How to spot an AI-built website: the tells of Claude Code and ChatGPT Codex

You recognize an AI-built website first by its defaults: the choices the tool makes when nobody told it what to do. In 2025, that meant a purple gradient on a white background, the Inter typeface and three rounded cards under a centred headline. In 2026, Anthropic itself flags new clusters: a large serif headline and a terracotta accent, or small all-caps labels above every heading. ChatGPT Codex, for its part, is mostly associated with cards, pills and frosted-glass effects. These clues suggest that AI was used, but they prove nothing. The only reliable proof lives in the code history, where Claude Code and ChatGPT Codex sign their changes by default. For a business, the point is not to hide AI: it is to make sure every visible choice on its site is a decision. The surest way is to have a human design it: AI then builds what a person decided.

On this page

Why AI-built websites look alike

Claude Code (Anthropic's coding tool) and ChatGPT Codex (OpenAI's coding agent) write a website's code from instructions in plain language. Our article on what AI changes in website development describes what they do and what they change for a business. This one looks at a more visible consequence: left to themselves, they often draw the same website.

Both vendors acknowledge it. In November 2025, Anthropic's applied AI team published a post on the design of interfaces Claude produces. It explains that safe choices (the ones that work everywhere and offend no one) dominate the web pages used to train models. Without instructions, a landing page generated by Claude therefore almost always used the Inter typeface, a purple gradient on white and minimal animation. Anthropic calls this "distributional convergence": the model returns to the centre of what it has seen most often.

OpenAI says the same thing in its design guide for GPT-5.4, published in March 2026: when instructions are vague, models fall back on the most frequent patterns in their training data. The same guide bans default typefaces (Inter, Roboto, Arial), purple on white and cards used for no reason.

There is some irony in this story. In August 2025, OpenAI's official guide to coding with GPT-5 itself recommended a specific stack for new web apps: Next.js, Tailwind CSS, the shadcn/ui component library, Lucide icons and typefaces such as Inter, Geist or IBM Plex Sans. Millions of projects followed that advice, which in turn became a default. The same month, Adam Wathan, the creator of Tailwind CSS, jokingly apologized on X: by making every button in Tailwind UI indigo five years earlier, he had supposedly made every AI-generated interface indigo. The quip is not a study, but it sums up the mechanism well.

The first tells: purple, Inter and three cards

The first generation of tells, from 2024 and 2025, is the best documented. It appears in Anthropic's instructions, in OpenAI's guide and in lists published by designers. You can spot it by a few traits:

• A purple or indigo gradient, often on white, sometimes as gradient text in the headline.
• Inter everywhere, or another neutral typeface such as Roboto or Arial.
• A big centred headline with two buttons below it, then a row of client logos.
• Three identical cards with an icon on top, a short title and two lines of text.
• Emojis instead of icons, in the navigation or in front of every list item.
• Glow effects on dark backgrounds and cards nested inside other cards.

These traits have been measured. In April 2026, developer Adrian Krebs analyzed 1,590 project pages submitted to Hacker News, a forum popular with programmers. His program checked for about fifteen patterns: Inter in a centred headline, purple, a badge above the main heading, cards with a coloured border on one side, numbered steps, rows of statistics, frosted glass. The result: 22% of pages showed at least four, 32% showed two or three and 46% showed one at most. Krebs puts the false-positive rate at 5 to 10%.

The public notices too. An analysis of 3.2 million Reddit posts, published in 2026 by a developer, found that comments like "they all look the same" or "screams AI" each appear in about 13% of discussions on the topic. It measures perception, not websites, but it shows your visitors could make the connection.

Design tells associated with Claude Code in 2026

Purple faded as vendors banned it. Other defaults took its place. The best source on this is Anthropic itself: its design skill (a file of instructions Claude Code loads when it designs an interface) was published on November 12, 2025, rewritten on June 9, 2026, then again on September 3, 2026. Each version names the defaults of the moment to help Claude avoid them.

In its September 2026 version, Anthropic writes that AI-generated design clusters around five families:

The same document notes other habits: a single word highlighted in the headline (in italics or colour), 01, 02, 03 markers on content that is not a sequence of steps, a big number with a small label to open the page and every section fading in on scroll. Anthropic writes that these animations, like a hover effect on every card, read as AI-generated.

These families do not target Claude alone: Anthropic is talking about AI-generated design in general. But its instructions exist to correct Claude. The first family carries the colour of its own brand. On Hacker News, a developer summed up in July 2026 what several had noticed: orange accents and serif typefaces now give the impression of an AI-built site. In April 2026, others noted that Claude had moved from purple to retro, terminal-inspired designs. These are user observations, not measurements, but they point the same way as Anthropic's document.

Design tells associated with ChatGPT Codex

For ChatGPT Codex, the most cited source is Uncodixfy, a set of rules published on GitHub in March 2026 and followed by thousands of developers. Its author writes that a UI made by GPT is recognizable immediately. Its list of Codex tells:

• panels floating over the background, with heavy shadows;
• very rounded corners and pills everywhere;
• frosted glass, blur and glow effects;
• grids of statistic cards, even on a page that does not need them;
• small coloured dots in front of labels, with no real status behind them;
• muted grey, all-caps, widely spaced labels;
• buttons and cards that lift on hover.

OpenAI confirms most of this portrait through its own bans. Its March 2026 guide says to use no cards by default and rules out cards in the opening banner, stat strips, logo clouds, piles of pills and floating dashboards. The design skill published for Codex the same day also rejected the generic software card grid as a first impression. OpenAI removed it from its repository on April 23, 2026.

Head-to-head tests paint a more nuanced picture. In October 2026, How-To Geek gave the same portfolio-page prompt to both tools. The reviewer preferred Claude Code's design but found its copy generic. In the same test, Codex produced a generic corporate website, with a stats card, but better copy. Other tests published in 2026 by XDA, designproject.io and Bito reach different verdicts, with different model versions. One finding keeps coming back, though: when the prompt specifies almost every design decision, both tools produce very similar results.

Claude Code and ChatGPT Codex side by side

This table sums up what vendors and observers attributed to each tool in October 2026. These are tendencies observed without design instructions, not rules.

The column that matters most is the last one. When a vendor bans a trait in its own instructions, it acknowledges that its model was producing it. The other attributions rest on developer observation and can change with every new version.

Tells in the code

Your visitors do not see the code. Your next provider will read it. Code tells therefore matter for maintaining the site and for the day someone else has to take it over.

The ones Anthropic documents in Claude

Anthropic's prompting guide openly describes a tendency to overdo it. In October 2026, it said that Claude Opus 4.5 and Opus 4.6 tended to overengineer: creating extra files, adding unnecessary abstractions or building in flexibility nobody asked for. The sample instruction it offers to correct course also targets comments added to unchanged code, error handling for cases that cannot happen and temporary files left in the project.

The same guide warns about a more serious habit: focusing on making tests pass rather than on a real solution. Anthropic described it as early as February 2025 in the Claude 3.7 Sonnet system card. In tools such as Claude Code, the model sometimes special-cased a test to return the expected value directly, or modified the test itself. It often left a comment announcing it. Anthropic attributed this behaviour to its training. In May 2025, the Claude 4 system card measured an average 67% decrease in this behaviour in Opus 4 compared with Sonnet 3.7, without making it disappear.

Two more harmless tells come up in user reports on Claude Code's public repository. The first: "You're absolutely right!", served up at every turn, which was the subject of a widely discussed issue in July 2025. The second: emojis in documentation headings, commit messages and even code, sometimes despite an instruction that bans them.

The ones OpenAI documents in ChatGPT Codex

OpenAI describes its models differently: through the instructions it gives them. The system prompt recommended in its ChatGPT Codex guide, in October 2026, forbids wrapping code in broad blocks that catch every error and hiding a failure behind a fallback value that looks like success. It asks for rare comments and plain characters by default. An instruction is not an admission, but nobody bans what a model never does.

The GPT-5 guide, in August 2025, also reported the experience of the code editor Cursor: GPT-5 wrote long reports on its work, while its code was sometimes hard to read because of its terseness, with many single-letter variable names. OpenAI's sample prompt told it to remove the comments it had added.

For a business, the tell that matters most is the one about tests. Code that works around a test gives the impression that everything works. That is why a person must review every change and why a site's automated checks are not enough on their own.

The signature in the code history

A site's style proves nothing. Its history, however, often contains a signature. Versioning keeps every change with its date, its description and its authors. Both tools leave a trace there by default.

One detail is often enough to tell them apart: Claude Code writes "Co-Authored-By" with capitals, while Codex's code writes "Co-authored-by" in lower case. Claude Code's documentation says the line carries the name of the model used. Codex's source code, published on GitHub, contains the instruction telling the model to add its line to every change.

The instruction file is another clue, though a fading one. AGENTS.md is an open format launched by OpenAI and handed to the Linux Foundation in December 2025. The foundation said at the time that more than 60,000 open-source projects had adopted it. It is read by Codex, but also by Cursor, Gemini CLI, GitHub Copilot and several other agents. Claude Code also reads it when a project has no CLAUDE.md. An AGENTS.md alone therefore does not point to Codex with certainty, while a CLAUDE.md almost always points to Claude Code.

These files and lines are not visible on the published site: they stay in the code repository. If your code belongs to you, you can look. Otherwise, ask your provider directly. That is also why the code repository should be in your name.

Tells in the website copy

The copy on an AI-written site has its own tells. They show faster than the design. The volunteers of English Wikipedia keep the most complete list, updated as models change. It includes:

• The em dash. In October 2026, Wikipedia had moved it to its historical indicators, valid from 2022 to September 2026, because recent models are learning to avoid it. The page cites a July 2026 analysis by The Economist: among recent models, only Claude used more em dashes than professional writers.
• Negative parallelism. "It's not a tool, it's a strategy." The construction turns up in every generated text.
• The rule of three. Three adjectives, three benefits, three examples, even when two or four would do.
• Overused words. Delve, crucial, pivotal, robust, showcase, tapestry. A study published in Science Advances in July 2025 used these excess words to estimate, across more than 15 million scientific abstracts, that at least 13.5% of 2024 abstracts had been processed with a language model.

French Wikipedia's help page adds clues of its own: em dashes, repeated bold, bullet lists everywhere, extra capitals in headings, question-shaped section titles, a final "Conclusion" section and a promotional tone. It also flags technical traces, such as a link ending with a "chatgpt.com" source tag.

These clues have a limit: human writers also use dashes and lists. Our note on Claude's watermark explains what the invisible marks added to generated text actually allow you to conclude.

Can you know for sure which tool built a site?

Not from the site alone. In the lab, researchers manage it on short programs: in June 2025, a team identified which of five models had written a C program with 95.40% accuracy (97.56% between two very similar OpenAI models). Another team exceeded 95% on JavaScript within the GPT family in October 2025. These results concern older models and raw code, not a complete site reworked by a team.

Three reasons make the diagnosis uncertain on a real site:

• The tells shift every few months. The purple of 2025 gave way to the serif headlines of 2026. Each new model version moves its defaults.
• The same design skills circulate between tools. Impeccable, a widely followed set of design instructions on GitHub, works with Claude Code as well as with Codex and Cursor.
• A human almost always edits. A changed typeface or a colour imposed by a brand guide is enough to blur the trail.

The safest diagnosis therefore remains the code history. When we examined a recent Quebec site built with Next.js in October 2026, we found a serif typeface popular with generated sites, about thirty all-caps spaced labels and numbered sections. These clues pointed to Claude. They proved nothing.

What these tells can cost your brand

A site with default design can work very well. The risk is commercial, not technical.

• Your site could look like your competitors' sites. If three companies in your industry use the same tool without direction, their sites can share the same palette, the same cards and the same layout. A visitor comparing them is left with price alone.
• Visitors may conclude the work was rushed. The "screams AI" comments found on Reddit show that part of the public makes the connection. For a business that sells its expertise, that impression can cost a quote request.
• Your brand gets diluted. A colour or typeface chosen by the tool replaces the one your brand guide called for. With each change, the site drifts away from your other materials.
• The code could cost more to maintain. Code that does too much or hides its errors takes more time for the next person who takes it over.

Design choices are also among the decisions that remain human. AI can produce ten versions of a home page in a few minutes. The choice between them tells your customers who you are.

The tells we fixed on our own site

Falia.co was built with Claude Code on the Astro framework. Some AI tells slipped in as the site was changed, and we fixed them:

• The em dash. In August 2026, we removed 208 of them from the site's visible text. The 345,000 words of our old site contained none: they all came from AI. Our writing rules have banned them since. Several automated checks look for them before every release.
• Pills. On September 21, 2026, we removed them: there were many and they made people think of AI. All our tags took the shape of our buttons, with 6-pixel corners. A check now blocks a release if a pill comes back: the day before, the AI had added a new row of them.
• The comma before "et". In French, the serial comma comes from English. The written rule did not last a day. A check has blocked it since September 8, 2026.
• Anglicisms in French. Phrases translated word for word from English have been blocked by a check since September 27, 2026.

The lesson fits in one sentence: a written instruction is not enough to hold back a tell. It takes an automated check that blocks the release. And our changes do carry Claude's "Co-Authored-By" line: we do not hide the tool, we decide what it produces.

Avoiding the tells: seven decisions

Whether your site is built by an agency, a freelancer or your own team, these seven decisions reduce default design. They take little technical skill and a lot of clarity.

Falia's site and our clients' sites follow this method: AI executes, a senior signs off. Our pillar article on AI and website development describes the full approach.

Frequently asked questions about AI-built websites