponsor undercover recordings of pharmaceutical executives, nor can mainstream platforms easily manage corporate blowback.

- **Non-Profit Structure**: The channel operates beneath a non-profit umbrella funded entirely by direct viewer support. - **Legal Defense**: Community funding covers independent research and essential media liability insurance, allowing Benn's legal representatives and insurance providers to handle corporate cease-and-desist letters and aggressive legal threats. - **Community Benefits**: The Patreon hub distributes released and unreleased audio, professional sound libraries, and access to a creative community without corporate interference.

## Notable quotes > [05:05](https://www.youtube.com/watch?v=QVXfcIb3OKo&t=305s) 75 to 90 percent of information that humans don't realize is missing when listening to compressed audio could potentially be used as a fingerprint when identifying if something was created with generative AI and potentially even identifying the content that was scraped in the dataset. — Benn Jordan

## Mentioned Benn Jordan, Venus Theory, Spotify, Suno, Udio, Google Magenta, Splice, Loopcloud, FL Studio Cloud, Tracklib, Akai, Zero-G, Ueberschall, YouTube Music, YouTube, SoundCloud, Apple Music, Amazon Music Unlimited, Qobuz, Deezer, Pandora, Peloton, TikTok, Meta, Snapchat, VKontakte, Patreon, BetterHelp, MyHeritage, NordVPN, Pink Floyd

## Video description What streaming service pays the best? What sample library has the worst copyright problem? How to tell if music is human-made or not. It all gets answered here, folks!

The Splice Problem: https://www.youtube.com/watch?v=pbOop_za_OY&pp=ygUbdmVudXMgdGhlb3J5IHNwbGljZSBwcm9ibGVt

Udio Generated Music: https://open.spotify.com/playlist/0lPkLx6bo2GegvWPy8OUZN?si=aD7auMu2Q5iXSfTV-xLl4g

Suno Generated Music: https://open.spotify.com/playlist/7pRdrM4ZyQFStXGGowzRGk?si=MkvsqPPQQ-qBo9k-4pXG0A

💗 Support this channel and join an amazing community: http://www.patreon.com/bennjordan

👀 Stalk me on social media for more frequent updates: https://linktr.ee/BennJordan

🔴 Subscribe To My Streaming Channel. I stream weekly! https://www.youtube.com/watch?v=geaIEWGhgsI

⚡Those lovely custom acoustic panels in the background: https://psyacoustics.com

timestamps: 0:00 - Intro 1:51 - Identifying Fake Music On Spotify 5:05 - Building an AI Music Detector 8:08 - Copyright Auditing Splice (and others) 13:19 - Which streaming service pays the most? 17:37 - The Future Of This Channel

[Source](https://www.youtube.com/watch?v=QVXfcIb3OKo)

中文简报

利用AI检测AI音乐（以及其他音乐行业硬核数据） Benn Jordan 对现代音乐行业面临的困境展开了实证深入研究。首先，他介绍了一套AI检测流程，该流程利用心理声学压缩失真特征，识别出抓取自Suno并在流媒体平台上以虚假艺术家身份进行变现的曲目。随后，他与 Venus Theory 合作，对 Splice、Loopcloud 以及老式采样CD等主流免版税采样库进行了审查，结果表明自动 Content ID 系统在所有平台上都产生了惊人比例的版权误判。最后，Jordan 借助其全资目录的独立发行商数据，详细拆解了现实中的单次流媒体播放版税，揭示出 Spotify 等平台几乎垫底，而 Qobuz 以及 Peloton 等精心策划的健身服务平台支付的报酬则高得多。对560首排名前列的Suno曲目进行抓取分析后发现，有549首（98%）正在数字流媒体平台上以伪装的艺术家身份进行变现，其中单首曲目就从共享版税池中获取了近70万次播放量。 (1:51) Benn Jordan 开发了一套对Suno音频准确率达100%的AI检测流程，其原理是识别训练模型所用的未经授权的有损网络音频中遗留下来的残留压缩失真。 (5:05) 他与 Venus Theory 对免版税循环乐段服务的审查显示，使用 Splice、Loopcloud 和老式采样CD拼接出的曲目引发了极高比例的自动 Content ID 冲突，所测试的平台中无一幸免。 (10:00) 流媒体分成分析表明，主流流媒体平台给出的单次播放收益垫底——Spotify为0.0029美元，YouTube Music为0.0027美元；而Qobuz可达0.0136美元，精心选曲的健身服务提供商Peloton更是以约0.03美元位居榜首。 (13:19)

Source quotes

[5:05] Benn Jordan: 75 to 90 percent of information that humans don't realize is missing when listening to compressed audio could potentially be used as a fingerprint when identifying if something was created with generative AI and potentially even identifying the content that was scraped in the dataset.

Original transcript

[0:00] This video is going to be kind of all over the place, but if you're a musician or even a music enthusiast, it's frankly jam-packed full of really useful information. For example, the most common question that I've gotten in 2024 by far is "What streaming service should I use if I want to support my favorite artists?" or more specifically, what streaming service pays artists the most? And I have loads of hard data to share with you on this, and the answer is, um, really unexpected. to guess which music streaming service pays the most per stream, and I can almost guarantee that not one of the viewers of this video will have guessed correctly. Another pretty exciting thing in this video, I've sussed out a way to catch and identify musicians and artists who download music off of AI music generator sites, claim it as their own, and then put it on Spotify and other streaming services to siphon royalties away from, you know, actual musicians writing their own music. an AI model that, so far, has been able to detect if a song has been generated by AI on Suno with 100% accuracy. So we're going to talk about that. What else? Also, myself and Venus Theory subscribed to every major loop and sample platform, downloaded a bunch of random loops, and made a whole bunch of songs, like close to 100 ultimately. Then I organized and ran them all through the same which of the bunch is the most and least likely to get you copyrighted due to other artists using the same loops or samples. And the answer that we found will blow your part of your brain that processes obvious things because it's not actually that surprising. So let's um... I guess this video is a little bit too disheveled or disorganized to have a title sequence. Some guys like football. Some guys like motorcycles. Some guys like anime. I like to generate strange sounds on AI music sites by the thousands, then upload them with an API that compares them to previously copyrighted music, and then try and isolate reliable methodology that would stand up in court should, I don't know, an artist or record label say, "Hey, assholes, you stole my work without asking and used it to raise hundreds of millions of venture capital dollars. How about you pay me now?" But really, this is a long work in progress that I've been doing intermittently since Google's Magenta came out. scraped music other people were writing on Udio or Suno, and then seeing if I could find the AI-generated music on streaming platforms. And, uh, yeah, it turns out that this is actually a pretty active and lucrative gray market. And I say "gray market" because most music streaming platforms have rather ambiguous terms about AI, but I would assume that a tech company like Spotify doesn't care if music was made by humans or not, but would also not be interested in paying royalties on bodies of work that aren't copyrightable. In my experiment, I exhaustively draped 560 of Suno's top-generated songs and staff picks, and out of those 560, only 11 were not being monetized by posing as a human artist on a digital streaming platform. I certainly don't want to call out a legitimate human artist who got copyright claimed. I'm going through this process to make sure that they are using UDO or Suno to upload music to Spotify. I'm doing it by ear, which is not fun. blue thoughts so they intentionally obfuscated the name and claimed that it's their own and what's even worse is that in the vast majority of cases the artist and song names were changed and the musician was a verified artist on spotify and then of course the artist is verified they even have their pick their new song that you could listen to their discography has a lot of albums huh wow look at all this shit jesus christ so this song for example has 688 946 plays That's significant. So if you're wondering, this is like $2,000 earned from the royalty pool. Now, when you upload or release a new song to a digital streaming platform, you get something called an ISRC number. And while the full 2024 stats haven't been released yet, if you analyze the ISRC registration data, it looks like, as of last June, that there was already a 15% increase in new songs uploaded to streaming services per day over the previous few years. the first half, that suggests that over 50 million new songs have been uploaded to Spotify in just the last year. And another piece of data worth considering is that music being registered from record labels is dropping and music being registered from independent musicians is rising exponentially. "Independent artists representing a larger portion of released music is really good news," said the version of Me in a Parallel Universe where AI music is unflooding into streaming services. And the universe that you and I share, unfortunately, If you're curious, [5:00] I quickly whipped up a few Spotify playlists of music that has been 100% generated on Suno or Udio. So as a sort of useful precursor, let's talk about MP3s for a second. The technology was quite the rage when I was in my late teens, but these days I feel like kids just don't appreciate the magic of psychoacoustic compression. You know? Let's just get this hard-to-digest fact out of the way first. Chances are that the vast majority of the time, you cannot tell the difference between a lower-quality MP3 and a lossless, uncompressed WAV file or CD. Me personally, on my finest monitors or headphones or amplifiers, cannot reliably tell the difference between a 256kbps MP3 file or a lossless file. An NPR a few years ago did a fun double-blind study, and 4 out of 5 people couldn't tell the difference between a 128kbps MP3 file and a lossless file, with the majority of the people selecting the lowest quality file as the one that they thought was the one with the highest quality. The point is, unless you are an extremely rare exception, you usually cannot tell the difference between compressed and lossless audio. So why are we talking about this? Well, when you listen to music here on YouTube or on Spotify or something you listening to a heavily compressed file usually between 7 to 1 and 10 to 1 meaning that if the lossless audio file is 50 megabytes the size of the streamed audio file is ultimately like 5 megabytes to 8 megabytes. So when you consider that the vast majority of people don't notice the difference when up to 90% of the information is missing, the amazing psychoacoustic magic trick of audio file compression becomes fully appreciated. Ultimately, our brains are just collecting pressure waves and then passing them through latent inhibition and then only really registering data that's useful to us. And audio-visual file compression has managed to make a surgical art out of exploiting that distinctly human information processing workflow. Earlier this year, it dawned on me that very little of what generative AI companies were non-consensually scraping off the internet was uncompressed or lossless. They scraped things like YouTube and Spotify and SoundCloud to make their datasets generate AI music. But a neural network of this type, even if it's a convolutional neural network, is a black box. You can't tell it to note these limitations in a functional way. Even if you could, you can't technically describe the differences you might hear in compressed files. You just say, it's glitchy sounding, or it sounds a little bit off, and that's useless information in machine learning My point here is that 75 to 90 of information that humans don realize is missing when listening to compressed audio could potentially be used as a fingerprint when identifying if something was created with generative AI and potentially even identifying the content that was scraped in the dataset. For obvious reasons, I'm not going to get too technical about my methodology, but I've managed to find the delta, or the isolated fingerprints of music generated on Suno. And I assume that this will be possible on other services as well, but I went for the one with the $500 million valuation first. That first initial test had 100% accuracy, so I had a call with some friends that can improve and further test it who I trust to not publish or leak the information that would make circumvention easier. I would not personally consider myself a sample flipper or loop pack user for my own music, but it's safe to say that if you make music on a computer, then you're probably familiar with loops and sampling. A few years ago, Venus Theory made an excellent video about the SPLICE problem. You should definitely watch that video. I'm going to link it in the description. Also the SPLICE problem isn't solely limited to SPLICE, by the way, that's just a side effect of market saturation. So if you use a service like SPLICE and you download a bunch of loops and put them together and make a song there a pretty strong chance that someone else has used some of those loops for their own song And you both have the legal right and license to use it but YouTube and Spotify and whoever else doesn know that As you probably know these sites have content ID systems that prevent you from just uploading Dark Side of the Moon and claiming royalties on it, but in this case, these systems are assuming that you stole music from the other Splyce user and they'll either block you for copyright fraud or, in YouTube's case, give some or all of your royalties to the other artist. often viewed as this unforeseen problem that is mostly unpreventable when it's completely and entirely preventable. The first and most obvious choice would be for platforms like YouTube and Spotify to recognize a whitelist for particular data and not flag it at all. Another workaround would be for a company like Splice to register ownership of all of their content before making it public and then not claiming any money or attribution for it. This mechanism works nearly flawlessly. Unfortunately, if you run into this problem with loop or sample services, it will be long after you've paid the loop library and the distributor, so there's not much incentive for anybody to do much about it. But I suppose if somebody had enough time on their hands, they could subscribe to all of the major loop library services and make close to 100 songs on them and then run them through a massive Content ID API and see which service has the lowest chances of landing a musician with a copyright claim headache. I wonder who would do such a thing? They specifically have a section so you don't [10:00] trigger content id yeah if you want access to the filter for undiscovered sounds or content or sounds that aren't as likely to trigger content id you do have to pay extra or you do have to get a creator or creator plus plan not a splice x studio one pro or sounds plus plan so loop cloud just keeps crashing it just keeps crashing the the standalone yeah and i i'm trying to use the bridge Generally, all the songs that we made were about a minute or so long, and the loops were pitched in time-stretched to fit in a way that just aesthetically worked for the song, which is what I assume most people do when they're using loops. So how did they do? Let's go from best to worst and use a little bit of algebra to get us a comparable score. And the higher the score, the more likely flagged for using a sample or loop on that service. I made 14 songs using loops from FL Studio Cloud, and that got 18 copyright matches with other published music, which scores us at 128. Venus Aria and I made 18 songs combined using Loop Cloud, which earned us 28 matches or copyright flags, which scores us at 155. I made 16 songs on Splice, just picking whatever I wanted for my nine quick songs using old Akai sample CD libraries from loop companies like Zero-G and Ubershaw, and that got me 29 matches, so 322. In my opinion, there is no real winner here. These are all really troubling numbers that should probably make you reconsider how you use loops. For example, if you're using a 10-second melody, perhaps use it as inspiration for your own melody, or write around it and then remove it later in your session. Or don't. I don't care. Okay, a few more things that you might find interesting. rare finds, a tag that will provide you with a curated list of sounds still undiscovered by most users. Using those exclusively in four songs got me six copyright flags. still pretty bad. Also interestingly, Splice has a new feature called Stacks, which uses their own BPM and Harmony tagging system to randomly combine loops to give you some interesting options. Anyway, I churned out 10 loops with this, mostly under 20 seconds in length, and got eight copyright pings. And I initially was surprised that this was faring better than even the rare finds were, but ultimately I think that the reason was because the songs were just much shorter. Finally, You might be wondering why I didn't include Tracklive, and that's because Tracklive operates completely differently. They're actually a holding company with investors from other holding companies that buys up or leases the licenses of music as a general investment. The website, Tracklive, combines those publishing holdings with Music for Hire to, ideally, make additional money on those holdings. And most of the things that you get on Tracklive requires you to register its use and, in some cases, split a small amount of streaming royalties. So unless there's an error of some sort if your music is getting flagged for using Tracklive didn't fill out your forms. So let's say that you get past the content ID and licensing nightmare and your song blows up How much are you gonna get paid? So I own 100% of my own music. Anything that was released on a label years ago I bought back. I own the publishing, the mechanical, everything. So these numbers that I'm about to give you are about as accurate as you could possibly get if you're looking for the pay per stream rate for independent musicians. I also ran these numbers by a few friends who also own all of their discographies and they were very similar. Alright, let's go from the bottom up. reporting, much less paying anything, or literally not one person listened to my music on them. Alright, so let's get into the services that did pay me, and unsurprisingly, social media is at the bottom of that barrel. The Russian social media site VK is paying me $0.00005 per stream, which means that I need 200 streams to get a single penny. Snapchat comes in at $0.0001. Meta is $0.003 per stream, and TikTok is a little bit higher at $0.0021. All of these are worse than every music streaming service on this list. Tied in their shame is YouTube Music and Pandora at $0.0027, and predictably near the bottom is Spotify with $0.0029. Apple Music is widening the margin a bit by paying more than double what Spotify pays at $0.006 per stream. Deezer is with a .0078. Now for the top three, all of which are very surprising to me. Paying a whopping $0.0096 per stream is Amazon Unlimited Music. [15:00] you want about Amazon, but from my data, it appears that they are paying like 3.5 times more than Spotify is. Coming in at number two is Kobuz, which is a French streaming service that I had never heard of until a few months ago. They're paying me .0136 dollars per stream, crossing the penny per stream threshold, which I think is the first time I ever looked at streaming royalties and felt fairly compensated. But then again, I don't think that optimism has ever rewarded me in this industry, and I'm sure like everybody decent market share, they'll start wringing the towels and we'll see those streaming rates go down just like they did with Spotify and everybody else. Speaking of that, here's Spotify's rates over time since I personally signed up. It dropped 10.9% from last year and over 50% in the last decade. If you adjust that for inflation, that would be a 68% drop in pay per stream, meaning that in 2024, artists got paid less than one-third of the price years ago, and they agreed to put their music there. Okay, sorry. Number one, the highest paying digital streaming service in the world right now is Peloton, the expensive stationary bicycle company. So you can use Spotify and other DSPs on your Peloton bike, but they also have their own Peloton Music, which is a curated and licensed catalog. So it doesn't seem to be something that you can just join, nor is it something that the vast majority of your listeners will have access 3 cents per stream. Or look at it this way: if every single person in the world who streamed my music made the small, tiny sacrifice of buying a $1,500 stationary bicycle and paying $40 a month to listen to music on it, I'd be making a cool $6 million per year. Okay, seriously, if you're an active watcher of my content, I feel like over the last five years I've made a pretty solid, conclusive, well-researched case and argument for why and other big streaming services are poisonous for artists and the evolution of music. Maybe now is a good time to just open up your phone or web browser and cancel that shit. Consider signing up for something that more fairly pays the artists that you enjoy listening to before they do what I did and stop releasing anything on there due to lack of incentive. Or even better yet, see if they have something more direct and involved like Patreon. That absolutely changed the game for me and allowed in hopes that it would inspire some carrots and sticks to be distributed to those services and companies that deserve them. And hey, Speaking of Patreon, a little bit over a year ago, I decided that I was going to go all-in on this YouTube channel with my finances, with my time, and my reputation. And in that time, it's been full of deep dives and risky journalism and actively untethered science education. And some of these videos have performed tremendously. And that performance definitely got the attention of companies like Netflix. scaling the way I've been scaling. For example, the Spotify bot video or the ADHD video or the Tesla video have all resulted in other organizations or law firms taking note and pursuing ways to either combat bad things or improve the good things. This sounds cheesy and trite and you probably hear it all the time from me and you probably hear it from other YouTubers as well, but if you're one of my Patreon supporters, this is because of you. It could have not happened any other way. Do you think that better help? or MyHeritage or NordVPN would sponsor a video where I'm secretly recording the CEO of an ADHD drug company? F*ck no. As I've mentioned in other videos, this channel operates under the umbrella of a non-profit organization. And this last year or 18 months of this channel can only be done with direct viewer support. It's not only helped pay for research, but it's helped pay for media liability insurance. And I've received plenty of legal letters and threats, and all of them got insurance company. And I just wanted you to know that if you've been an active supporter of this channel, all of this is made possible by you. And I want to thank you from the bottom of my heart and my penis. And if you want to join my Patreon, it has plenty of released and unreleased music and pro audio samples and yada yada. And more importantly than all that, it has a really healthy community that actively inspires members to keep creating music and art and feeding their passions. Thank you for watching. Keep creating. Bye.