The Infinite Catalogue: How AI Rewrote Stock Photography — and Why I Stopped Paying For It

It's a weekday morning in autumn 2025. Santa Maria Capua Vetere has that undecided sort of weather where the cold air cuts your face in the shade, but the sun warms you the second you turn the corner. I'm under the pale stone portico a stone's throw from the Criminal Court, the white 1960s building with those distinctive horizontal bands of windows staring down at you. Past the courthouse, I head down Via Albana and onto Via Melorio. A few steps further and I reach the entrance of the block where EmmeU Creative Studio is based. The smart modern building takes me in and carries me up in the lift, in the metallic hush of the cabin rising. Inside, the smell of coffee is already drifting from the little moka pot left on the electric ring in the next room — a ritual that repeats itself identically every morning. I sit down, take off my black rectangular glasses, run a hand through my grey beard and unlock my phone.
The first thing I see is the Freepik newsletter. It's announcing yet another update to the generative catalogue, partly folded by now into Magnific, the upscaling tool Freepik picked up in 2024. Another millionth milestone of images generated and made available. Up until a couple of years ago, reading an email like that would've had me thinking about some minor photo agency getting bought out. Now I know exactly what it means. It means the machine has learnt to clone commercial aesthetics. A year ago I had two subscriptions running side by side: one with Envato, one with Freepik. Recently I decided not to renew Envato. Freepik's still on my card, but I wouldn't rule out dropping that one too before the year's out.
I've already written about this short circuit, this fake sense of security the tech giants are selling us, in Il paradosso della fortezza. There, I talked about fear-based marketing, the way they're pushing us towards AI by waving the copyright bogeyman about. But this morning I want to pop the bonnet. I want to walk you through the technical mechanics of how the stock platforms have transformed from the inside, quietly, while I carried on paying monthly subscriptions to download images I thought had been shot by actual human beings.
There was a time when stock photography had a certain grubby dignity to it. You knew what you were buying: the time you didn't have to organise a shoot yourself. You'd search for a photo of a business meeting and get the usual five grinning models with blindingly white teeth, but you knew there'd been a set, a photographer, catering, lights rigged up in some office in Seattle or Milan. You were paying for that prior graft. It was a clear deal. Me, the agency, paying you, the platform, to use the sweat of a photographer who'd already done the donkey work. But today? Today I'm paying to access a server that no longer holds photographs at all, just statistical projections of photographs past.
The shift wasn't an explosion — it was an infiltration. The platforms didn't wipe their archives to replace them with AI. They simply sat the two side by side, blended them, folded them in together. They stopped being archives and became generators. And I, like thousands of other creatives, carried on paying commercial licences for images that don't exist, spat out by invisible prompts. The question I asked myself over that coffee in Santa Maria Capua Vetere, looking out towards the courthouse, was a simple one: if the value's no longer in the shot but in the generation, why am I still paying toll to the old gatekeepers? Where do the images I'm sifting through every day in my searches actually come from? And, more to the point, who's on the hook if something goes wrong?
The catalogue that duplicated itself
The technical mechanics of this mutation are fascinating and brutal in equal measure. Up until 2022, a professional commercial photographer would organise a studio shoot, produce maybe 100 perfect shots of a subject — say, a craftsman at work in his workshop — and upload them to whichever stock platform. From those 100 shots, the photographer hoped to earn a steady income through micro-licences. It was a volume market.
Today, that same photographer sees his material treated in a completely different way. Those 100 shots aren't just the finished product for sale any more; they've become feed for the machine. They've become training data. The platforms worked out they didn't need to keep paying new contributors for fresh images of craftsmen at work. They could use the images already in their possession as seed — as a seed or base reference — to train generative variants endlessly.
The technical difference between a seed image used for training and a generated variant is fundamental to understanding today's market. The seed image is the real photo: it holds the sweat, the grain of the film or sensor, the optical imperfections of the lens, the real light that hit a real face. The generated variant (the new output) is a mathematical calculation. It's an image that keeps the stylistic and chromatic consistency of the seed, but depicts a subject that doesn't exist. It's an echo. And it's an echo that costs nothing to produce, yet the platform sells it back to you at the same price as the original.
Nearly every major platform has built this internal fork. I've watched separate sections spring up and then gradually merge: the Shutterstock AI Image Generator, Adobe Stock's formidable integration with Generative AI Firefly, Getty and iStock leaning on the NVIDIA Picasso engine — rebranded NVIDIA Edify as of September 2024, now part of NVIDIA AI Foundry — launching Generative AI by iStock. And then Freepik AI. Worth a technical footnote I couldn't have made back in 2023: as of late 2024, even Envato, which plenty still consider an independent player, is owned by Shutterstock. The stock market is consolidating at exactly the same moment it's automating — two processes feeding off each other.
But here's where a technical point comes in that separates the amateurs from the professionals: not every generative AI hidden inside the stock platforms is the same. There's a hard line between engines classed as commercially safe — trained exclusively on content already licensed, owned by the platform, or in the public domain — and open engines, trained wild on billions of images through indiscriminate web scraping.
It's a distinction any joiner from Caserta would recognise instantly: you don't use the same workbench for prized oak and scrap poplar. Mix the two and the sawdust from the cheap timber ends up ruining the finish on the expensive piece. The stock platforms had to build closed pens to reassure corporate clients that the sawdust of the open web wouldn't end up in their renders. And they had me believing that fence justified the ticket price. But the truth is the catalogue duplicated itself at zero marginal cost to them, while my subscription stayed exactly where it was — if not went up.
Two training philosophies, two levels of risk
Before slotting even a single generated image into a client campaign, I stopped myself. I wanted to know exactly what I was buying. In our line of work you can't afford to be sloppy about where your assets come from. Here's what I checked before trusting any of it, digging into the two philosophies driving the models on the market today.
The closed model: Adobe Firefly
Adobe made a surgical and commercially shrewd choice with Firefly. From day one, they officially stated that their generative model is trained exclusively on content held in Adobe Stock, openly licensed files, and public domain material whose copyright has expired. It's a sealed ecosystem. No web scraping, no living artists' work lifted from ArtStation or DeviantArt.
That choice led them to set up a compensation scheme, the Firefly bonus, for Adobe Stock contributors whose material ended up in the training set. This bit matters, because it's the one structured case where photographers get paid not just when a user downloads their photo, but simply for having helped train the algorithm. It's a fraction of a penny, sure, but it sets a solid legal principle.
For higher-end clients, Adobe's gone further still: they offer an IP indemnity for Enterprise clients. In plain terms: if someone sues you because a Firefly-generated image looks a bit too much like their protected work, Adobe covers the legal costs and damages. It's the perfect shield.
But there's a sharp edge I'm not going to soften. Firefly is commercially safe by contract and by law, not because the results are always technically flawless. Precisely because its training pool is limited and clean, the engine stumbles exactly where the open competitors soar. All the same chronic technical limits of every closed generative engine remain: hands often veer into body-horror territory, working readable text into images is pot luck, and reproducing logos or complex geometry takes dozens of goes. You're buying legal safety, but paying for it in technical frustration.
The open model: Stable Diffusion and the London ruling
On the other side of the fence is the wild west of open models, Stable Diffusion and Midjourney leading the charge, trained by scraping every available pixel on the internet. But that approach hit a significant legal wall. The case of Getty Images v Stability AI, heard before the High Court in London and reaching a landmark judgment on 4 November 2025 ([2025] EWHC 2863 Ch), set the rules of the game.
Let me walk you through the ruling plainly but precisely, because it overturned quite a few people's assumptions. Getty had accused Stability AI of stealing millions of images from its archive to train Stable Diffusion. The British judge established a watershed technical-legal principle: Getty lost the main claim for direct copyright infringement on the final images. Why? The judge technically recognised that the AI model, once trained, no longer contains the protected images. There's no hidden folder of JPEGs sitting inside the neural network; there are only mathematical weights, statistical relationships between pixels. Generating an image doesn't mean "copy and paste".
That said, Getty did win partially on the trademark question. Because the training datasets contained so many Getty photos with the classic watermark stamped across them, the AI had learnt that "a high-quality professional image" had to include some sort of blurry smudge in a corner, clumsily reproducing the Getty mark. That was found to be infringing. It's a rare, one-off sort of case, but it shows unmistakably that there's residual risk in using open models.
Why does this ruling matter both for a small studio in Caserta and for an agency in Shoreditch? Because it's the first real UK precedent on the subject. If you work with English clients, UK case law is what counts. And the ruling says one thing clearly: responsibility for unlawful training sits with the developer of the model, not with the creator typing the prompt. I'm a user of a tool, not an accomplice to theft. But if that output throws up someone else's trademark, the problem suddenly becomes mine too.
When generative AI wins a brand the crowd — and when it sinks it
In marketing, no theory holds up until it meets the public. The stock platforms promise us instant perfection, but generative AI applied at scale is opening up gulfs between brands that know how to handle it and those getting swallowed by it. To talk about it properly, we need facts, dates and names you can check. No vague impressions.
On the downside: Coca-Cola and the cost of two years of generated "Holidays Are Coming"
For two years running, Coca-Cola paired its legendary Christmas ad of illuminated trucks with versions generated entirely by AI — not as a full replacement, but as a central part of the campaign's public narrative. In 2024 they had a go with "Real Magic AI", the platform built in collaboration with OpenAI. In 2025, within the wider "Refresh Your Holidays" campaign, they ran a traditional ad shot with real actors, "Feliz Navidad", produced by WPP/OpenX, alongside the AI spots. But it's the generative side that made the noise: handed to two specialist studios, Secret Level and Silverside AI, who produced over 70,000 clips in around thirty days — a core team of 5 AI specialists handling prompting and selection, within a project that still involved around a hundred people across Coca-Cola, WPP and the studios. The saving wasn't so much on total headcount as on time: a month instead of a year.
The first year was a technical shipwreck, close to textbook stuff. In the 2024 ad the truck wheels slid across the snowy tarmac without turning, the backgrounds seemed to "melt", the faces and hands of the extras were twisted and grotesque, Father Christmas only flashed by briefly, often with one badly proportioned hand. In 2025 Coca-Cola fixed some of these errors with a precise creative choice, not just a technical one: they shifted the weight of the storytelling onto anthropomorphic animals — polar bears, hedgehogs, seals, pandas — specifically to sidestep the uncanny valley of human faces, and rebuilt Father Christmas in the classic style of Haddon Sundblom's 1930s paintings rather than generating him in motion from scratch. The truck wheels, in 2025, turn in a decidedly more convincing way. But continuity slips remain — animals that shift slightly in appearance from one shot to the next, textures wobbling between glossy and plasticky, some shots very polished, others more "cartoonish" or under-rendered.
Public reaction, in both years, was harsh. Reading the comments online, the recurring words are damning: soulless, lifeless, creepy, digital slop. A joke went round the world on X, later picked up by trade press: "This is the best advert Pepsi has ever made" — used precisely as shorthand for the reputational damage.
There's a huge technical lesson in all this, and it's not just about marketing. In 2024 AI automated production while wiping out artistic direction on the details that actually matter — hands, wheels, continuity. In 2025 Coca-Cola showed they'd learnt something: they got round the known technical limits of 2024 with a directorial choice, animals instead of human faces, not just with more computing power. And yet, even after fixing the most glaring errors, they couldn't recover the emotional warmth of the original 1995 spot — which suggests the underlying problem wasn't only technical, but one of creative direction applied to a tool meant to speed things up, not replace whoever decides what story to tell. Coca-Cola ignored a precedent from just a few months earlier: the Toys "R" Us ad made with OpenAI's Sora, unveiled at the Cannes Lions Festival on 24 June 2024, which had already met the same backlash — "creepy", "dehumanising", compared to Stephen King nightmares by more than one commentator — the same sense of cheap plastic being sold as childhood magic.
On the upside: the Adobe Firefly model, where transparency becomes architecture
Here I need to go back to looking at system architecture rather than individual campaigns. Adobe states publicly that Firefly is trained exclusively on licensed content — chiefly the hundreds of millions of Adobe Stock assets — plus openly licensed and public domain material whose copyright has expired. No dataset scraped from the web without permission, according to their official line.
On top of that there's a compensation scheme for Adobe Stock contributors whose material helped train the model, and a contractual IP indemnity for Enterprise clients who've signed up to the relevant entitlement: cover that extends to copyright, trademark, patent and image rights claims, with a few contractual exclusions (substantial client modifications, unlawful uses, material the client supplied for custom training).
Adobe isn't the only major player moving in this direction — some enterprise tiers of Microsoft Copilot and Google offer similar forms of indemnity too — but it remains the most explicit among the big names aimed specifically at the creative world in stating all three things together: training on licensed content, contributor compensation, traceable indemnity.
This technical angle is vital. When I'm talking to clients like Berto, the Campanian craftsman who just wants the job done, or Bobby, the Manchester manager who reads contracts with a magnifying glass, this is the option that holds up best against reality. If a client asks me today, "Umberto, but is this generated stuff actually above board legally?", I can answer in a few checkable lines: Firefly is trained on licensed Adobe Stock content, contributors get paid, and — with the enterprise entitlement — Adobe indemnifies the client against IP claims on the generated outputs, save for the contractual exclusions. It's not fluff. It's documented accountability.
What changes for those of us working between Caserta and London
This technological and legal fracture isn't just theory debated in courtrooms. It's changed how I sit down at the meeting table, and it carries a radically different weight depending on the postcode of the client sitting across from me. The comparison isn't a flourish; it's the litmus test of the market.
In Italy: Berto and the wrong question
In Italy, I work with clients like Berto. Business owners with rough hands, who still remember the smell of the factory floor, perhaps making pasta in Gragnano or running a canning business in Portici. For Berto, "stock photo licences" on an invoice has always been an annoying mystery. A cost he couldn't make sense of. "Why do I have to pay two hundred euros for a photo of a wheat field when there are thousands on Google?", he'd say ten years back. These days, Berto's discovered ChatGPT on his phone. He's watched his son generate images on Midjourney. And his question has changed: he no longer asks you why stock costs so much. He asks, jabbing a finger at you: "Hang on, can't you just generate this background image yourself, for free, on the computer?".
The trouble is Berto doesn't know what he's really asking for. He wants speed without weighing up the risk, he wants lower costs without understanding rights usage. To him, a generated image is just a JPEG he shouldn't have to pay for.
In the UK: Bobby and the culture of preventive risk
In the UK, the picture's the polar opposite. I work with counterparts like Bobby, running businesses between Manchester and Hackney. Over there, preventive legal culture is fierce. IP indemnity clauses aren't tucked away in the footnotes — they're the heart of the negotiation. The Getty v Stability ruling out of London, which I mentioned earlier, has made generative AI a hot topic, already on the table in English contract conversations in a way that, in Italy, simply hasn't landed yet. In Campania we're still arguing over whether AI is stealing designers' jobs (if you've had that doubt yourself, have a read of L'AI sta uccidendo il graphic design? No.). In London they're arguing over who foots the solicitors' bill if an algorithm's plagiarised a shot.
Same delivery, two worlds: Southwark versus Santa Maria Capua Vetere
Let me tell you about a technical scene that played out a few weeks back. I was working on a visual asset for a campaign for a client based in Southwark, near Borough Market. When it came to delivering the final files, their legal department sent me a four-page form where I had to justify in writing which generative engine I'd used, certify that the base images contained no prompts infringing third-party copyright, and confirm I'd used a commercially safe engine with corporate indemnity.
I didn't have to invent anything on the spot. I already had the Firefly documentation sitting ready in a drawer, exactly like a reportage photographer who always carries two spare SD cards in his pocket: you never know when you'll need them, but when you do, there's no time to nip out and buy some. I replied attaching the Adobe clauses in twenty minutes. Job done. A week later, I deliver a similar job for a client in Santa Maria Capua Vetere. Same type of asset, same generative technique. I send the files over. The reply was a WhatsApp voice note: "Beautiful, Umbè, stick 'em straight on the site." The question of where the image came from never even came up. Two worlds, the same trade, two levels of awareness a world apart.
The real workflow: from generation to a file ready for the web
Theory's all well and good for conferences. But when you've got a 6pm deadline and a blank Photoshop canvas staring you down, you need a method. The stock platforms sold us the illusion of the instant click. The reality of AI generation in the studio is much closer to old-fashioned workshop graft. Here's the real workflow I use, from generation through to the finished file ready for publishing, no filter.
Stage 1: choosing the engine and drawing the line
It all starts with one golden rule: if the image needs to show a close-up portrait of a professional looking down the lens, or a real product belonging to the company, the generative engine gets switched off. Never generate subjects that need to convey direct human trust. Premium stock photography or a dedicated shoot remain unbeaten there. But if I need an atmospheric image, a complex texture, a conceptual background (e.g. "a server lit up blue, wrapped in mist"), then the generative terminal opens up. Which engine I pick comes down to safety: commercially safe systems if the client's corporate, open systems only for internal concept boards.
Stage 2: iterating the prompt (and the patience of pasta)
The idea that AI nails it first go is a fairy tale for people who don't actually use it. It's like Gragnano pasta: it needs to cook for exactly the right time, not a minute more, not a minute less. In my experience, the first generation only tells you how the machine's misread your brief. The second sorts out the lighting. The third fixes the perspective. The fourth adds the missing detail. In practice, a usable image for a professional campaign never comes out before the sixth or seventh targeted generation. That means spending time calibrating word weights, ruling out negative terms, forcing specific aspect ratios.
Stage 3: relentless post-production
A generated image is never finished as it stands. Never. It always has that plasticky aftertaste, that artificial contrast screaming "I'm fake" a mile off. This is where the graphic designer's craft comes in. I import the generated image into Photoshop and take it apart. I apply a heavy colour correction to bring it inside the client's brand palette. I add digital noise, film grain to break up the sterile perfection of the synthetic pixels. I fix inconsistent details by hand. I often run an upscale with dedicated algorithms to make sure it holds up for print or on high-res retina screens. (For more on this specific clean-up stage, as distinct from generating from scratch, I'll point you to Denoise, Firefly o Luminar Neo: there I talk about post-production on real shots, whereas here the raw material's synthetic from the off).
Stage 4: converting and cleaning up the metadata
Before publishing, the last technical stretch. Files from AI engines carry the prompt information or the platform's proprietary tags in their EXIF metadata, along with the C2PA provenance tag (Coalition for Content Provenance and Authenticity), which verifiably states the image was generated or altered using AI. A year ago a bug in my publishing editor was corrupting those very tags on upload, which had me briefly thinking I'd need to strip them out — a bug that's long since been fixed. Now I always keep the C2PA tag, deliberately: it's the proof of transparency I want on every file I publish. I only strip out the superfluous EXIF metadata (the text prompt, the platform's internal tags), fill in the correct IPTC fields for SEO, and convert to modern formats (WebP or AVIF) to keep things light and fast-loading. Transparency about AI provenance isn't optional.
TECHNICAL BOX: Text-to-image vs image-to-image
In a context of consistent branding, writing text (Text-to-Image) is only useful at the outset. To scale up, I use Image-to-Image. I take the first generated image or a real shot from the client, load it up as a style reference or structure reference, and ask the AI to generate variations from there. It's the only way to get 10 different images for a website that all look like they were shot with the same lens, at the same time of day, by the same photographer.
📦 SUMMARY BOX: the 3 questions to ask before generating an image for a brand
- Is the engine I'm using trained on licensed data or on open scraping? (Your contractual safety hangs on this).
- Who takes the legal hit if the output triggers a dispute? (If the answer's you, switch engines or get a waiver signed).
- Does this image genuinely need to meet a real need for authenticity (a person, product, specific place), or is it an atmospheric image that could be swapped out? (If it needs to be true, switch off the AI and pick up the camera).
Conclusion
There's a huge misconception floating around communication agencies and freelancers alike: believing that using generative AI to create business visuals is a shortcut, an acceptable compromise born of laziness. It isn't. In my experience, and looking honestly at the budget figures I manage between Italy and the UK, generative AI is — in the right cases — the technically cleaner and more honest choice compared with a stock market that's rotten from the inside by now.
I stopped paying for traditional stock photography because the trust that underpinned it broke down. The platforms charge a toll to access images that have been purged of any humanity, iterated endlessly by algorithms trained on the underpaid or unpaid work of photographers I used to respect. And if I need a synthetic image, I'd rather generate it myself, in my studio in Santa Maria Capua Vetere, with full control over the genesis of the prompt, the virtual lighting and the final post-production, rather than fishing blind in an endless catalogue whose true origins nobody's tracking any more.
AI isn't a fallback. It's a new scalpel in an operating theatre that had stayed faithful to the flick-knife. It demands skill, knowledge of the regulations, and a deep professional ethic. When a client looks at the final result today, they shouldn't see the effort behind the prompt or the copyright doubts; they should see a clear message, correctly positioned, speaking exactly to their target audience.
Will real photography survive? Absolutely. It'll survive wherever there's a need to document what's real: the sweat on Berto's brow in his factory, the rough texture of his pasta, the genuine expression on an employee's face, the smell of the workshop. Everything else — the filler aesthetic, the backdrop, the blurred-out background of visual mediocrity — has already been swallowed up by the infinite catalogue. And frankly, I've got better things to do than pay for a ticket to a museum where the paintings change themselves every time you turn your back.
Frequently asked questions
Are AI-generated stock images legally safe to use in a commercial campaign?
It depends entirely on how the engine was trained. "Commercially safe" platforms like Adobe Firefly, trained only on licensed or public domain content, carry very low risk and often include legal indemnity for Enterprise clients. Open engines, trained through indiscriminate web scraping, carry higher residual risk, as shown by the Getty Images v Stability AI case.
What exactly did the Getty Images v Stability AI ruling establish?
On 4 November 2025 the High Court in London dismissed Getty Images' main claim for direct copyright infringement, ruling that a trained AI model no longer contains the original images once training is complete. Getty did, however, win partially on trademark infringement, because some generated images reproduced the Getty watermark present in the training data.
What's the difference between a "closed" AI engine like Adobe Firefly and an "open" one like Stable Diffusion?
A closed engine is trained exclusively on content owned by the platform, licensed, or in the public domain, drastically reducing legal risk, though with some technical limits when rendering complex detail. An open engine is trained by scraping vast quantities of images from the web, including copyright-protected work, offering greater creative freedom but higher residual legal risk.
Do photographers get paid when their images are used to train an AI model?
Only in some cases. Adobe, for instance, pays Adobe Stock contributors whose material was used to train Firefly, through the Firefly bonus scheme. Not every platform adopts this kind of compensation model for training, and it's worth checking before choosing a tool for professional use.
When does it NOT make sense to use an AI-generated image instead of a real shot?
When you need to represent a real, identifiable subject: a professional portrait, a specific client product, a recognisable place tied to the brand. In these cases the authenticity of the shot can't be substituted, and a generated image risks introducing technical inconsistencies or, worse, sending a message of fakery where trust is what's needed.
Why keep the C2PA tag on AI-generated images rather than removing it?
The C2PA tag (Coalition for Content Provenance and Authenticity) verifiably certifies that an image was generated or altered using AI tools. Keeping it is a transparency choice towards the client and the reader, and it's in line with the disclosure requirements for generative AI use demanded by some platforms and regulations.
Why is the stock photography market consolidating right now?
Because generative automation drastically cuts the cost of producing a catalogue, making it more cost-effective to buy up smaller competitors than compete with them. A concrete example is Shutterstock's acquisition of Envato in 2024, which happened just as both platforms were folding AI generation tools into their catalogues.
By accepting you will be accessing a service provided by a third-party external to https://insightadv.it/
IT
EN 















































Comments