Every design tool you've ever used has trained you to expect certainty. You pick a colour, you get that colour. You set a margin, it stays set. So it's no wonder that when AI gives you something close to what you wanted — but not quite — it feels broken.
It's not broken. You're just reading it wrong.
What's actually shifting
There's a growing conversation around what some are calling "probabilistic design" — a mindset that challenges how we interpret AI outputs. The core argument is straightforward: stop treating AI-generated suggestions as right-or-wrong answers. Start reading them as weighted possibilities, each carrying a different degree of confidence.
Think of it this way. When an AI tool suggests three layout options, it's not saying "these are the three best layouts." It's saying "based on patterns I've learned, these are three plausible candidates — and I'm more confident about some than others." That's a fundamentally different proposition. And the way you evaluate those options should reflect that.
Meanwhile, Adobe has rolled out a wave of updates across its ecosystem aimed squarely at the control problem — that maddening gap between what you intend and what generative AI actually produces. The goal: more predictable, reliable results that keep designers in the driver's seat rather than white-knuckling through random outputs.
Why this changes how we work
These two developments seem like they're pulling in opposite directions. One says "get comfortable with uncertainty." The other says "here's more control so you don't have to be." But I think they're complementary — and understanding why is the key to working well with AI.
Adobe is right that the tools need to improve. You shouldn't have to fight your software to get a usable result. But better controls don't eliminate uncertainty — they narrow it. You're still working with a system that generates outputs based on probability, not deterministic rules. The slider might get more precise, but it's still a slider, not a switch.
Here's the thing we don't talk about enough: in design, we're trained on deterministic tools. Draw a rectangle, get a rectangle. Every time, exactly the same. But AI is fundamentally non-deterministic. The same prompt, the same settings, can produce different results on different runs. If you walk into that expecting the reliability of Figma or InDesign, you'll be perpetually frustrated — and you'll blame the tool when the real problem is the expectation.
The designers who'll do the best work with AI aren't the ones who master the cleverest prompts. They're the ones who learn to think in ranges rather than absolutes. Who build systems that flex rather than snap. Who treat AI outputs as starting points for evaluation, not finished deliverables.
This has direct product implications too. If your product uses AI to recommend, predict, or personalise anything, you should be communicating confidence levels to users — not just bare results. A "you might also like" recommendation that's 90% confident is a very different signal from one that's 40% confident, but most interfaces present them identically. That's a design failure. It comes from not thinking probabilistically about what AI is actually doing under the hood.
Tool spotlight: Adobe's control updates
Adobe's latest updates deserve attention if you work in their ecosystem. The headline feature tackles precision in generative imagery — giving you more reliable outputs that better match your creative intent. If you've found Firefly results feel like a coin toss, these are aimed squarely at that frustration.
My honest take: better controls are necessary but not sufficient. The real upgrade isn't in the software — it's in how you interpret what comes back. Use Adobe's tighter controls to narrow the range of possibility, then apply your own eye and judgement to evaluate the results. The tool gives you better odds. You still have to make the call.
One thing to try today
Next time an AI tool gives you something unexpected, don't immediately re-prompt or tweak the settings. Pause. Ask yourself: is this actually wrong, or is it just different from what I assumed?
Start noticing which results feel 80% right versus 40% right, and pay attention to what separates them. Train yourself to evaluate AI outputs as probabilities, not prescriptions. That single shift in thinking will make you more effective with every AI tool you touch — not just the ones with better sliders.