# First U.S. Entertainment AI Fraud Sentence Hinged on Bots > On October 6 a federal judge sentenced Michael Smith to 18 months in prison for using bot accounts to stream AI-generated songs, in what both sides call the first U.S. criminal prosecution of its kind, and the crime was the fake listening. The same week, a dancehall veteran described how a distributor's AI filter flagged his human-made track, new provenance research showed the limits of after-the-fact detection, and Italy opened a probe into Suno's terms of service. For artists who use AI and those who do not, dated records of how a track was made are becoming the strongest protection. Content type: article Source URL: https://www.agentpmt.com/articles/first-u-s-entertainment-ai-fraud-sentence-hinged-on-bots Markdown URL: https://www.agentpmt.com/articles/first-u-s-entertainment-ai-fraud-sentence-hinged-on-bots?format=agent-md Updated: 2026-10-07T14:27:07.096Z Author: Pancakes Tags: AI Agents In Business, AgentPMT, Authentication For AI, Security In AI Systems, News --- # What the First U.S. AI Music Fraud Sentence Means for Artists Who Use AI Michael Smith is going to federal prison for 18 months. Judge John G. Koeltl handed down the sentence in Manhattan on October 6, added two years of supervised release, and ordered Smith, 54, of Cornelius, North Carolina, to forfeit the full loss figure both sides had agreed on. Smith pleaded guilty in March to one count of conspiracy to commit wire fraud, and prosecutors and the defense have both described the case as the first criminal prosecution of its kind in the U.S. From 2017 to 2024, according to prosecutors, Smith used thousands of fake accounts on Spotify, Apple Music, Amazon Music, and YouTube Music to stream hundreds of thousands of AI-generated songs billions of times. The songs were cheap to make. The audience was entirely invented, and the audience is what put him in front of a judge. The detail for working musicians surfaced years before the courtroom. Spotify told Music Business Worldwide in 2024 that its preventative measures had limited Smith's earnings from the platform to about $60,000, a sliver of the total alleged at the time. Fraud detection spotted a fake audience long before prosecutors arrived, and the Mechanical Licensing Collective (The MLC), which pays publishing royalties, halted his payments in 2023. That screening now sits in front of every release, AI-assisted or human-made, and the artists who move through it fastest are the ones who can show how their tracks were made. ## A wire fraud case with a very large catalog Streaming services such as Spotify pay rights holders from a pool tied to revenue, split by each track's share of all streams in a period, so a fake stream shrinks the slice of everyone paid from that pool. The Music Fights Fraud Alliance (MFFA) told the court that "\[e\]very fraudulent stream perpetrated by Mr. Smith diluted the value of every legitimate stream occurring during the same accounting period." Generative AI gave Smith a catalog large enough to hide in. Prosecutors quoted a 2018 email in which he wrote that he needed "a TON of content with small amounts of Streams" to avoid attention, since spreading fake plays across that many tracks keeps any single song from looking like a hit. It works until someone looks at the accounts. In April 2023, prosecutors said, Smith's bots used family plans to stream his AI-generated music 80.9 million times on YouTube Music, while Taylor Swift's entire catalog drew 9.3 million family-plan streams there that month. An unknown catalog outdrawing Swift on a product designed for households is precisely the pattern behavioral fraud systems are built to flag. U.S. Attorney Jamie McDonald said that "by flooding music streaming platforms with automated bots in the place of consumers, and fake songs in the place of creativity, Smith robbed millions in royalty payments from genuine artists and their fans." He also said Smith "exploited super intelligence technology to generate a fraud," a vocabulary upgrade from the plain "artificial intelligence" his office used at the March plea. Smith's lawyers argued that "it is not inherently unlawful to use AI to create songs and place those songs on streaming platforms," citing a September 2025 Spotify statement that for royalty purposes "all music is treated equally, regardless of the tools used to make it." The count Smith pleaded to concerned deceiving platforms with bot accounts. The sentence puts a price on faking listeners and leaves the old seminar question, can artificial intelligence be creative, exactly where it was. IFPI, the global recorded-music trade body, said the result "sends a clear message that streaming fraud is a crime and has serious consequences," and added the industry also needs to "prevent, detect and disrupt fraudulent activity." Detection is where this case reaches artists who never met Michael Smith. ## Catching fake listeners is easier than catching AI audio Behavioral detection asks whether a real person is listening, and the Smith case shows it working. Content detection asks whether a model made the audio, which drags the AI creativity vs human creativity argument out of the seminar room and into a distributor's review queue, and it gets that call wrong in ways artists are starting to document. Papa Michigan, the Jamaican dancehall veteran, found out ahead of his album's August release. TuneCore, the distributor handling _So Many Roads_, flagged track seven, "Who Can Be Against Me," saying its content review team had "identified indicators suggesting the use of generative AI tools that rely on datasets that are not properly licensed." Michigan says no AI touched the song at any stage. He posted video of the studio and the musicians who played on it, filed a formal appeal describing it as an entirely human studio recording, and asked for review by a human moderator. Then he pulled the song and released _So Many Roads_ on August 14 without it. The Gleaner's report does not say how the appeal ended. An automated call the artist says was simply wrong cost him a track on his own album. Evon Mullings, general manager of the Jamaica Music Society (JAMMS), told The Gleaner that "the scenario of some human-made music being rejected (false positive) by digital distributors is becoming increasingly common." Distributors are under pressure from the major streaming services to tag AI-generated and AI-assisted music before it reaches retail, he said, and have responded with aggressive automated detection. With drums locked perfectly to the grid and vocals under heavy Auto-Tune, he said, "detection algorithms can mistake the lack of 'human imperfection' for robotic generation." Producers have chased tight timing and pitch-perfect vocals for years. Now that polish can read as a machine's signature. His remedy is evidence. A flagged artist can appeal with "proof of creation," which he says includes "video evidence, high resolution DAW screenshots, or offer to send individual raw audio stems." A DAW, or digital audio workstation, is the software where a track is recorded and arranged; stems are the separate audio files for each instrument or vocal. Every item on that list is created during the session, which is why it outweighs any argument made after a rejection. Mullings adds that the major services want music made with AI labeled, as fully AI-generated or AI-assisted, rather than banned, though many keep it out of editorial playlists, recommendations, and charts. The label carries a cost. It is also a lane for creative uses of AI, which is more than a rejection notice offers. Impersonation adds pressure. Sony Music told the Financial Times it has filed more than 260,000 takedown requests against AI deepfakes of its artists, according to Music Ally. Those are requests, with no count of removals, and at that volume keeping the same content from reappearing means automated matching. "We will keep enforcing our artists' rights and expect platforms to act quickly and to prevent the same content from reappearing, but it's becoming an uphill struggle," the label said. ## Proving origin now has a research agenda Four arXiv preprints posted between October 2 and October 6 take on the same task: showing where media came from, and how far to trust the answer. The most useful frame comes from a provenance agenda by Zheng Gao, Xiaoyu Li, Zhicheng Bao, Yang Song, and Jiaojiao Jiang. It separates passive inference, a detector studying finished output and guessing, from message recovery, reading an embedded watermark, and from authenticated provenance, a signed record of production events. Its open questions include "hybrid local contribution," where part of a work is human and part machine, and it treats agent-composed workflows as a production route of their own. The paper covers images and video, but the hybrid case describes plenty of music too. In those terms, a distributor's AI filter is passive inference. Papa Michigan's studio video sits much closer to authenticated provenance, minus the signatures. A benchmark by Seyedmahdi Kazempourradi, Ramtin Mojtahedi, and Behrang Mohseni of Original Pictures Technologies tested the backups behind Content Credentials, the C2PA standard that attaches a signed record of how a file was made. When re-encoding strips those records, "soft bindings," an invisible watermark or a registry fingerprint, reconnect the file. The two failed on different copies, so neither covered everything alone, and fingerprints did better on audio. One caution travels beyond images: the detector of every TrustMark image-watermark variant fired on some images carrying no watermark at all, so the authors advise verifiers to check for the specific payload they expect. Voice cloning got its own paper. Weizhi Liu, Yue Li, Hui Tian, and Zhaoxia Yin introduced Thrive, which embeds identifiers into synthesized speech so a clip can be traced to the user who requested it, and reported high recovery accuracy even after the audio passed through reconstruction models that can degrade watermarks. That is attribution the provider builds in at the moment of generation. A compliance paper by Fernando Delbianco, Fernando Tohmé, and Hugo Acciarri addresses Article 50(2) of the EU AI Act, which requires providers of generative systems to make synthetic outputs machine-readable and detectable. Working on text watermarks, the authors argue strong watermarking cannot survive adaptive removal and propose detectors that report "watermark supported," "not supported," or "inconclusive," with separate error levels and a signed report aligned to the European Commission's 2026 Code of Practice. An "inconclusive" verdict sounds like a shrug. For an artist, it beats a rejection letter, because it admits the machine does not know and sends the case to evidence. Across all four papers, after-the-fact detection comes out as a probability, and records made during production come out as proof. As entertainment automation spreads, origin checks will not stay in music. Fortune's reporting on Asian film and video shows how quickly the role of AI in entertainment is widening there. "It's increasingly impossible not to use AI in video production," Danming Xie, chief AI scientist at iQiyi, told Fortune. "The conversation's shifting more toward how deeply creators use it." How deeply is precisely what a production record can show. ## Making origin evidence during the session The practical move for creators is to produce authenticated provenance themselves, while the work is happening. On AgentPMT, that starts with [Quantum-Safe File Attestation](https://www.agentpmt.com/marketplace/quantum-safe-file-attestation). It takes a file's SHA-256 hash, a short code computed from the file's exact contents that changes if a single bit changes, binds it to metadata and a timestamp, and signs the bundle with ML-DSA-65, the NIST-standardized post-quantum signature algorithm, inside a hardware security module. Anyone can verify the resulting certificate offline with open-source tools, without trusting AgentPMT's servers. The [Document and File Certification with Post-Quantum Digital Signatures](https://www.agentpmt.com/agent-workflow-skills/document-and-file-certification-with-post-quantum-digital-signatures) workflow makes it repeatable: your AI agent takes the file, runs the attestation, and hands back the certificate. For a producer, certify the session export, stems, rough mixes, and master as each is finished, and the release ships with a dated, tamper-evident trail showing those exact files existed at those times. Paired with studio video, that is a stronger version of the "proof of creation" Mullings describes, ready before any distributor asks. The boundary is narrow on purpose. An attestation proves a file existed in that exact form at that time. It does not prove a human played it, and it says nothing about what any third-party model was trained on. It turns "I recorded this in May" from a claim into a checkable fact about files and dates, which a passive detector can never supply. For AI-assisted work, the second record is the run itself. When AI steps run through an agent workflow, such as edits handled by the Video and Audio Editing Agent, [AgentPMT's audit logs](https://www.agentpmt.com/marketplace/agentpmt-audit-logs) record every action the agent takes, with the full request and response and a timestamp. That is the production-route record the Gao paper describes, and it lets an artist label a track "AI-assisted" with specifics: which step, which tool, which inputs. We have covered why [ordinary application logs fall short](https://www.agentpmt.com/articles/agent-logs-diary-you-need-flight-recorder-observability) and how [disclosure terms reached film and TV delivery contracts](https://www.agentpmt.com/articles/ai-entertainment-industry-now-runs-on-proof-of-consent). ## Who owns what the generator makes The same week, a regulator looked at the rights people keep in what they generate. On October 6, Italy's antitrust authority opened an investigation into Suno, saying clauses in its terms of service could create "a significant imbalance in the rights and obligations of users," Reuters reported. The regulator says the terms give Suno "broad discretion" to unilaterally change its contract, service, and subscription prices. It called the license users grant excessively broad, including a waiver of moral rights (a creator's rights to be credited and to object to distortion of a work) that may conflict with Italian copyright law, and flagged mandatory U.S. arbitration, a class-action waiver, and Massachusetts as the exclusive jurisdiction. A consultation open to business groups, chambers of commerce, and consumer associations follows in the coming weeks. Suno was not immediately available for comment. This is consumer protection applied on behalf of people who create with AI, which is what a maturing field looks like. The probe concerns one company's contract and has only just opened, yet the lesson travels: before you build a catalog on any generator, read what the license says you own, what you hand over, and where you would have to argue about it. We covered the training-data side in [what the AI music copyright cases settle for creators](https://www.agentpmt.com/articles/entertainment-ai-trained-on-21-million-songs-files-show). Choosing tools is where builders have the most control. AgentPMT routes model calls through a gateway, so a team can pick and swap the model at each workflow step without a rebuild, and the run record shows which model produced which element. TuneCore's notice drew its line at tools built on improperly licensed data, so when a distributor or a contract treats one generator's output differently from another's, that record answers the question. Reading each generator's license still falls to the artist. ## After the sentence, the paperwork Italy's consultation on Suno's terms opens in the coming weeks, the EU's Article 50 detectability duties are being worked out through the 2026 Code of Practice, and distributor filters will keep screening every upload meanwhile. For the artificial intelligence music industry, the week clarified where the lines sit: a federal court punished a fake audience, labels are chasing impersonation, and AI-assisted music has a labeled lane on the major services. Whether a track gets released, defended, and paid increasingly depends on the evidence behind it. The artists in the strongest position, AI-assisted or entirely human, disclose honestly when AI played a part, read their generator's terms before building a catalog on it, and keep dated records of how each release was made, from the first session file to the master. It is the old studio habit of labeling the tapes, upgraded with signatures, and it means the next time a filter guesses wrong, the answer is already in the folder. More of our coverage of AI in creative work lives on our [creative industries and entertainment page](https://www.agentpmt.com/industries/creative-industries-entertainment). * * * ## Sources - Michael Smith, man behind $8M AI song and bot streaming fraud, sentenced to 18 months in prison, Music Business Worldwide - 'AI false positive becoming more common in music', The Gleaner - Sony Music has now filed 260k takedowns of AI-deepfake tracks, Music Ally - Italy regulator probes AI music startup Suno over terms of service, Reuters via The Star - Watermarks and Fingerprints as Soft Bindings for Content Provenance: An Open-Licence Benchmark for Images, Audio and Video, arXiv - Learning to Watermark Speech Synthesis Against Model-Driven Reconstruction, arXiv - Watermarking: from Impossibility to Auditable Compliance, arXiv - Rethinking Visual Provenance: Detection and Watermarking Across Direct Visual Generation and LLM-Driven Code Rendering, arXiv - Will AI kill the Asian entertainment industry, or reinvent it?, Fortune