For three years the argument between generative AI companies and working illustrators has been stuck in the same groove. The companies scraped the open web and called it fair use. The artists said their life’s work had been ingested without permission, credit, or a penny in return. Nobody moved, and nobody seriously tried paying artists for the material.
A small text-to-video startup called Pippa is now testing a different opening move: what if you just paid them? The company has launched with a royalty scheme that sends money to illustrators every time someone generates an image or a video clip in their style. It is the clearest test yet of a question the whole industry has been avoiding, which is whether paying artists is actually enough to bring them across the line, or whether money was never the only thing standing in the way.
The answer matters well beyond one startup. If paying artists turns out to be a viable acquisition strategy for training data, it reshapes how every model developer sources material. If it turns out that paying artists buys compliance but not enthusiasm, the industry has a much harder problem than a pricing problem. Charles Pulliam-Moore’s reporting for The Verge on 2 August 2026 laid out the numbers, and the numbers are worth sitting with.
Table of contents
- Pippa’s Bet: Paying Artists Instead of Scraping Them
- The Royalty Math Behind Paying Artists Per Generation
- Why Paying Artists Is Not the Same as Earning Their Trust
- The Training Data Problem Paying Artists Does Not Yet Solve
- Seedance 2.5 and the Quality Question
- What the Lawsuits Mean for Paying Artists as a Business Model
- Can Paying Artists Survive Contact With Scale?
- What Businesses Should Take From the Paying Artists Experiment
- Frequently Asked Questions About Paying Artists and AI
- The Verdict on Paying Artists
Pippa's Bet: Paying Artists Instead of Scraping Them
Pippa was founded by Hogan Shrum, a marketing executive, and Sean Wright, who came from operations. The product itself is not exotic. Users type a prompt, and the platform returns short animated video clips suitable for social posts, personal projects, and light creative work. It launched publicly in May 2026 with iOS and Android apps, and opened a dedicated artist portal the following month.
What separates it from the field is the accounting. Rather than treating training data as a free input, Pippa treats it as a licensed supply chain with an ongoing revenue share attached. The pitch to illustrators is that paying artists is built into the unit economics rather than bolted on after a legal threat.
The artist portal as a recruitment channel
The portal, opened on 17 June 2026, is the mechanism through which the company signs contributors. Artists submit work, agree licensing terms, and are told they will be compensated for both their time and their talent. It is a deliberate inversion of the scraping model: instead of harvesting quietly and apologising later, the company asks first and puts the payment schedule in writing.
That framing is doing a lot of work. Paying artists through a formal portal creates a paper trail, and a paper trail is exactly what the companies currently facing litigation cannot produce.
A very small starting roster
The scale is modest and the company has not hidden it. As reported, Pippa had four signed artist licensing agreements at launch with four more in negotiation, and roughly 800 paying subscribers on tiers running from $14.99 to $99.99 per month.
Eight artists is not a dataset. It is a pilot. But it is a pilot with real contracts behind it, which is more than most of the industry can claim when it talks about ethical sourcing.
Why the founders chose video first
Video is the format where the objection is loudest and the supply is thinnest. Still-image models had years of open scraping before anyone organised resistance. Video arrived after the fight had already started, which means paying artists for motion work is both more necessary and more feasible than retrofitting consent onto an image corpus assembled in 2022.
The Royalty Math Behind Paying Artists Per Generation
The headline rates are specific enough to test. Pippa pays $0.005 per generated image and $0.003 per second of generated video. A fifteen-second clip therefore pays $0.045. Crucially, the company says 100% of that goes to the artist, with no platform cut taken off the top.
On top of the per-generation rate sits a 5% royalty pool funded from subscription revenue. That pool is divided among participating artists each month on a pro-rata basis, weighted by how heavily each contributor’s style was used. So paying artists happens twice: once per generation, and once as a share of the subscription business.
Running the numbers on a single artist
Work the arithmetic forward and the picture sharpens. At $0.045 per fifteen-second clip, an artist needs roughly 22,200 clips generated in their style to clear $1,000 in a month from per-generation royalties alone. At $0.005 per image, the same $1,000 requires 200,000 generations.
Those are large numbers for a platform with 800 subscribers. Paying artists at these rates only becomes meaningful income at a volume Pippa has not reached yet, which is the honest reading of the model rather than a criticism of it.
What the subscription pool actually contributes
The 5% pool is easier to size. Take 800 subscribers at a blended $30 per month as an illustration, and monthly revenue lands near $24,000. Five percent of that is roughly $1,200, split across a handful of artists.
Divided among four signed contributors, that is a few hundred dollars each per month before per-generation royalties. It is not a salary. It is a signal that paying artists is a line item rather than a marketing slogan, and at pilot scale a signal is most of what can be offered.
How the rates compare to other creative platforms
Context helps here. Spotify pays somewhere around $0.004 per stream. Adobe Stock returns 33 to 35 percent of each licence to the contributor. Pippa’s $0.003 per second sits in the same order of magnitude as a music stream, while its 100% pass-through on per-generation royalties is structurally more generous than a stock library’s revenue split.
The comparison also carries a warning. Streaming economics are precisely the arrangement musicians spent a decade complaining about. Anchoring a compensation model to per-play micropayments imports that argument wholesale, and paying artists fractions of a cent has a well-documented history of disappointing the recipients.
Why Paying Artists Is Not the Same as Earning Their Trust
The most revealing detail in the reporting has nothing to do with rates. Sean Wright noted that Pippa offers its partner artists the option to submit work under a pseudonym, because some are worried about being seen as “crossing the picket line.”
That single accommodation tells you the money is not the whole problem. If compensation resolved the objection, no one would need an alias. The pseudonym option exists because paying artists does not shield them from the social cost of participating, and the social cost is being priced by their peers rather than by the platform.
The picket line is real
Illustration is a referral-driven profession built on reputation inside small communities. A contributor who is publicly associated with an AI platform risks losing commissions from clients and standing among colleagues who view any collaboration as a defection. Paying artists a per-generation royalty does not compensate for a shrunken professional network.
Consent, credit, and control are separate asks
There are at least four distinct grievances in play: unpaid use, uncredited use, loss of stylistic distinctiveness, and eventual market substitution. Paying artists addresses the first directly and the second partially. It does very little about the third and nothing at all about the fourth.
An illustrator whose style becomes a preset that anyone can invoke for $0.045 has not lost income today. They have lost the scarcity that made the style commercially valuable in the first place. That is the fear no royalty rate answers.
Why pseudonyms cut both ways
Anonymity protects the contributor from backlash, but it also strips the model of the thing that would make it persuasive as an industry precedent. A public roster of well-known illustrators endorsing the arrangement would move opinion. A roster of pseudonyms proves the money is flowing while leaving the legitimacy question wide open.
The Training Data Problem Paying Artists Does Not Yet Solve
Here is the uncomfortable part. Pippa’s models are not built exclusively from licensed material. The company acknowledges using open models that carry initial training on the broader set of content available online, with a stated plan to transition toward artist-provided datasets over time.
That means the ethical foundation is aspirational rather than achieved. Paying artists sits on top of a base model whose own provenance is the same murky scrape everyone else relies on, and no royalty scheme retroactively cleans that.
Clean-room training is extraordinarily expensive
The reason nobody has built a fully licensed video model is arithmetic. Competitive systems need millions of samples. Licensing that volume at rates artists would consider fair pushes data acquisition costs into territory that makes the resulting product uncompetitive against rivals who paid nothing.
This is the structural trap. Paying artists properly at foundation-model scale may simply cost more than the market will bear, which is why every well-funded lab has chosen scraping and litigation risk instead.
The transition plan needs a timeline
A commitment to move toward artist-provided data is only as good as the schedule attached to it. Without a published milestone, the promise functions as reputational cover for continuing to use a scraped base indefinitely. Paying artists for the top layer while the foundation stays untouched is a genuine improvement, but it is not the clean break the framing implies.
What honest disclosure would look like
The most credible version of this model would publish the proportion of outputs attributable to licensed material, the size of the licensed corpus, and the date by which the scraped base is retired. Those three figures would let anyone judge whether paying artists is a transition strategy or a permanent veneer.
Seedance 2.5 and the Quality Question
Pippa plans to incorporate ByteDance’s Seedance 2.5 model into the service. The draw is practical: Seedance 2.5 generates up to 30 seconds of video and allows that footage to be refined without regenerating from scratch, which removes one of the more painful workflow limits in short-form AI video.
This is where the ethical proposition collides with product reality. Subscribers do not buy a licensing philosophy. They buy output quality, clip length, and iteration speed, and a startup that cannot match OpenAI’s Sora or Google’s Veo will not retain the subscriber base that funds the royalty pool.
Borrowed capability, borrowed provenance
Integrating a third-party foundation model solves the capability gap and reopens the provenance gap. Seedance is ByteDance’s model, trained under ByteDance’s data practices. Paying artists on the Pippa side does not extend any consent guarantee to what the underlying model learned before it arrived.
Iteration without regeneration matters more than length
The thirty-second ceiling gets the headline, but the ability to adjust a clip without starting over is the more consequential feature for anyone doing real work. Regeneration roulette, where each attempt discards everything good about the last one, is the main reason professionals abandon these tools after a trial.
Quality is what makes the royalty pool grow
Every argument about paying artists eventually routes back to subscriber count. The per-generation rate is fixed, so artist income scales only with usage volume, and usage volume scales with how good the product is. A better model is, counterintuitively, the strongest pro-artist feature on the roadmap.
What the Lawsuits Mean for Paying Artists as a Business Model
The legal backdrop is not incidental. Artists have brought coordinated copyright actions against Google, Meta, and Anthropic seeking damages over training data. Those cases will set the price of unlicensed material, and that price determines whether voluntary licensing looks cheap or absurd in hindsight.
Meanwhile OpenAI has argued to the Trump administration that restricting access to training data risks the United States losing the AI race to China. That argument reframes licensing as a strategic liability rather than an ethical baseline, and it is the single biggest obstacle to paying artists becoming an industry norm.
Litigation sets the reference price
If courts award substantial damages, every lab will recalculate. Licensing suddenly becomes cheap insurance, and paying artists moves from differentiator to table stakes. If the cases settle quietly for modest sums or fail outright, the incentive to license collapses and Pippa’s model becomes a curiosity rather than a template.
The geopolitical argument is hard to counter
The competitiveness framing is effective precisely because it is unfalsifiable in the short term. No one can prove that licensed training would or would not have slowed a national AI programme. Paying artists as policy has to survive that argument, and so far its advocates have not found a compelling rebuttal.
Small companies license, large ones litigate
There is a pattern worth naming. Startups with no scraped corpus and no legal war chest license, because licensing is cheaper than a lawsuit they cannot fund. Incumbents with enormous existing datasets litigate, because relicensing what they already hold is far more expensive than defending it. Paying artists is currently a strategy of the small.
Can Paying Artists Survive Contact With Scale?
Assume the model works at pilot scale. The interesting question is what happens at a hundred times the size, because that is where every revenue-share arrangement in creative industries has historically broken down.
The 5% pool is the fragile component. It is fixed as a percentage while artist count grows without limit, which means the per-artist share falls as the roster expands. Paying artists from a capped pool creates a dilution treadmill: recruiting more contributors improves the dataset and simultaneously reduces what each contributor earns.
The dilution problem in plain terms
Four artists sharing a $1,200 pool receive $300 each. Four hundred artists sharing a pool that has grown to $120,000 still receive $300 each only if subscriber revenue scaled perfectly in step with the roster. If artist recruitment outpaces subscriber growth, which it usually does in the land-grab phase, every existing contributor is worse off.
Style attribution is technically contested
Pro-rata distribution by style usage assumes you can measure how much any given artist contributed to a given output. In practice, diffusion models blend influences in ways that resist clean attribution. Paying artists proportionally requires an attribution method that is defensible, auditable, and not simply an opaque internal score.
What a durable version would need
A model built to last would probably fix a floor per contributor, publish its attribution methodology, cap roster growth against revenue, and offer term-limited licences that artists can withdraw. None of these are exotic. They are the mechanisms collecting societies developed over a century for exactly this problem.
What Businesses Should Take From the Paying Artists Experiment
For companies deploying generative tools rather than building them, this story is a procurement signal more than an ethics debate. Provenance is becoming a contractual question, and the vendors who can answer it will have an advantage that has nothing to do with output quality.
Any organisation building an AI strategy around generative media should be asking suppliers where their training data came from, what indemnity is offered against infringement claims, and whether the answer would survive disclosure. Paying artists is one vendor’s answer, and it is a more auditable answer than most.
Ask vendors the provenance question in writing
The practical step is unglamorous. Add a data-provenance clause to procurement templates, request the licensing position in writing, and treat a vague answer as a risk finding. Vendors who are paying artists will say so with documentation, and vendors who are not will produce fair-use boilerplate.
Indemnity is the real protection
Compensation upstream does not protect a downstream user from a claim. Contractual indemnity does. If a supplier is confident that paying artists has cleaned its corpus, it should be willing to indemnify customers against infringement actions, and that willingness is the cleanest test of whether the ethical claim is genuine.
Keep watching the model layer
Because Pippa plans to sit on top of Seedance 2.5, buyers should note that provenance claims do not automatically flow through a stack. Our AI models, tools and releases hub tracks how these underlying systems change, and the underlying system is where the actual data question lives.
Budget for the possibility that licensing wins
If litigation raises the cost of unlicensed data, tools built on scraped corpora may become more expensive or disappear. Organisations that standardised on a single such vendor will face a migration. Treating machine learning tooling as replaceable rather than foundational is sensible hedging while the legal position is unsettled.
Frequently Asked Questions About Paying Artists and AI
How much does Pippa actually pay per generation?
The published rates are $0.005 per generated image and $0.003 per second of generated video, so a fifteen-second clip pays $0.045. The company states that 100% of that amount goes to the artist with no platform commission deducted, which is unusual among creative marketplaces.
Is any other major AI platform paying artists per generation?
No large text-to-video or text-to-image platform currently operates a comparable per-generation royalty. Adobe pays contributors through stock licensing revenue shares, and several startups have announced licensing partnerships, but paying artists a royalty on each individual output remains rare.
Does paying artists make a model legally safe?
Not by itself. Legal exposure depends on the provenance of everything in the training pipeline, including any third-party foundation model underneath. Because Pippa acknowledges building on open models trained on broadly scraped content, paying artists for the licensed layer reduces risk without eliminating it.
Why would an artist refuse money for their work?
Because the objection is not purely financial. Artists cite loss of control over their style, absence of credit, professional backlash from peers, and the prospect of being replaced by a tool trained on their own output. The pseudonym option Pippa offers exists precisely because paying artists does not neutralise the social cost.
What is Seedance 2.5 and why does it matter here?
Seedance 2.5 is ByteDance’s video generation model, capable of producing up to thirty seconds of footage that can be refined without regenerating from scratch. Pippa plans to integrate it to close the quality gap with larger rivals, which is what determines whether subscription revenue funds the royalty pool.
Could this model scale to a large platform?
It is untested. The 5% subscription pool dilutes as the artist roster grows, and pro-rata distribution depends on style attribution methods that are technically contested. Paying artists at pilot scale with eight contributors is a very different exercise from doing it with eight thousand.
The Verdict on Paying Artists
Pippa deserves credit for making the commitment specific. Published rates, a 100% pass-through, and a named royalty pool are falsifiable claims, which is more than the industry’s usual gestures toward responsible sourcing. Anyone can check whether the money moves.
But the reporting also makes the limits plain. The base model is still built on scraped material. The roster is eight artists. The subscriber base is 800 people. And the pseudonym option quietly concedes that participation carries a reputational cost that no per-generation rate covers.
So is paying artists enough? On this evidence, it is necessary and insufficient. It resolves the compensation grievance and leaves consent, credit, stylistic control, and market substitution entirely intact. The startups that eventually win over the illustration community will probably need to offer all five, and the ones treating paying artists as a complete answer are solving the easiest quarter of the problem.
What makes the experiment worth watching is that it is falsifiable. In a year we will know whether 800 subscribers became 80,000, whether four artists became four hundred, and whether the scraped foundation was actually retired. Those three numbers will say more about the future of paying artists than any amount of position-taking. The full account of Pippa’s launch and its royalty structure is worth reading in Charles Pulliam-Moore’s report for The Verge.
