AI and your Swiss CV — what is allowed, what shows, and what you still have to write yourself

Using AI in an application — what applies

No rule forbids AI and no duty requires disclosing it. Nobody has been caught for using AI — people are caught by letters that could have been sent to any company. As of 7 September 2026.

What Swiss law actually says

There is no Swiss provision on how an application must be produced. No declaration is required, no application form in the live adverts asks for one, and no case law turns the use of a writing tool into a defect. The provision that does apply is about content. Art. 28 of the Code of Obligations, in the federal English version, reads: «A party induced to enter into a contract by the fraud of the other party is not bound by it even if his error is not fundamental.» An employer who discovers an invented degree can therefore walk away from the contract — and the same is true whether the invention was typed or generated.

The counterpart provision protects you. Art. 328b limits what an employer may do with your data: it «may handle data concerning the employee only to the extent that such data concern the employee's suitability for his job or are necessary for the performance of the employment contract», with the Data Protection Act applying in all other respects. That is the legal reason a Swiss CV can be short on private detail without looking evasive — and a useful thing to know when a generated template offers you fields for marital status and religion.

For the tools themselves, no special regime exists. In November 2023 the Swiss data protection commissioner published a statement that the federal Data Protection Act is «technology-neutral» and therefore directly applicable to AI-supported data processing, and that providers must make the purpose, the functioning and the data sources of such processing transparent. In plain terms: the same rules that govern any other service govern this one, and the obligation to tell you what happens to your input sits with the provider.

Where a model actually helps an applicant in Switzerland

The strongest case for using one has nothing to do with writing style. It is that most people applying here are applying in a language they did not grow up with. Of the live adverts, 41.6 per cent are posted in German, 20.2 per cent in English, 9.7 per cent in French and 1.6 per cent in Italian — so an English-speaking candidate in Zurich who only reads the English-language adverts is looking at a fifth of the market. A model that translates an advert accurately, and then translates your finished lines back, widens the field you can see. That is a real gain, and it costs nobody anything.

The second real gain is structure. A Swiss CV has an expected order and an expected length; the federal application-tips page sets the outer bound at three pages and asks for PDF. Handing a model your raw history and asking it to fit that shape is faster than doing it by hand, and it does not touch a single fact. Structure is exactly the kind of work a general-purpose model does well: bounded, conventional, verifiable at a glance.

The third is the version problem. Most applicants need several versions of the same CV, one per advert, differing in emphasis rather than in content. Producing those by hand is where people give up and send the same document twenty times. Producing them with help is where a model earns its keep — provided each version is checked against the advert it was made for.

What no model knows about the employer you are writing to

There is a persistent belief that a model «knows» the company. The structure of this market says otherwise. 45.4 per cent of the employers with a live advert have exactly one; the median employer has two; the ten largest posters together account for only 21.4 per cent of the adverts. The typical recipient of a Swiss application is a firm about which almost nothing has been written outside its own website.

The same granularity shows in how roles are described: 82.2 per cent of advert titles occur exactly once across the whole live set, and the single most frequent title accounts for 0.6 per cent. Ask a model for «a suitable job title» and it will pull toward the middle of everything it has read. The title that matters is sitting in the advert in front of you, and copying it is not gaming a filter — it is simply the more accurate statement.

So the division of labour writes itself. The general parts — order, length, grammar, translation, tightening — can be delegated. The specific parts cannot: the name on the advert, the location, the workload, the two requirements the advert lists first, and what you actually did about each of them. A draft produced without those inputs fits every employer, which is another way of saying it fits none.

The last pass, which is not optional

The federal application-tips page names three checks before you send: spelling and formatting errors, completeness of the documents, and correct, unambiguous file names. All three are exactly what a generated draft fails when nobody looks again — a model produces well-formed paragraphs, but it does not know whether your diploma is actually attached or what your file is called.

Spend on that pass the time you saved on the draft. The market gives you room: the median live advert on SwissJobs.app is nineteen days old when you see it, so ten extra minutes cost you nothing measurable. Read it aloud, hold it next to the advert, check the file names, and confirm every number in it against a document you could produce if asked.

The honest summary is not a legal one. A CV is not an essay; it is a set of checkable statements. If you have the statements, using help to phrase them is unremarkable — Swiss employers use the same tools every day. If you do not have them, no tool will hold the gap together past the first interview.

Why no tool reliably marks a CV as machine-written

The clearest statement comes from education, where the same question has needed a practical answer for three years. The ZHAW, in its teaching-and-learning guidance, records that AI detection software is explicitly forbidden at the institution because it does not deliver reliable results. That is not a review of individual vendors; it is an institutional rule forbidding its own teaching staff from using tools whose verdicts it cannot stand behind.

The Université de Neuchâtel gives the reason in terms of the error it produces. Asked whether generated text can be identified, its teaching-support service answers flatly: not reliably. The tools, it writes, are «peu fiables et susceptibles de générer des faux positifs, accusant à tort les étudiant-e-s» — unreliable and prone to false positives that wrongly accuse people — and it therefore advises teaching staff not to trust them. The damage sits with the falsely accused rather than the correctly caught, and that burden falls hardest on people writing in a language that is not their first.

For a CV there is a second reason that has nothing to do with the quality of any detector. A CV is mostly structured data rendered as text. The proportion of genuinely free composition is small, and detection methods need volume before their signal rises above noise. There is also a historical confusion at work: the flat, verb-first register people now read as machine output is the register that application guides have taught for a generation. Recognising it proves that someone followed advice, not that a model was involved.

What Swiss hiring teams actually notice

Answering the question usefully means moving from style to local knowledge. The Swiss market is linguistically mixed but not arbitrary: among adverts whose language is unambiguous, 56.9 per cent are German, 27.6 per cent English, 13.3 per cent French and 2.2 per cent Italian. An English CV is therefore nothing unusual here — more than a quarter of those adverts are themselves in English. What stands out is something narrower: a document that reads as though it were written for a market in which every place name gets translated.

That is where the measurable differences sit. Across all live adverts the city is written «Zürich» 88 per cent of the time and «Zurich» 12 per cent; for Geneva it is 92.8 per cent «Genève» against 7.2 per cent «Geneva». Bern, Basel and Neuchâtel appear in their local form almost without exception — «Berne» shows up in a single advert, «Basle» and the German «Genf» in none. These spellings are a convention rather than a preference, which is what makes them a signal of belonging, and an internationally trained model breaks them consistently.

The reverse case is just as instructive, because it kills a rule of thumb that circulates widely among applicants moving to Switzerland. The tag m/w/d is often said to identify a German employer. In the recorded data it does not: 43.8 per cent of German-language advert titles carry a gender tag at all, and 71.5 per cent of those use m/w/d. Someone who «corrects» that form in a covering letter is correcting away from the market. The difference between a useful adjustment and a harmful one is whether anyone counted.

Who reads your CV first, and how much time they have

The fear that a system discards a CV before any person sees it applies, in the Swiss market, to a minority of vacancies. 27.2 per cent of live adverts lead into a recognised hiring system — the largest are Workday at 7.9 per cent and Prospective at 7.1 per cent, followed by Greenhouse at 2.7 per cent, SmartRecruiters at 1.9 per cent, Ashby and Umantis at 1.3 per cent each and Lever at 1.2 per cent. The remaining near three-quarters go to the employer's own system or to a small, often domestic one.

Those destinations are extraordinarily fragmented: 45.1 per cent of them are reached by exactly one advert, and the ten largest together take only 32 per cent of adverts. There is no single filter to optimise against. There is a very long tail of small recipients where a person opens the dossier at the end of it. Worth noting too is who most often sits in between: the largest single poster in the index is not an employer but an intermediary, accounting for 10.6 per cent of live adverts. In those cases a consultant reads first and will walk through your entries on the phone regardless.

The pace argues against the snap-judgment story as well. The median live advert is 19 days old; 62.2 per cent are at most 30 days old, but only 13.8 per cent are a week old or less. Vacancies stay open for weeks and dossiers are read across those weeks. That is the substantive answer to the question: whoever reads your CV has time to compare it with the advert, with your reference letters and later with what you say in the room — and it is those comparisons that expose something, not a hunch about style.

When the other side automates the decision: what you may demand

The question has a reverse side that is rarely discussed. Applicants ask whether their use of AI will be noticed; Swiss data protection law regulates the opposite case, the employer's use of automation. Art. 21 of the Data Protection Act requires the controller to inform the person concerned about «any decision that is based exclusively on automated processing and that has a legal consequence for or a considerable adverse effect on the data subject», to let them state their point of view on request, and it adds: «The data subject may request that the automated individual decision be reviewed by a natural person.»

The right of access completes it. Under Art. 25 you are entitled in every case to the personal data being processed as such, the purpose, the retention period and, «if applicable, whether an automated individual decision has been taken and the logic behind the decision». The Act states that no one may waive that right in advance, that the information must be provided free of charge and that it is as a rule supplied within 30 days. Whether a given rejection really was exclusively automated is a question of fact — the access request is how you settle it. Note that the English text on fedlex is a translation published for information only and has no legal force; the German, French and Italian versions govern.

Put together, the picture is undramatic. On one side there is no dependable tool that could flag your CV as generated. On the other there is a process in which factual contradictions surface over weeks, and in which an automated rejection owes you both an explanation and a human review. The practical consequence follows: the draft may come from a model, but the place names, the job title, the workload percentages and the numbers have to be yours and have to match your certificates. Once they do, whether anyone «can tell» stops having consequences.