The answer is no. Leaving it there would be dishonest. Something is happening to design work, and "AI doesn't have taste" is a comfort rather than an argument.
The question hides three claims. AI can make websites. Making websites was the designer's job. So the job goes away. The first is true. The second was never quite true. The third is where it breaks.
What it already took
Start with the strong version. Give a model an empty folder and it writes a working site, and the code is usually fine. Plain markup. A stylesheet that keeps one spacing scale for a whole file. It'll hand you eight versions of a hero section before you've opened a design file, and draw the illustration on the way past.
People used to get paid for that. Landing page templates. Brochure sites. The hundredth pricing table. The pass where somebody turns an approved design into real markup. Most of it was junior work, and a lot of it is gone.
Anyone calling this pure augmentation hasn't priced a five page marketing site lately.
Nobody is hiring juniors
Two studies are worth knowing. Neither says what the headlines say.
Xiang Hui and Oren Reshef looked at Upwork after ChatGPT and the image generators arrived. Freelancers in the exposed categories lost about 2 percent of their monthly contracts and about 5 percent of their earnings. Image work took roughly the same hit as writing. Real, and smaller than the noise around it. It's also one platform, and the authors say plainly that what they measured was small, well defined gigs. That is the easiest work in the economy to hand to a model.
The Stanford number is the one to sit with. Brynjolfsson, Chandar and Chen read ADP payroll data and found no widespread displacement. One group is the exception. Take 22 to 25 year olds in the jobs most exposed to AI, and compare them with 22 to 25 year olds in jobs it barely touches. The exposed group sits about 19 percent below where it would be if it had kept pace. A year earlier that gap was 15 percent. Not layoffs. Companies stopped hiring them.
So the honest answer isn't "your job is safe". The job is about as safe as it was. The way into it is closing. A profession that stops hiring juniors runs out of seniors later, and nobody has fixed that.
Which version to ship isn't in the prompt
Now the part the fear gets wrong.
Ask for a clean pricing section and you get one. Ask which plan should carry the recommended badge and you're on your own. Same for whether the page should open on pricing at all. Same for the client logo row, where one logo is a wide wordmark and the other six are square marks, so the row looks wrong on this page and nowhere else.
That isn't taste in the vague sense. It's knowing what the page is for and who lands on it. None of that is in the prompt. It's in your analytics, and in whatever you know about the people who arrive.
A model can write a hundred versions. It can't tell you which one to ship.
The second problem is duller and costs more. A model can't see the result. It knows the code it wrote, not what your browser did with it. It has no idea the sticky header now covers the section when someone follows a deep link into it, or that the results count shoves the filter bar sideways when the number goes from one digit to four. Somebody opens the page and spots that. That somebody is a designer.
Developers misread their own speed
Worth knowing before you plan a career around anybody's productivity claim.
METR ran a randomized trial with experienced open source developers, on code they knew well. With AI tools allowed, they took 19 percent longer. Afterwards they said the tools had made them 20 percent faster.
Sixteen developers, 246 tasks. Small, and it was early 2025 tooling on mature repositories, which is the hard case. It isn't proof the tools don't work.
Read it for the other thing. The people using them couldn't tell which way the effect ran, and that includes the ones who liked them. Every confident figure about how many designers a model replaces was taken with the same instrument.
Sentences are bad at position
Making the thing is cheap now. Deciding what it should be is still yours. What sits between the two is telling somebody else exactly what to change, and that is now most of the job.
You do that in sentences, and sentences are bad at position. "A bit more air above the section title, but not on mobile" leaves four decisions to the agent. Each wrong one comes back as another screenshot.
Pixy is for that half. Add the script tag, or install the Chrome extension and skip even that. Press ⌘⇧E and the page in front of you becomes editable. Your stylesheet, your content, your logged-in state. Pull the frame in to 390 and fix the phone layout at the width where it's actually broken, with your real CSS reflowing through its own breakpoints while you work.
Each change goes over as what it was: the element, the property, the value before and after, and the width you were sitting at. Spacing that belongs to a rule lands on the rule, so every match updates instead of one card drifting away from the others. Whatever you tried and dropped stays in your browser, because only the changes you keep reach your agent. Design mode is off until you turn it on, and turning it off puts the page back exactly as it was.
What the job is now
Will AI replace web designers? No. It has already replaced a lot of web design production, which is a different sentence and a colder one.
What's left is the part that was always hard. Work out what the page is for. Open the real thing and find out whether it does that. Say exactly what to change. All three are worth more now that making things is cheap.
Whatever you were about to explain in three sentences, open that page from your projects and press ⌘⇧E. Three sentences take longer than the change does.
Enough theory. Try designing with AI.
Use Pixy to make a design change on this page yourself.