When Everything Starts Looking AI-Generated

I started recognizing “the AI look”

Some time ago, I began seeing images that gave me an odd feeling. They did not need a watermark, signature, or explanation to feel familiar. Not because I could reliably tell how they were made —I do not think that is possible from a result alone— but because they repeated a combination of visual decisions I had already seen too many times.

The same product hovering in a dark scene. The same extreme contrast. The same carefully placed edge glow. Particles arriving to confirm that, yes, this is futuristic. Huge type announcing an epic story the product may not need. Everything polished; everything strangely interchangeable.

This is not exclusive to generated imagery. The internet knew how to copy formulas long before I could ask a tool for twenty variations in a minute. The difference is speed: a recognizable visual formula can now appear, multiply, and fill a feed before anyone asks whether it actually belongs to the brand, the product, or the problem.

The paradox of nearly infinite options

The promise of these tools is enormous: more alternatives, more sketches, more references, more ways to test directions that once needed more time or budget. That promise is real. I use AI every day for research, writing, debugging, interface prototypes, comparisons, and getting unstuck.

The paradox appears when a tool with almost unlimited possibilities is used with the same kind of request and approved with the same kind of judgment. Then we do not get unlimited directions. We get a very efficient factory for attractive averages.

The output does not have to be bad. That is exactly why it is tricky: it is often good enough. It is well lit, well rendered, and ready to post. But if changing the logo, product, or campaign name would not meaningfully change the composition, it may not be a direction. It may be a mold.

The prompt has an aesthetic too

Words like “professional,” “premium,” “modern,” “eye-catching,” and “for social media” look like instructions, but they are often entrances to a very crowded place. They are not bad words. They are simply vague. If I do not explain who a piece is for, what it should communicate first, what tension it should hold, which conventions it should respect, or which ones it should avoid, I am asking for an average solution with a high-quality finish.

There is no need to assign mystical intentions to a model to understand this. Tools respond to patterns and context. If my context is generic, my output is likely to be generic too. A prompt is not a secret spell; it is a record of which questions I asked before I asked for production.

Sometimes the answer to “make it more premium” is not more metal, more shadow, and a more expensive gradient. Sometimes it is stepping back and asking what premium means in that specific case: precision, calm, materiality, time, service, trust? A tool can help explore those possibilities. It cannot choose which one matters without someone defining the ground.

Producing is not deciding

Using AI to produce is not the same as using AI to decide.

AI can help me produce a great deal: an alternative list, a first structure, an interface prototype, a copy variation, an implementation hypothesis, or a consistency review. It can also make quick work of explorations that used to take too long to do by hand.

Decision-making is a different kind of work. It means rejecting a technically immaculate solution because it has no voice. It means keeping a convention because it improves reading. It means noticing that an effect gets attention but steals attention from what matters. It means reviewing a result at 375 pixels, with a keyboard, without JavaScript, or with reduced motion. It means admitting that something spectacular can still be wrong.

Delegating production can be a very good idea. Delegating judgment without review is, at minimum, a fast route to something that resembles a lot of what already exists.

Conventions are not the enemy

I am not interested in turning this into an absurd cult of originality. Conventions do useful work. A button should look actionable. Food advertising is allowed to make food look appetizing. A payment screen should reduce doubt before it tries to surprise anyone. Patterns exist because people have learned to read them.

The issue begins when a convention is used in place of a decision. When every launch needs the same hyper-dramatic hero, every product needs particles, and every “technical” interface needs a cyan line network even when nothing is being connected. A convention stops being shared language and becomes automatic decoration.

A piece does not need to reinvent a button to have personality. It can keep clear hierarchy while taking specific decisions about composition, rhythm, material, language, or interaction. Originality is not making everything different. It is making something answer this case rather than any case.

It happened while building this portfolio

This is probably the uncomfortable part to write. While building this portfolio, some requests were very tempting: “more creative,” “more spectacular,” “more unhinged.” With enough energy, those instructions quickly produce panels, boxes, cyan lines, dashboards, generic technical systems, and a respectable amount of glow. As a capability demo, they work. As identity, not necessarily.

The work improved when I changed the process. Instead of asking for more intensity in the abstract, it became: idea → direction → mockup → critique → visual reference → implementation → comparison → correction. The difference was not that AI stopped participating. It participated a lot. It helped generate, discuss, explore, and build.

What changed was that the first output stopped having special status. Home found a more specific relationship between portrait, rail, and exits. Approach stopped being text with an illustration and started showing INPUT → CONTROL → FLOW. Work stopped arranging projects as equal cards and started showing capability → committed route → evidence. Not everything worked on the first try, and that became useful information rather than a failure of the process.

The decision-machine note tells that change from the system-building side. It also connects naturally to my note about discovering GSAP and wanting to animate everything: a new tool can make every surface feel like an excuse to use it. Judgment begins when I ask what happens if I do not.

The first response does not have to be final

Technically good outputs have a very comfortable trap: they feel like they close the conversation. An image can have light, volume, a coherent palette, and an apparently resolved composition. An interface can have smooth animation and acceptable hierarchy. None of that automatically makes it memorable or appropriate.

Accepting the first answer because it is already “good” is understandable. There is time pressure, budget pressure, fatigue, and an overwhelming number of options. But precisely because convincing finishes are becoming easier to produce, it is worth saving time for one more question: which decision in this result would not have happened by default?

If the answer is hard to find, the piece may still be promising material. There is no shame in that. The problem would be calling it final because polishing is easier than thinking again.

When looking professional stops being enough

The ability to create visually polished assets is being democratized. That raises the floor of quality, and I think that is good news. People with less time, less budget, or less specialist training can make ideas tangible, communicate better, and learn faster.

It also changes the bar. If many pieces can already look professional with relative ease, looking professional stops being enough of a differentiator. Value starts moving toward selection: what not to show, what to prioritize, what tone to avoid, which detail to repeat until it becomes language, and which glow to turn off because it is not helping.

This is not an argument for replacing professions or declaring winners and losers. It is an observation about the work that remains around the tool. Technical quality opens the door. Identity decides whether someone remembers walking in.

What I am trying to do differently now

My practice is not a universal method. It is a discipline I try to apply when time allows. I use AI to explore more than one direction. I ask for variations with constraints, not just “better” versions. I compare against concrete references. I critique the results I like most. I change composition, discard effects, remove elements, and check whether every part has a function.

I also try not to confuse consistency with repetition. A visual language can include rails, nodes, materiality, and motion without forcing every route to become the same machine. Coherence should allow identity, process, and evidence to behave differently.

The question that helps me most is simple: if I remove this, does the experience lose meaning or only lose shine? Sometimes the answer is uncomfortable because it reveals that the most visible element was the least necessary. Sometimes it confirms that an unusual choice was carrying an important idea. Either way, looking with intent is more useful than adding another prompt.

I still want to use more AI, not less

I do not want an anti-AI posture. I want to keep using these tools more and better. I want them to reduce research friction, make alternative exploration viable, and help me see problems I would otherwise miss. It would be absurd to give up such a useful lever simply because it can also produce clichés at industrial speed.

AI can turn a mediocre idea into a technically impressive image in seconds. That is remarkable. It also means the finish alone no longer differentiates. The growing value may be in noticing when something resembles everything else too closely, and having the patience to say: this is not finished yet.

The goal is not to use less AI. It is to avoid using it as a substitute for judgment.