I have been skeptical of AI logo tools for two years. Not because I am protecting my job — honestly, if a €30 tool could reliably produce a good brand identity, I would want to know. I would factor it into my workflow and price accordingly.

So I ran the test properly. Same brief, three tools, no cherry-picking. I am writing this the day after, while the results are still annoying me in specific and useful ways.

The brief

I kept it honest. Not a trick brief designed to expose AI weaknesses — the kind of brief I actually receive, stripped to its essentials.

A solo brand and UX designer based in Marseille. European sensibility, not corporate American. Works across digital and print. The mark needs to read clearly at 16px as a favicon and hold up on a business card at 8mm. No gradients — vector only. Personality: precise, warm, slightly editorial. Not a tech startup. Not a creative agency. Something in between.

That is a reasonable brief. Specific enough to constrain the output, open enough to allow for interpretation. I ran it through ChatGPT with image generation, Midjourney, and Looka. I gave each tool the same text. No additional prompting, no iterating — first output only, to simulate what a non-designer founder would experience.

ChatGPT — technically competent, completely anonymous

The output was clean. I will give it that. The letterforms were consistent, the proportions were considered, the spacing was not embarrassing. If you printed it at 300dpi, nothing would break.

The problem is that it looked like every other logo generated by a language model trained on design portfolios. The mark was a geometric monogram — a stylized "A" or "AO" depending on how you squinted — set in a sans-serif that felt vaguely like Futura but was not quite Futura. There was a thin rule underneath. The color was a dark navy that suggested professionalism without committing to anything.

It answered the brief technically. European sensibility? Sure, in the sense that nothing in it was aggressively American. Warm and editorial? No. It was neutral to the point of invisibility.

The more specific problem: the file it produced was rasterized at 1024px. Vector only was in the brief. When I pointed this out, it generated another version — also rasterized. It does not actually understand the constraint because it does not understand what a vector is at an operational level. It understood the word.

Midjourney — beautiful, unusable

Midjourney produced the most visually striking result. It also produced something that exists only as a dream.

The mark was genuinely interesting — an abstract form that played with negative space in a way that felt considered. The rendering was immaculate. The color palette was a specific warm charcoal and off-white that I have not seen used that way before.

It is also completely fictional as a logo. The letterforms it generated have no relationship to any real alphabet. They are aesthetically derived from letters but they are not letters. If you tried to redraw them in Illustrator, you would be designing a new typeface from scratch. The proportions that make the form beautiful at Midjourney's render resolution completely collapse at 16px — which was explicit in the brief.

Midjourney is extraordinarily good at generating images that look like logos. It is not good at generating logos, because logos are not images. They are systems. They need to work at 8mm and 800px. They need to be handed to a printer who will ask for a PDF with outlined fonts. Midjourney has never seen a printer.

The output was the most inspiring thing I received from the test. It was also the least useful.

Looka — the honest one

Looka is a purpose-built logo tool, which means it plays a different game. It does not try to generate a mark from scratch. It combines existing graphic elements — icons, fonts, layouts — based on the inputs you give it, then presents a range of options.

The output was predictable in a way that is actually its strength. Within eight minutes I had twelve variations, each technically correct: proper vector format, real fonts with proper licensing, scalable marks, color system built in. Nothing was visually exceptional. The icon options were generic — a small diamond, a minimal monogram, an abstract leaf shape. The typography was safe.

But here is what Looka understood that the others did not: it separated the mark from the wordmark and showed me how they work together. It gave me a favicon version automatically. It showed the logo on a dark background and a light background. It thought about deployment, not just appearance.

If I were a founder with no design budget and a launch in two weeks, Looka is where I would start. The output needs a designer to make it good. But it is a real starting point, which is more than I can say for a beautiful hallucination.

What this actually means

The framing that designers use — "AI is coming for our jobs" — misses where the actual risk and the actual opportunity are.

The risk is not that AI will replace good logo design. None of these outputs would have passed a real creative brief review. The risk is that clients who do not know what good looks like will use these tools and believe they have a brand. They will launch with a Looka mark that has not been adapted for their context, or a ChatGPT PNG they call a logo, and wonder in eighteen months why nothing feels cohesive.

The opportunity is the other side of that same observation. The gap between what these tools produce and what a real identity system does — how it behaves across contexts, how it carries personality at every scale, how it builds recognition over time — that gap is not closing as fast as the discourse suggests. It is where the work actually lives.

Where AI is genuinely useful in the logo process: early exploration and concept pressure-testing. Feed a brief into Midjourney and you get a visual vocabulary — color temperatures, formal languages, spatial relationships — in minutes. That is legitimately useful for a designer who knows how to read the output and extract the signal from the noise.

Where it is actively dangerous: as a final output given directly to a client who will treat it as a finished brand.

The question I keep coming back to

After running this test, the thing that stayed with me was not the quality of the outputs. It was the speed at which all three tools gave me something that looked like a decision.

That speed is seductive. And it compresses the time a client is willing to spend sitting with a brief before asking to see something. Which means the designer's job in 2026 is partly to protect the thinking time — to explain why the best answer does not come back in eight minutes, and why the thing that comes back in eight minutes is not the best answer.

That is not a new argument. Clients have always wanted things faster. But the tools now make "faster" look like "done," and that is a different problem.

The logos the AIs made are fine. The conversation they are replacing is not.