The Consent Economy

In my last post, What We See When We Look, I shared that statistically most people can't reliably tell AI-made work from human-made work, and that the label changes opinion more than the work itself does. This one looks at the people who are applying the label, and who profits (and who doesn’t) once they do.

Sometime in early March, a medieval historian named Verena Krebs opened Grammarly's newest feature and found a colleague waiting for her. The tool, called Expert Review, let her pick from a list of "experts" who would critique her writing in their own voice. One of the names on the list was David Abulafia, the Cambridge historian of the Mediterranean world. Abulafia had died two months earlier. Nobody from Grammarly asked his permission while he was alive, nor contacted his estate after. 

The team at Grammarly had built the feature to critique essays and articles "from the perspective" of real people, using their published work as raw material. The list read like a syllabus: Stephen King, Neil deGrasse Tyson, Kara Swisher, Carl Sagan. The problem was that none of them had agreed to it. Swisher's response, on finding her own advice being generated on her behalf, was to threaten to "go full McConaughey" on the company. A disclaimer buried deep in Grammarly's support page said the "experts" had no actual affiliation with, or endorsement of, the feature bearing their names.

By March, journalist Julia Angwin had filed a class action lawsuit against the parent company Superhuman in federal court, alleging violations of New York's and California's right-of-publicity laws. Grammarly disabled the Expert Review feature within days. The company still runs a "humanizer" tool that rewrites AI text to sound human-written, and an AI grader that guesses what a professor wants to hear. It’s awkward from a business perspective, squeezed on one side by lawsuits and on the other by frontier labs giving away equivalent editing inside their own chatbots for free. And though this is small case, as these things go, it brings attention to the personal voices being monetized, without compensation, that nobody agreed to.

The majors making moves

Suno and Udio were sued in mid-2024 by all three major record labels for training on catalogues nobody licensed to them. Udio settled with Universal Music Group first, in October 2025, and the terms shifted the company virtually overnight: downloads were switched off within a day, with just a 48-hour window the following month for users to download what they'd made prior to the new terms, and the platform was rebuilt as what UMG called a "walled garden", music that streams inside Udio and nowhere else, with artists who opt in being paid for training and for use. Warner settled with Suno a month later with friendlier terms that let users keep downloading and distributing. Universal and Sony never settled with Suno, and the reason became clear. In May, they moved to add 61,026 recordings to the case after audio fingerprinting turned up matches in Suno's training data, pushing the potential damages, calculated under a fixed per-work copyright penalty, from roughly $84 million to over $9 billion. Suno called the move a delay tactic. A fair-use hearing, on the question of whether training a model on copyrighted recordings without a licence counts as a permitted use, is scheduled for this month. As of this writing, nobody knows how it will play out, but a recent code leak that revealed the targets of Suno’s vast data scraping efforts probably won’t help their case.

Midjourney has been going through a similar ordeal. Disney and Universal sued the company in June 2025, then Warner Brothers followed that September. Midjourney argued fair use, telling the court that "copyright law does not confer absolute control over the use of copyrighted works." The company has never taken outside investment, brought in roughly $300 million in revenue in 2024 according to PitchBook, and is reportedly closing in on $500 million now.

Big names get their cut

The celebrity embrace of AI is tied to a well-established business model. Martin Scorsese's deal with Black Forest Labs (BFL), the company behind the FLUX image model, was brokered through BroadLight Capital, an investment firm co-founded by Scorsese's own manager, Rick Yorn, who is also a BFL investor; CAA co-founder Michael Ovitz, another BFL investor, helped close the deal and appears alongside Scorsese in the company's own promotional video. Yorn now profits from every director who follows his client's lead. Black Forest Labs was valued at $3.25 billion in December, and FLUX has been downloaded more than 400 million times and is used in Adobe, Canva, Meta, and Microsoft products.

Last November ElevenLabs, valued at $6.6 billion, launched an "Iconic Voice Marketplace" that lets companies license the voices of Michael Caine, Maya Angelou, and Alan Turing, among two dozen others living and dead. Matthew McConaughey, an ElevenLabs investor since the company's founding, isn't on the marketplace himself, but he narrated the launch. What the announcement didn't dwell on was that ElevenLabs had just settled a lawsuit filed by two actors who'd accused the company of cloning their voices without permission. ElevenLabs wisely turned the thing it got sued for into the thing it sells.

The detection game

AI detection has now become its own industry, and the best of it turns out to be genuinely good. When the Commonwealth Short Story Prize's regional winners were flagged for likely AI use in May, The Atlantic asked Pangram, the detector that flagged them, to run the prize's entire fifteen-year archive back through its model. Almost nothing else came back positive. Only four stories in fifteen years were flagged, three from this year and one from last, including the piece published in Granta that kicked off the whole controversy, which scored 100 percent. Pangram's co-founder says the tool has a false-positive rate of just one in ten thousand.

Deezer took a more proactive route and built detection algorithms right into their platform, with 13.4 million tracks tagged since June 2025, and 85% of streams for AI tracks treated as fraudulent and stripped of royalties. Their detection technology is now being licensed out to other platforms. The approach is perhaps warranted. Back in March, Michael Smith pleaded guilty to defrauding streaming platforms of more than $8 million, using hundreds of thousands of AI-generated songs and a bot network built to mimic real listeners. It was the first federal conviction of its kind and he’s due to be sentenced sometime this month.

Disclosure vs detection

Provenance is having a moment too, as a business tool. At Google's developer conference in May, OpenAI, Kakao, and ElevenLabs agreed to adopt SynthID, Google's invisible watermark for AI-generated images, video, and audio. More than 100 billion files already use it. Midjourney and Black Forest Labs have not adopted it and don’t plan to, though Midjourney’s founder and CEO David Holz says all images generated by their tools disclose AI in the files' metadata. And Meta, which has AI image tools embedded in Facebook, Instagram, and Whatsapp, has its own watermarking system. Their watermarks are substantially less effective once an image is cropped however, with detection falling as low as 55% in a recent test. 

The music industry just proposed its own version of the labelling idea. The RIAA and IFPI, backed by A2IM, the Recording Academy, SAG-AFTRA, and the Human Artistry Campaign, proposed a two-tag labelling system for streaming: an uppercase "AI" for tracks that are wholly AI-generated or where AI performed the lead vocal or key instrumental parts, and a lowercase "ai" for tracks that are substantially human-made but lean on AI for some elements. Streaming platforms haven't committed to either tag yet. DiMA, the industry lobbying group, gave what its own trade press called a cautious welcome, saying it was waiting on better metadata from the music labels before committing to anything. Spotify and Apple Music already run their own systems regardless, with transparency credits being provided since April, and transparency tags being applied since March. None of these systems say what role AI had in the lyrics, the composition, or the cover art, only the recording itself.

One-time payment

Only one AI lab has ever paid the artists whose work was used to train its models, and it didn't do it voluntarily. Anthropic, the company behind Claude, which provides research and editing assistance for this blog, agreed to pay $1.5 billion into a settlement fund after a court found that while training language models on legally acquired books was fair use, downloading millions of them from pirate sites like Library Genesis was not. That distinction is the whole case: Anthropic won on the big question and still had to write the biggest cheque in copyright history, because winning on principle and having pirated the files are two different facts.

But even this payment hasn't fully happened yet. At a May hearing, the judge declined to grant final approval on the spot, despite a 92.77 percent claims rate and an implied payout of roughly $3,100 per work, final sign-off was still pending as of this writing. 

Every other lab is still arguing over whether it owes anyone anything. The exceptions are the ones that never got sued because they never needed to be. Adobe Firefly and Getty built their models on licensed stock libraries from the start. Bria pays contributors directly for training data. Exactly.ai lets individual artists train models on only their own work and keep both the model and what it makes.

A proposal to the backlash

In late June, Mark Cuban, an AI investor with no reputation as a skeptic, posted a long message on X about why communities are fighting AI data centres. His real point wasn't about power grids. It was that the backlash Silicon Valley keeps treating as a NIMBY problem actually reflects current anger at what AI is doing to ordinary livelihoods, and no amount of political spending will make that anger go away.

Cuban referred to artists specifically, and urged the big labs to engage directly with the arts and creative unions, not with the music or film companies. Don't bother paying famous people to endorse what you're building, he added, calling the idea dumb. Talk to artists instead, and ask them what they actually need. "Every creative I know is TERRIFIED about what AI will do to their profession," he wrote. When a user pushed back by resurfacing an old, more optimistic post of his, Cuban didn't back down from that position either. He still believes AI nets out as a win for creators over time, he said, and that people, not machines, are the ones who can actually tell a good story. But only if those people still have jobs.

The price of permission

Some companies are trying to do the right thing. Tess.Design launched in 2024 and built a company entirely around paying the artists whose style trained its models. They offered a 50 percent royalty and advances of $300 to $4,000 for the first artists who signed on. Some artists took the deal. Others turned it down over concerns of brand dilution, ideological objection, or the simple belief that art made by hand is worth more, regardless of what it pays. Thirty-seven artists eventually joined, but it wasn’t nearly enough. The company shut down this January, after less than two years. They couldn't make it work as a business.

Canadian indie music distributor LANDR is investing $1M and has so far managed to sign up over 30 thousand artists for its Fair Trade AI program, which promises to pay artists 25% of net licensing revenues for opting in and submitting tracks to train AI music models. Payouts are promised monthly, and each submitted track gets a $5 advance towards future royalties. It's yet to be seen how or if this model will prove lucrative for the artists, but it's an honest effort to make it work. 

And that’s where things still stand right now. Offered to be paid honestly, some artists are saying yes. But the business models, and the companies trying to make it work, have yet to find a groove. Clearly there’s a lot of money trading hands at this point, and that suggests there's an opportunity to strike a balance and make it work. It's just that where the money is actually flowing, and where it needs to get to, still hasn’t been completely figured out.

Research and editing assistance provided by Claude Fable 5. Cover image generated with Midjourney 8.2 Preview.