The Licensed Model as Corporate Redemption Arc
Walk into any design studio or artist’s workspace right now and you’ll catch the exact moment someone mentions Adobe Firefly 3 with a particular tone of voice—part relief, part resignation. Since its 2025 release, the tool has positioned itself as the ethically conscious alternative in a field that’s become synonymous with artistic theft. Adobe trained Firefly 3 exclusively on Adobe Stock’s licensed image library, a collection exceeding 300 million assets that the company explicitly markets as “commercially safe.” The pitch is seductive: unlike its competitors, this AI model was built on legitimate, compensated imagery. The artists whose work lives in Adobe Stock agreed to be part of this. The photographers signed contracts. The illustrators understood the terms. So we’re supposed to feel better about it, right?
This is where I find myself bristling at the framing. Yes, using licensed imagery rather than scraped data from the open internet represents a meaningful legal distinction. But let’s be precise about what that distinction actually means and what it conspicuously does not solve. The licensed training approach addresses one specific vulnerability: Adobe cannot be sued by artists for unauthorized use of their work during the model’s construction phase. That’s real protection for Adobe. Whether it constitutes actual justice for the creative class is an entirely different question, one that requires us to look beyond the corporate press release and into the actual ecosystem where artists are trying to survive.
The Compensation Question Nobody’s Really Answered
Here’s what we know with uncomfortable clarity: a 2025 MIT Media Lab study documented that 67 percent of professional illustrators reported losing commissions they directly attributed to clients switching to AI image generation tools. Sixty-seven percent. That’s not a margin of error or a demographic subset—that’s the majority of working professionals in a field that’s already been hollowed out by technological disruption and the devaluation of visual labor over the past two decades. The illustrators whose work trains Adobe Firefly 3 may have been compensated for licensing their images, but that compensation structure was designed around a fundamentally different use case: their work being displayed, featured, sold as individual assets. It was not designed around the scenario where their aesthetic, their stylistic decisions, their years of accumulated creative intelligence could be algorithmically extracted and reproduced at scale.
This matters because the licensing fee structure does not dynamically respond to the actual value being extracted. An illustrator licensed an image to Adobe Stock for perhaps $50 or $100. That image, along with hundreds of millions of others, then becomes part of the training corpus for a generative tool that competes with the illustrator’s livelihood. The original $50 transaction cannot possibly account for that secondary harm because it happened in a different era, when this use case didn’t exist. The licensed model doesn’t solve this temporal problem. It just allows Adobe to claim ethical high ground while the mathematics of harm remain unchanged.
The Copyright Office’s Cautious Language and What It Actually Means
In February 2025, the U.S. Copyright Office issued updated guidance on AI-generated imagery that seemed designed to satisfy nobody and clarify nothing. The ruling established that AI-generated images remain uncopyrightable unless a human author exercises “sufficient creative control” in the output. That word “sufficient” is doing a lot of work. Sufficient according to whom? How do we measure it? What constitutes the threshold between using a prompt as a tool and using a tool as a proxy for actual creative decision-making?
This guidance arrived amid ongoing litigation that still hasn’t been fully resolved. The class-action lawsuit Andersen v. Stability AI, which named Stable Diffusion, Midjourney, and DeviantArt as defendants, reached a partial settlement phase in 2025, but the artist compensation discussions remain active and contentious. We’re watching this play out in real time: the legal framework, the corporate strategies, the artist advocacy responses all happening simultaneously, each one influencing the others. What the Copyright Office decided is that the burden of proving sufficient creative control falls on whoever wants to claim copyright. What that means in practice is still being determined by lawyers, judges, and increasingly exhausted visual artists.
Scale as Its Own Kind of Problem
Midjourney reported over 20 million registered users as of the third quarter of 2025. These users were generating an estimated 3 million images per day across the platform. Three million images. Per day. That’s a scale that makes it almost impossible to discuss individual artist harm in the aggregate, which might be precisely why the discourse has shifted toward theoretical discussions about consent and licensing rather than actual economic impact on actual people trying to pay rent.
When you’re operating at that scale—when the tool is this accessible, this fast, this cheap—the problem stops being about whether the training data was licensed. The problem becomes about market saturation. It becomes about the client who used to hire an illustrator for $2,000 now using Midjourney for $20 a month. It becomes about the aesthetic homogenization that happens when millions of people are prompt-engineering the same model simultaneously. Scaling a tool doesn’t just displace workers; it actively diminishes the visual culture we’re all swimming in.
Where the Conversation Should Actually Go From Here
Here’s what I’m genuinely uncertain about, and I’m willing to sit with that uncertainty: whether there’s an ethical way to scale AI image generation tools while maintaining space for human artists to sustain themselves. I’m not convinced there is, at least not under the current economic conditions. But I’m also not prepared to argue that the solution is to ban the technology or pretend it isn’t happening. The question that matters is whether we’re going to demand that companies designing these tools actually think through the downstream effects, or whether we’re going to accept licensing agreements as sufficient moral cover.
Adobe’s licensed training data is better than scraping without permission. It’s just not good enough to stop here. Better is not the same as good, and it’s certainly not the same as just. What we need is pressure on platform designers to think beyond their legal liability and toward actual support structures for displaced creative workers. We need the U.S. Copyright Office AI Guidance 2025 to evolve as these tools do. We need the ongoing artist compensation discussions in litigation to set meaningful precedents rather than just settling quietly. And we need to resist the comfortable narrative that using licensed data resolves the ethical questions. It doesn’t. It just repositions them.
What are you actually observing in your own creative spaces? Are you a working artist watching this unfold? Are you a client rethinking how you commission work? This conversation lives in the friction between what’s technically possible and what we’re collectively willing to accept. I’m interested in where you think that friction actually is.