On July 16, 2026, Lorde used Instagram to tell a technology platform to keep artificial intelligence away from her art. Spotify, the world's largest music streaming service with more than 600 million monthly active users, had deployed an “About the Song” feature — a beta AI tool that draws on third-party sources to generate contextual notes — that produced a fabricated account about one of her tracks, wrongly linking a stage performance to an entirely different song. “I’m going to go out on a limb and say we don’t want this,” she wrote. “Reducing a song to an AI-generated meaning right at the source feels like it limits free interpretation.”

Spotify removed the text. A spokesperson described the product as a beta feature that would improve through feedback. Lorde’s objection, however, was not only about an inaccurate blurb. It pointed to a broader concern: the growing presence of machine-generated interpretation in music, and the platform assumption that AI curation can be placed by default alongside artistic work.

Artists without global visibility face a quieter version of the same problem. Owen Lyman-Schmidt, one half of the Philadelphia gutter-folk duo Makeshift Hammer, found in 2024 that an entity named “Carey Dupont” had taken his band’s recordings, made minor alterations, and uploaded them to Spotify with AI-generated album art and nonsensical track names. The fake artist had reached 650,000 streams. Makeshift Hammer had 40,000.

The two incidents reflect different sides of the same issue: Spotify’s increasingly complicated relationship with artificial intelligence, and the wider music industry’s effort to respond to technology that challenges the basic idea of creative authorship.

The Fraud Machine

The issue is not limited to isolated bad actors. Analysts estimate that AI music generation has become part of a fraud ecosystem worth $4 billion.

Suno, a major AI music generator, reached $300 million in annual recurring revenue by February 2026 and was valued at $2.45 billion. Its 2 million paid subscribers generate an estimated 7 million songs each day, an amount equal to Spotify’s entire historical catalog every two weeks. Most of that output will never be heard by a person. Some of it, however, is used for fraud.

Fraud tactics have changed. Rather than using bot farms to replay a small set of songs millions of times, which is more likely to trigger detection, fraudsters now upload vast numbers of AI-generated tracks and stream each one only a few thousand times. The approach allows them to collect royalties while drawing less scrutiny. Apple Music confirmed that it demonetized 2 billion fraudulent streams in 2025, redirecting nearly $17 million away from illegitimate activity and toward legitimate artists. Deezer reported that as many as 85% of streams on AI-generated music were fraudulent, and by 2026 more than 30,000 fully AI-generated tracks were being uploaded to its platform every day.

The most prominent case is Michael Smith, a North Carolina musician who pleaded guilty in March 2026 in the first federal criminal streaming fraud case in U.S. history. Smith used AI to create hundreds of thousands of songs and deployed bot networks across thousands of accounts. At the peak, the operation generated more than 660,000 streams per day and fraudulently collected more than $10 million in royalties. Smith agreed to forfeit more than $8 million and faces sentencing on July 29, 2026, a proceeding expected to indicate how U.S. authorities will treat AI-enabled music fraud.

Smith’s case was unusual in scale, but not in method. Makeshift Hammer’s case shows how a similar scheme can affect independent musicians. A bot farm can scrape recordings from services such as SoundCloud, alter them just enough to avoid detection, repackage them with AI-generated artwork and near-match titles, and then generate automated plays. “Carey Dupont” had no website, no social media presence and no concert history, only a Spotify profile and 650,000 streams. When Lyman-Schmidt filed takedown notices with Spotify and SoundCloud, he received automated responses and, from Spotify, silence. Three and a half months after the initial complaint, and only after his story was published, Spotify told him the tracks would be removed “pending an investigation.”

The harm is also economic. Spotify pays artists through a “streamshare” model, in which roughly 70% of the platform's music revenue is distributed as royalties and divided proportionally among all rights holders based on their share of total streams. If a scammer adds 650,000 fake streams to a pool that previously contained 40,000 legitimate ones, the same overall payout is effectively divided 94% to the fraudulent account and 6% to the original artist. The platform does not lose money. The fraudster receives royalties. The musician absorbs the cost.

The Sheriff and the Outlaw

Spotify has not ignored the problem. In September 2025, the company revised its AI policy around transparency, a strict ban on unauthorized voice cloning, and stronger spam filtering. During the preceding 12 months, Spotify removed more than 75 million tracks for spam and low-quality content.

On April 30, 2026, Spotify introduced the “Verified by Spotify” badge, a green checkmark intended to help listeners identify authentic human artists rather than AI personas. The badge verifies artist identity, not the nature of the music. A verified human artist can still upload fully AI-generated tracks, while AI-persona artists are explicitly ineligible. Spotify is therefore drawing a boundary around who an artist is, while declining to define what the artist makes.

That distinction matters. Spotify does not want to arbitrate artistic merit. It wants engagement. AI, despite its problems, can generate engagement.

The tension became clearer on July 14, 2026, two days before Lorde’s complaint, when Spotify launched a ChatGPT-like music assistant for Premium users. On the Q1 2026 earnings call, co-CEO Gustav Söderström said Spotify has “the capabilities and technologies we need” to build a derivatives business on top of existing artists’ music. He did not disclose what those capabilities were trained on. Creators have no disclosure system, no audit trail and no compensation mechanism to determine whether their work is used in Spotify’s generative systems.

Spotify is developing AI tools that interpret, recommend and potentially regenerate music, while also using AI tools to detect AI-generated fraud. In that sense, it occupies both sides of a dispute its own platform model helped make possible.

An Industry Caught Flat-Footed

The music industry’s response has combined legislative lobbying, improvised legal strategies and reluctant commercial deals. Throughout that response, the imbalance between major industry players and working musicians has become more visible.

At GRAMMYs On The Hill 2026 in April, the Recording Academy backed three bipartisan bills: the NO FAKES Act, which would create a federal right protecting voice and likeness from AI deepfakes; the TRAIN Act, which would give creators visibility into AI training datasets; and the CLEAR Act, which would require transparency in AI systems. The urgency reflects a specific legal gap. Copyright law protects recordings, but not voice itself, a weakness that AI voice cloning has made more consequential. None of the three bills had reached a floor vote by mid-2026.

In January 2026, folk musician Murphy Campbell discovered AI-cloned versions of her voice distributed on her own Spotify profile. The perpetrator then used a distributor’s Content ID access to file copyright claims against Campbell’s original YouTube videos, causing her to lose monetization on her own material. She described the situation as being “in a weird limbo where I’m telling robots to take down music robots made.” Sony Music has since requested the removal of more than 135,000 songs created to impersonate artists on its roster, a number the company says is likely only a fraction of the actual total. In April 2026, Taylor Swift filed three trademark applications with the USPTO, including two sound marks covering her voice, to address what attorney Josh Gerben called “a gap that trademarks may help fill” now that AI can imitate an artist without copying an existing recording.

Major labels have pursued a different route. Universal Music Group and Warner Music Group reached licensing agreements with AI music company Udio covering label-controlled rights, with plans for an artist opt-in platform. The approach suggests that if AI cannot be stopped, it can be licensed. Independent music groups have been more resistant. The Independent Music Publishers Forum has urged publishers to reject AI licensing agreements that do not allocate at least 50% to songwriters, warning that the financial benefits of AI deals may not reach the creators whose works helped train the systems.

Even with lobbying and commercial negotiations underway, the most important decisions may come from courts. The United States has no comprehensive federal AI law, leaving judges to address questions that legislatures have not yet answered. The AI music economy remains provisional as it awaits two significant rulings scheduled within the same week. On July 29, 2026, Michael Smith faces sentencing in the first federal criminal streaming fraud case. Two days later, on July 31, the Munich Regional Court is expected to deliver a verdict in GEMA v. Suno, based on Germany’s text-and-data-mining exception. Separately, in Sony Music Entertainment v. Suno, now before U.S. District Judge F. Dennis Saylor IV in Massachusetts, a fair-use summary judgment hearing was held in July 2026. It is the first federal case to address whether training a generative AI model on copyrighted recordings qualifies as fair use. No ruling has been issued. Together, the cases will help determine whether AI training data becomes a paid licensing category or remains largely free to use. If training is found to be fair use, AI music generation could expand under the current economics. If not, the business model could change rapidly.

These legal fights are significant, but they can obscure the practical imbalance experienced by musicians. Lorde posted an Instagram Story and received a response from Spotify within hours. Owen Lyman-Schmidt spent three and a half months filing takedown notices, consulting lawyers and publishing an investigative article before the platform acted on an apparent case of theft. Major labels negotiate licensing deals with AI companies. Independent musicians are offered Discovery Mode, a Spotify program that lets them accept a lower royalty rate in exchange for playlist exposure.

Spotify’s architecture reinforces that hierarchy. Its streamshare model rewards volume over value, catalog scale over curation, and automation over accountability. AI fits easily into such a system. It can produce vast amounts of content at near-zero marginal cost, supply algorithms with constant novelty and shift the financial impact of fraud onto artists the platform says it supports.

What distinguishes this shift from earlier disruptions such as the synthesizer, the sampler or the MP3 is its speed and opacity. AI does not only replicate or distribute music. It can interpret it, impersonate it and potentially replace it, while legal and ethical frameworks remain behind the technology.

Lorde asked for an opt-out. Makeshift Hammer asked for justice. So far, the platform has provided neither. It has continued to offer more features, more AI and limited answers.

Until that changes, the ghost in the stream is not the AI. It is the artist watching their work disappear into a machine they never asked to build.