"Preserve the unique soul of your work." That's the promise Adobe is making with its new custom Firefly models — and it's either the most exciting development in brand design this year, or the most dangerously misleading framing we've heard in a while. Possibly both.
Let's talk about what soul actually means when we're talking about design.
Training AI on your style — what does that even mean?
Adobe's pitch is straightforward: you feed Firefly your brand assets, your visual language, your colour palettes and compositions, and it learns to generate outputs that feel like you. For studios and brands drowning in content demands, this sounds like a dream. Consistency at scale, without hiring ten more junior designers.
And honestly? For certain use cases, I think this could be genuinely useful. If you need fifty social media variations that all feel on-brand, or a quick concept that sits within established visual guidelines, a custom-trained model makes sense.
But here's where I get uneasy.
What Adobe is calling "soul" is really pattern. An AI can learn that your brand uses a particular type treatment, favours warm tones, tends toward asymmetric layouts. It can replicate those surface characteristics with impressive consistency. What it can't do is understand why you made those choices — the brief that led to them, the audience you were designing for, the trade-offs you considered and rejected.
The soul of a designer's work isn't in the output. It's in the reasoning. And that's the bit no model can train on.
Meanwhile, AI systems are getting harder to see through
While Adobe is making AI more personal, Microsoft is making it more complex — and potentially more opaque. Copilot is now orchestrating multiple AI models together, routing tasks between GPT and Claude depending on what it thinks will produce the best result.
From an engineering perspective, this is clever. From a UX perspective, it raises some uncomfortable questions.
If you're using an AI assistant and you don't know which model is actually generating your output, how do you calibrate your trust? Different models have different strengths, different failure modes, different biases. When the system is making that choice for you, silently, you lose a layer of understanding about what you're working with.
For designers building products that use AI — and that's increasingly all of us — this is a real design challenge. How do you communicate to users what's happening under the hood without overwhelming them? Do users even need to know? I'd argue yes, at least sometimes. Transparency isn't just an ethical nice-to-have. It's what allows people to make informed decisions about when to trust the output and when to question it.
This multi-model approach is going to become standard. The UX patterns we establish now for communicating model behaviour will matter for years.
The most powerful design this week used no AI at all
While we're talking about soul, let's look at a piece of work that has plenty of it — and not a single algorithm in sight.
Brazilian agency Porta took the country's flag and stripped out all its colour, replacing the iconic green, yellow and blue with shades of grey. The message: as Brazil loses its forests, its rivers, its natural world, it loses the very colours that represent them.
That's it. No generative fill, no custom models, no prompts. Just a simple, reductive idea executed with conviction.
It works because it's specific. Someone understood the symbolic relationship between Brazil's flag and its environment, and found the most economical way to make that connection visible. It's the kind of work that makes you stop scrolling — not because it's technically impressive, but because the thinking behind it is sharp.
This is what I mean when I talk about soul. It's not a visual style that can be trained into a model. It's a point of view.
The bigger picture
We're watching AI tools become simultaneously more personalised and more complex. Custom models that learn your visual language. Multi-model systems that choose their own approach without telling you. These aren't bad developments — but they do demand that we stay clear-eyed about what's actually happening.
Personalised AI generation is a production tool, not a creative partner. It can help you scale decisions you've already made. It can't make new ones worth scaling.
And as AI architectures get more layered — models talking to models, systems choosing systems — the designers who understand what's happening beneath the interface will have a significant advantage over those who just accept the output.
What to take away
If you're exploring custom AI models for brand work, go in with realistic expectations. They're brilliant for consistency and speed. They're not capturing your creative soul — they're capturing your visual habits. There's a difference, and it matters.
If you're designing AI-powered products, start thinking now about how multi-model systems change your transparency obligations. Users deserve to understand what they're interacting with, even if the explanation is simple.
And if you want a reminder of what real creative soul looks like — find that desaturated Brazilian flag. Sometimes the most powerful design is just a good idea, clearly expressed.
No model required.