Can customers tell when a website or app was designed by AI?

Last updated 2026-10-06

TL;DR

Often, yes. AI-made websites and apps share recognizable defaults, such as purple gradients, pulsing badges, glass cards, stock fonts and hype taglines, and designers now spot them quickly. The bigger trust risk is invented facts in AI-written copy. Let AI draft, and keep brand, design system and fact-checking with people you hold accountable.

Customers can often tell when a website or app was designed by AI, because tools such as Claude, ChatGPT, and the site builders that run on them reach for the same defaults. A layout that looked polished in March looks familiar by September. Looking familiar is a small cost. Publishing claims nobody checked is the large one, and it is the part an owner can control.

1. Gradients everywhereOften purple, on buttons and backgrounds
2. Rainbow colorNo dominant color, no accent discipline
3. Pulsing badgesAn animated "active" dot that is always active
4. Rounded cardsThe same soft card on every section
5. EmojiDecorating headings and buttons
6. MisalignmentIcons and dots slightly off their lines
7. Stock fontsInter, JetBrains Mono, and "//" separators
8. Leaked contextBuild details the customer never needed
9. GlassmorphismFrosted glass panels as the only style
10. Hype taglines"Elevate your workflow" and its cousins

Can customers actually tell a design came from AI?

Many customers can tell a design came from AI, even if they cannot name why. The September 2026 essay "Tells of a Slop UI" lists ten repeated defaults, including purple gradients, pulsing badges, frosted glass panels, and hype taglines. Designers recognize them at a glance, and ordinary visitors register them as a site that feels generic or slightly off.

The essay drew more than 230 comments on Hacker News. One commenter described using Claude to design a screen for an Unreal Engine game and being impressed at first. Less than a year later the same developer was embarrassed by how "Claude" it looked: the blinking dot in the corner, the double slashes in headers, the font choices. Now those patterns show up everywhere, and the commenter assumes the audience notices too. Another commenter, a musician, made the point owners should hear. A trained ear names the fault. A casual listener just says it does not sound right.

Does an AI-looking design cost you customer trust?

An AI-looking design costs some trust, because people judge credibility by appearance first. Stanford's Persuasive Technology Lab asked 2,684 people to rate real websites, and "design look" appeared in 46.1% of their comments on credibility, more than any other factor. A site that looks mass-produced invites the question of what else was mass-produced.

The Stanford study is from 2002, but the mechanism has not changed: visitors decide in seconds, before reading a word. Sameness alone is not fatal. A commenter in the same thread compared the current wave to the early years of Bootstrap, when thousands of sites looked identical and usability still improved overall. Customers forgave Bootstrap sites that worked. They will forgive an AI-assisted layout that works and sounds like you. They are less forgiving when the page also gets facts wrong.

Why are AI-written claims riskier than AI layouts?

AI-written claims are riskier than AI layouts because a model fills gaps with plausible inventions. In a September 2026 benchmark of 16 models, extraction runs invented a value for 70.7% of fields that did not exist on the page. Adding the sentence "Do not guess" cut that to 20.2%, which still leaves one invention in five.

The benchmark counted 405 invented values out of 573 missing fields without the instruction and 116 out of 574 with it. Its authors note it was one run on synthetic pages, so treat the numbers as a direction. The direction is clear enough for an owner. An AI-written "About us" can carry a founding year, a client count, or a certification nobody gave it. On Hacker News, practitioners reported that prompt rules like "always ground your responses" help but never close the gap.

What should AI draft and what should people own?

AI should draft layouts, first-pass copy, image variations, and code, where speed matters and mistakes are cheap to spot. People should own the brand, the design system, every factual claim, pricing, legal text, and anything a customer relies on. Accountability decides the split: a model cannot answer for a claim, so someone on your team must.

Glyph, a software developer who writes at blog.glyph.im, argued in a post discussed on Hacker News this week that a serious AI product would make checking for mistakes a first-class feature rather than fine print. Your process can do that today. The commenter in the slop thread who built a design system first, then told the AI it could only assemble approved components, reported far fewer stray styles. That is the pattern: people set the rules, AI works inside them.

How do you review a vendor's AI-built deliverable?

Review a vendor's AI-built deliverable by checking what AI gets wrong, not whether AI was used. Ask the vendor to show the design system behind the screens, to list every factual claim with its source, to confirm the copy contains no invented numbers, and to name the person who checked it. Then check a sample yourself.

1. Count the tellsHold the ten defaults above against every screen
2. Ask for the design systemColors, fonts, and components, written down and yours
3. Get a claims listEvery number, name, date, and promise, with its source
4. Spot-check five claimsPick them yourself; one invention means check them all
5. Read it aloudCut any sentence that sounds like a tagline, not like you

Put the claims check in the contract, not just the kickoff call. A vendor who used AI well will hand over a design system and a sourced claims list without complaint. A vendor who cannot has probably shipped the model's first draft. If the deliverable is large enough that you want someone else to run this review, an independent second opinion costs less than rebuilding a site customers have already judged.

Related guides

• What are the red flags in a software development proposal?
• Where does AI fit in my product?
• When is an AI API wrapper enough, and when do you need more?
• The Second Opinion: an independent review of a technology decision
• 10 Tells of a Slop UI
• Glyph: What would a serious AI product look like?

Key takeaways

• AI design tools share visible defaults, and designers and many customers now recognize them on sight.
• Visitors judge credibility by design first, so a generic look invites doubt about everything else on the page.
• Invented facts in AI-written copy are the larger trust risk, and prompt instructions reduce them without removing them.
• Let AI draft layouts, copy, and code; keep brand, design system, and every factual claim with named people.
• Review a vendor's AI-built work with a design system, a sourced claims list, and your own spot checks.

Frequently asked questions

Is it bad to use AI to design my website?

No. AI is a fast way to produce layouts, drafts, and code. The problems come from shipping its defaults unchanged and publishing its claims unchecked. Give the tool your brand colors, fonts, and components, have a person edit every screen, and verify every number and promise before the site goes live.

What are the most common signs a website was made with AI?

Purple gradients, pulsing status badges, identical rounded cards, frosted glass panels, emoji in headings, stock fonts such as Inter, slight misalignments, build details in customer copy, and taglines with words like elevate or effortless. One or two mean little. Several together suggest nobody reviewed what the tool produced.

Does telling ChatGPT or Claude not to guess stop made-up facts?

It helps but does not stop them. In a September 2026 benchmark of 16 models, adding "Do not guess" cut invented values from 70.7% to 20.2% of missing fields. One in five remained. Every factual claim still needs a person to check it against a real source before publishing.

Should I ask my agency whether they used AI?

Ask how they used it and how they checked it. Most agencies now use AI tools, and banning them raises costs. Require a written design system, a list of factual claims with sources, and the name of the person who reviewed the copy. Those answers matter more than whether AI touched the work.

Back to Blog