Illustrators are screaming “bloody murder” at generative AI, writers are striking at Large Language Models, specific AI laws and regulations are being drafted and discussed and implemented all over the world. In the meanwhile AI is everywhere, absolutely everywhere, people on all kinds of fields are reporting productivity boosts, cost reductions, marketeers and spammers are having the best time of their lives, and providing generative AI is approaching, at record speeds the shortlist of biggest grossing business models in human history. Social networks and the web are being flooded with AI content, to the point of redefining the value of content itself, and the politicians… Well, let’s not even go there. Something is going to happen, are you stressed already? Let’s think about this:

Take artists as an example, we’ll go on a trip where we explore what is really happening to them now, then we try to understand a bit these AI models, what they do, and why the art that was “stolen” can’t be “unstolen” anymore. Finally, we use that knowledge to imagine a possible new business model that can be (at least) a first step towards a fairer world and dampen the chaos.

Listening to the wind of change (The Scorpions)

A global survey of more than 10,000 freelance designers found that 61% say AI has already affected their income, and interestingly, more reported gains than losses. Over half of those designers now actively use AI to streamline workflows or spark creativity and are expecting income increases in 2025.

Meanwhile, in the UK, a survey of nearly 800 illustrators revealed that 26% have lost work to AI, and 37% have seen their income drop. Only about 12% of illustrators are using AI tools themselves, which means most are at a very real risk of losing their livelihoods.

What emerges is a clear, growing divide: commercial designers and agencies are embracing AI as a powerful tool to enhance their craft and boost income, while many traditional illustrators and other fine-artists are struggling with lost jobs or diminishing pay. This is not just about adoption though, due to the state and capabilities of current tools and business specifics, It might be easier for some to take advantage without being “replaced” while others might have to heavily adapt their business and skill in order to survive.

There is also a cultural/political barrier there. Fine-artists are aggressively responding to these changes, which is an understandable reaction. Accepting adoption might make their court cases and their (very necessary) legislation lobbies weaker. As a side-effect, if you are an artist and your reaction is to try to understand the situation by experimenting with it and/or integrate AI on your workflow, even if it is to artistically criticize it, you’ll very quickly feel threatened by your own community and tagged as a traitor, a thieve or a “non-artist”. There’s a fear that there is no turning back from publicly sharing anything that AI has touched in any way. This doesn’t help us come up with co-operative solutions.

Life, uh… finds a way. (Dr. Ian Malcolm, Jurassic Park)

When a transformative technology arrives, like the internet did in the 90s, it starts in chaos. No one knows how it should be monetized and big disruptions start to emerge. But once a profitable model is understood, it takes over and stabilizes the system. With the internet, that model turned out to be advertising.

Ads now pull in over 1 trillion dollars globally a year, with around 800 to 900 billion in digital ad spending, a sum that continues to surge 7% to 9% every year.

Let’s keep in mind that the internet completely disrupted the main distribution channels of advertising, newspapers, TV and Radio.

Nevertheless, digital advertising thrived because it captured and monetized what online platforms inherently gather: behavioral data. Every click, scroll, or video view became a monetizable action, fueling the massive attention economy.

Same-same… but different… but still the same (James Franco / Meme culture)

Fast-forward to today: AI didn’t arrive empty-handed. Unlike the early internet, it started with users already paying. Subscriptions, API access, enterprise contracts, revenue streams emerged from “day one”, under the control of a few massive companies that completely control the technology and the resources needed to operate it.

Only in a few years, the generative AI market is estimated in the hundreds of billions, quickly catching up with advertising, with projections reaching $3–4 trillion over the next decade. Those are insane numbers.

But here’s the twist: While the monetization is clear and seems to scale very rapidly, the two “fuels” that keep these machines running are drying up quickly. These “fuels”, in my opinion are the key to the emerging business models: GPU power (micro-chips) and training data (content).

He who controls the Spice, controls the Universe (Vladimir Harkonnen, Dune)

So, about that “fuel”, the GPU part is boring in this context, even though Sam Altman predicted it would replace money all together. That’s for giants like NVIDIA, OpenAI, AMD and Google to fight about and throw money at… But training data, that’s what I want to really talk about: Up till now, the internet provided AI model trainers with an infinite pool of human created content, all topics, all formats, mostly commented and contextualized, lots of gray copyright situations, lot’s of cleanup needed, but all “free” to use (at least in secret). It was a total feast.

For artists, actually providing their art online, for free, as much as possible was pretty much the only way to get work, to the point of clients using the infamous “Exposure as a currency” bargaining tactic.

A bit of a tangent on privacy here: If you ever think (we all do… mostly, I’m not judging): “Privacy?? Why should I bother? I do nothing wrong!! I am proud of what I do!” well… things change, and every time you give up your privacy, you’re giving up some power, so you better get something back that is worth it. Keep in mind, you never know the future price of what you are giving away. But well… here I am writing this for free… Artists have been giving up their work in exchange of advertising for their “services”, notoriety, and community. And that was OK, but then AI came, and had a feast on the data, learned their unique skills and now can do amazing stuff that only you could do before. And they don’t want to pay you anything for it. I know this is enraging, seems unfair, unethical, and might even be illegal in some cases… but that damage has been done beyond repair. Your past work, your skill, has been “stolen” and used up, you can sue and even get (maybe, in very rare cases) a small compensation, but I believe (and that’s the whole point of this article), that protecting and monetizing your future work and your place in the creation of these technologies is where the value is now.

Ok, that brings us back to our previous point… training data. What seemed like an infinite buffet is quickly becoming a problem. AI researchers mostly support the “scaling theory” that basically says, the more data, and the bigger the model, the better it will perform. It’s all about size and quantity. The more the better. That together with the potential of the tools and the enormous investment being made, models grow fast and consume pretty much all the “low hanging” data on the internet. We do create a lot though, every day about 720.000 hours of video are published (on YouTube alone) and about 2 million news articles are released on the internet. But now the value of this data is known and people are fighting hard to legislate and block companies from using this data to train AI. The pressure is on. Soon AI trainers will run out of quality data to use and the crazy pace of advancements might slow down.

There are some mitigation strategies already in motion to deal with this data scarcity, one is putting effort into developing strategies and techniques to improve the models beyond just training, and the other is using the models themselves to generate synthetic training data. But these are not without major risks and they wont scale well enough. The best bet for AI companies, in my opinion, is to generate incentives for creative people to keep creating, more and more, and let them use the data to train new models. This means giving creators a fair piece of the pie. If there is a cooperative environment, AI companies will get ultra-high quality data, in a controlled annotated and contextualized format under their own control, which is the absolute dream for researchers.

It’s not a simple problem, but we need to start somewhere, and this is where I would like to contribute with thinking.

So what do we do now?

The same thing as every night, Pinky, take over the world! (Pinky and the Brain)

So now we know where the value lies and we know that, to make the best out of AI, we should find a way to work together. That makes it the time to start thinking of business models that align all the incentives. I’ve searched and searched, asked all the AI models, but couldn’t find one that made me happy, so I went into my thinking-pod (aka. the shower) and got to work… Here’s what I came up with:

What if we augment generative AI models with a new type of social network that serves the purpose of providing data to train the models while still serving the popular purpose of social networks: to allow people to be entertained, communicate and share information (and, unfortunately, the more nefarious political mass mind control we see these days). Enter the Generative Network. It’s basically marrying ChatGPT with Instagram in holy matrimony.

We would decouple the current incentives of social networks to serve advertisers by capturing attention and spreading marketing messages and replace those with generating the most creative, useful or varied human-powered data for AI systems to feed on. The money would flow in from the usage of the AI model (with subscriptions, enterprise deals, pay-per-generation, etc…) and part of the pie would be fed back to the creators on the network for providing continuous creative input in the form of high quality, properly licensed, multi-modal, training data for the models to keep improving.

This would look like a normal social network but less constrained in terms of format and algorithm. People would post pictures, paintings, music, spoken word, poems, full books, stories, videos, films, documentaries… and then they would be paid in micro-transactions (just like monetized accounts on youtube for example) but instead of by the amount of views or engagement, by relevance to what is being generated on the AI model side of things. All accounts in good standing would be monetized and would be paid whenever someone generates something that could have been “inspired” on their work.

The problem is, unintuitively, that when a model generates something, no one can tell where the “inspiration” come from in the training data.

There is no spoon (The Matrix)

The training data (your art, be it images, text, audio, or whatever else) is not stored inside the AI model. ChatGPT, for example, does not keep a massive database behind the scenes from which it copies and pastes content. Instead, it generates completely new and original outputs based purely on mathematical operations optimized to produce results that make sense to us, somewhat like our brains do.

Training data is a consumable, it’s used in the training phase, before the model is ever released, and then effectively “discarded.” Once the model is trained, it no longer directly needs or uses your original data.

Let’s quickly go through it cause there is one term I need you to understand:

Initially, the AI model simply outputs random noise, no matter what you ask it. It can’t produce anything meaningful. Then, during the training phase, the model sees many examples of inputs (prompts, images, text, etc.) and their corresponding “appropriate outputs” (answers to the input, aka.: your art) Here’s where something called embeddings (remember this word!) comes into play.

Think of embeddings as a way to convert the input data (words, pixels, sounds) into lists of numbers, so that AI can “digest” and process them mathematically. These embeddings group similar concepts close together in mathematical space (think of a graph but instead of 2 dimensions, it has thousands), helping the model recognize patterns and relationships between ideas, images, sounds and words. To put it simply, conceptually related stuff sticks together numerically. “Cat” will be close to “Furry”, “Kitten” or “Animal” in most dimensions, and a bit further away from “Tree” or “Computer”. Although in some other dimensions, “Cat” might be relatively close to “Computer” because of the crazy obsession the internet has with cats.

The model then slowly fine-tunes its internal math operations to transform any input into something as close as possible to the appropriate output. It does this billions or even trillions of times with lots of different inputs/outputs until it gets remarkably good at predicting the “right answer.” Once training is complete, those mathematical operations are “frozen,” and the model is ready for use.

In the inference phase, when you actually use the model to generate new stuff, you provide it with a new input it has never seen before. The model converts this input into embeddings, runs its learned mathematical operations on them, and generates an entirely original output based on predictions of what would be a suitable response.

Models are far more complex under the hood, but here’s the essential takeaway: your data is not stored or remixed by the model. Instead, AI learns by example to predict how you or any other artist might respond to a similar request. Because of the complexity of these mathematical operations and how the values are tuned over millions of examples, no one can precisely identify the exact source of inspiration behind any specific output, especially because every piece of data on the training set changes the values (weights) set by the others, influencing the whole system. This means the model needs all the data to generate anything at all, just like an artist needs everything they ever saw, experienced or learned to make a painting.

I’m Winston Wolf. I solve problems (Pulp Fiction)

Although we can’t know exactly what part of the training data contributed to a generation, that might not really matter, we just need to find a way to give the creators their share of the pie and incentivise them to continue creating. Under that premise, we could definitely get by using the embedding part of the model in a smart way.

What if we feed whatever is generated back to the embedding part of the model, turning it into the coodrinates to the multi-dimentional map in the mind of the model and find the pieces of content on the Generative Network that are the closest, and then pay those creators a bit of money. It would not be exact but it would be a great (and fair) way to decide who to pay, and how much, for the provided training data. It would be an automated “credit” system that would be hard to mess with since it’s frozen as part of the model. And it would also work just as well to “pre-pay” for data that is not yet used for training. Imagine you ask the model for a unicorn flying over a field in the style of Van Gogh, this would likely “credit” paintings similar to Van Gogh but also anything with unicorns or fields, including texts that describe what a unicorn or a field looks like or even anything with birds (because they fly too and have wings).

In this world, parallel to social networks, “the algorithm” would be the embedding model and the success function would move from “what people want to see” to “What people want/need to generate”. For AI companies, it would be a way to differentiate from others, right now, the models are very different but the training data is mostly the same, by running and supporting the generative network, the AI companies would stay in control of their own training data and therefore their “moat”. This would also make it very appealing to invest on making the generative network a success for it’s users/creators. For artists, this could mean getting paid to truly develop your art in your own style on whatever direction you want it to go. It’s creation for the sake of creation. And there is a lot of potential for extra revenue channels (read the extras if you’re interested)

I hope this all made sense, It’s a new idea, still constricted to my brain and a handful of others, I would love to discuss it with whoever wants to, so reach out in LinkedIn, X, BlueSky or Instagram.

Cheers, and good luck!

Extras

If you found this idea interesting, there is plenty more to it that would not fit most readers attention span or likely interest. But if you are one of those that want to dig deeper, then here we go, let’s geek-out:

Let the games begin (Ancient greeks at the Olympics)

The interesting part of building the business model like this is that it has so much room for mechanics and dynamics (gaming industry language). So many things to play with to balance out all the incentives and create something truly better then any existing social network:

• Pricing would be key of course, just balancing the amount paid out by generation depending on the income generated. Also works the other way, the AI model could charge the users dynamically more (or less) for asking things that fall within a “crowded space”.
• Adjusting the radius of the payout depending on how crowded the space is. This would incentivise people to create truly original work by “going” into areas that are not very crowded.
• Styles, prompts and original artwork could be featured and advertized through the social network. Creating a controllable feedback environment to encourage popular use-cases where the model currently fails.
• Data labeling and context enrichnment by using likes, views, and engagement metrics.
• Artists incentivised to have a lot of art in their own style so a generation hits most of his works at once. Accumulating more value quickly.
• Fine-tuned, specialized versions of the models for specific contributors/artists pay out more to the source.
• “Hire” the closest artist to tweak/curate/re-generate/overpaint/brainstorm on a generated piece.
• Embedding could also help detect, and flag copy-catting (In any case, at least original will also get rewarded) This is also an incentive for artists to get into the system and be “registered” as the original creators of something.
• De-monetizing areas of the embedding map that are dangerous or unethical such as political misinformation.

I could just go on and on… but it’s late. Let’s sleep on it.