AI and web development: where do we really stand?

We're often asked whether AI "writes websites" now. The honest answer, after running it in production on our own projects: it writes half of one, and that's the easy half.

Here's a specific, measured case study from this site.

The problem: three languages required

playdigital.ch exists in French, English and German. It's not a preference, it's a build constraint: our route registry refuses to compile if a page doesn't have all three versions. The rule protects SEO — no orphaned pages, no incomplete hreflang — but it turns every publication into three.

Practical result: writing an article means writing three. Editorial output simply stopped.

What we built

An automated translation pipeline connected to Claude Haiku, triggered by command and continuous integration. You write in French, the English and German variants appear.

The interesting part isn't the model. It's everything around it.

An exhaustive field policy

An article isn't a block of text: it's about thirty fields. Some translate (title, excerpt, body). Others must never be translated (translationKey, which links the three versions; publishedAt; author identifiers). Still others must be retargeted: a cover image path follows the slug of the target language.

We wrote a policy that classifies each field by its path in the document, not by its name. An unknown field doesn't pass: the pipeline stops rather than guess. That's the difference between a tool and a damage generator.

Explicit refusals

The pipeline refuses to produce if:

• the Markdown structure has drifted — a heading added, a list turned into a paragraph;
• mandatory SEO fields come back empty;
• the target variant has already been reviewed and approved by a human.

No generated variant enters published. They all arrive in review, and publication requires a review recorded with a name and date.

What the machine got right

Across thirteen articles translated into English and German, we compared each variant to its source:

• identical number of headings in all 26 cases
• identical number of paragraphs in all 26 cases
• length variance from −4% to −22%, which simply reflects the compactness of English and German compared to French

Structure, technical terminology, Swiss typographic conventions (ss and never ß): all passed through. No broken tags, no untranslated fragments.

What the machine got wrong

And this is where it becomes instructive. Human review found, among other things:

• "der Akteur" — literal translation of "l'acteur", which means nothing in German in this context;
• "dieses Gerät" for "ce dispositif" — the German word denotes a physical device;
• "Domänenname" instead of "Domainname", the term actually used;
• "die KI-Vorschauansichten von Google" — Google's AI Overviews, which Google doesn't translate into German;
• a sentence describing the location of the robots.txt file "at the end of the domain name", faithful to the original French and factually wrong in both languages.

None of these errors is automatically detectable. All are visible to someone who reads the language.

What we take from this

AI didn't eliminate the work, it displaced it. The bottleneck is no longer production, it's review. A professional translator produced and verified; today production is nearly free and verification remains entire.

Three practical consequences for a client project:

1. Budget for review, not translation. It's the line item that persists, and the only one that protects your credibility.
2. Demand verifiable safeguards. A pipeline that doesn't know how to refuse will one day silently produce false content.
3. Never let anything publish itself. For us, a generated variant carries "in review" status until a human signs off.

We continue to use this pipeline — it made multilingual publishing possible for us. But anyone selling you "AI writes your website" is selling you the easy half and hoping you don't notice the other.