Corrext and generative AI. Useful for understanding, but binding documents need a framework
ChatGPT, Azure OpenAI GPT and Claude rephrase, summarise and explain a legal passage better than a conventional translation engine. This page sets out what they do well, and what they cannot prove on their own on a binding Swiss document.
Request a demoThe short answer
Yes, a general-purpose LLM translates and rephrases a legal passage very well.
What it does not provide on its own: sourced Swiss terminology, a guaranteed legal reference, a review interface, a quote, and data processing in Switzerland. In Corrext, Azure OpenAI GPT and Claude remain available, under your organisation's governance.
- No sourced Swiss terminology
A term is rendered plausibly, with nothing to show how the Federal Gazette or fedlex.admin.ch actually renders it. In Corrext, that is the job of Fast lookup CHnell.
- No guarantee on legal references
An article, an abbreviation or a court can be produced with full confidence and still be wrong, with no warning sign in the answer.
- No review interface
No validation chain: no review level, no legal translator to order, no certification possible. In Corrext, review is ordered in Translation projects.
What general-purpose LLMs do remarkably well
They are the best tools legal experts have ever had for working on the meaning of a text. Saying so does not weaken our argument: it is the honest starting point for the comparison.
- Rephrasing and simplifying
Turning a dense clause into a readable paragraph for a client who is not a legal expert, or the reverse, in the register requested.
- Summarising and explaining
Bringing out the structure of a 50-page written submission, explaining a contractual mechanism, isolating each party's obligations.
- Handling ambiguity
Offering several readings of an ambiguous wording and spelling out what separates them, where a conventional translation engine silently picks one.
- Exploring drafts
Quickly producing several versions of a clause to feed an internal discussion, before settling on the version that will be negotiated.
Twelve criteria, based on real work on a case file
What we say about general-purpose LLMs is limited to what is public and stable. They are generic models, remarkable with language. They have no sourced corpus of Swiss law and no integrated review or certification chain, and their consumer plans process documents outside Switzerland. In Corrext, Azure OpenAI GPT and Claude are configured as third-party engines hosted abroad.
| Generative AI alone | Corrextwith access to Azure OpenAI GPT and Claude | |
|---|---|---|
| Legal quality | ||
| Rephrasing, summarising, explaining a passage | Excellent | Rephrasing in five styles, including Formal legal and Financial |
| Swiss legal terminology in context | No source shown | CHnell: aligned segments, 30 domains, official sources |
| Verifiable Swiss legal references | Plausible, never sourced | One-click check with Fast Lookup, with ‘Show in context’ |
| Comparing several engines on the same sentence | One model only | Up to four alternatives, each labelled with its engine |
| Confidentiality and governance | ||
| Where documents are stored | Outside Switzerland | Swiss servers, in every case |
| Where the text is processed | Outside Switzerland | Exclusively in Switzerland with LexMachina or in Highly sensitive mode |
| Warning before a document is sent | None | Attorney-Client privilege notice shown when the engine is chosen |
| Governance across the organisation | ½Depends on the enterprise plan | Engine policy applied in the tool, sensitive mode, ISO 27001 |
| Review and delivery | ||
| Human review ordered in the tool | No | Light review, Full review, Double review, internal or external |
| Quote and delivery time before you start | No | Calculated from the standard pages and the detected legal category |
| Certified, notarised or apostilled translation | No | Double review & certified translation; translation of commercial register extracts, up to the apostille |
| Batches of files in their original format | ½Depends on the plan, one file at a time | Up to 25 files of 50 MB, format preserved, plus PDF to Word |
Available and observed½Partial or depends on the planNot available
Comparison made on the date this page went online, from the publishers' public information and our own observations in Corrext. The products compared change over time: only their official documentation is authoritative.
An example of an error, and the source that settles it
A banking law sentence, translated from French into Italian. ChatGPT keeps the French abbreviation of the Banking Act; LexMachina uses the one found in federal texts, and CHnell shows its source.
Source · French
L'établissement assujetti veille au respect du secret bancaire au sens de l'art. 47 LB. En cas de faillite de la banque, la liquidation suit les règles de la LP, sous réserve des dispositions de la LB.
ChatGPT
L’istituto assoggettato vigila sul rispetto del segreto bancario ai sensi dell’art. 47 LB. In caso di fallimento della banca, la liquidazione è disciplinata dalle norme della LEF, fatte salve le disposizioni della LB.
LexMachina
L’istituto assoggettato vigila sul rispetto del segreto bancario ai sensi dell’articolo 47 LBCR. In caso di fallimento della banca, la liquidazione è retta dalle disposizioni della LEF, fatte salve le disposizioni della LBCR.
The proof, in CHnell
powered by Neur.onIl rimando alla LBCR non comprende le disposizioni in merito a depositi privilegiati (art. 37 a- 37 c LBCR), garanzia dei depositi (art. 37 h -37 k LBCR) e averi non rivendicati (art. 37 l LBCR).
Le renvoi à la LB ne concerne pas les dispositions sur les dépôts privilégiés (art. 37 a à 37 c LB), la garantie des dépôts (art. 37 h à 37 k LB) et les avoirs en déshérence (art. 37 l LB).
GS-EFD FF20147235 Message-concernant-la-loi-sur-linfrastructure-des-marchés-financiers- - Source: fedlex.admin.ch - Data Set: Neur.onIn Swiss Italian, the Banking Act is abbreviated as LBCR: ChatGPT keeps the French abbreviation, LexMachina uses the one found in federal texts, and CHnell shows its source.
Compare the engines on your own texts in the demo- Real outputs recorded on 18 October 2026: ChatGPT on chatgpt.com, LexMachina in Corrext.
- In CHnell, 905 official segments from the Federal Gazette and fedlex.admin.ch align LBCR with LB.
- A one-off, dated example: judge for yourself on your own texts.
Should a law firm ban ChatGPT?
Banning them achieves nothing except invisible use: a team member in a hurry opens a tab, pastes an extract, and nobody ever knows. Corrext approaches the problem differently. Azure OpenAI GPT and Claude are offered internally, alongside LexMachina and DeepL Pro. Choosing them displays the Attorney-Client privilege notice, which restates your organisation's policy: full pseudonymisation, high-profile or price-sensitive matters excluded. Your teams keep the power of these models, and you keep the record and the rule.
HomeAMYou have chosen to use third-party translation engines hosted abroad. Before uploading your file(s), your organization requires you to remove any references that could potentially disclose the identity of clients, any other related parties, or the subject matter at hand. Third-party translation engines hosted abroad may exclusively be used for documents that have been completely pseudonymized. The use of such engines for high-profile or price-sensitive matters is strictly prohibited.
Le débiteur en demeure doit des intérêts moratoires au taux de 5% l'an, conformément à l'art. 104 al. 1 CO, sans qu'une mise en demeure formelle soit nécessaire lorsqu'un terme a été convenu.
Fast LookupThe debtor in default shall owe default interest at the rate of 5% per annum in accordance with CO Art. 104 para. 1, without the need for formal notice when a term has been agreed.
The engine is chosen each time a text is sent, without leaving the platform.
A reminder of your organisation's policy: pseudonymisation, sensitive matters excluded.
LexMachina or Highly sensitive content: everything stays in Switzerland, in DE, FR, IT and EN.
Frequently asked questions about general-purpose LLMs and Swiss law
Can I use ChatGPT to translate a contract?
To understand a clause, grasp its scope or prepare a discussion, yes: a general-purpose LLM does this very well. For a contract that will be signed, sent to a third party or filed with an authority, no, not on its own. It provides no sourced Swiss terminology, no guarantee on legal references, no review by a legal translator and no data processing in Switzerland. In Corrext, these models remain available, but the binding version goes through Light, Full or Double review.
How do I check a Swiss legal reference?
Do not ask the engine that has just produced it: ask a source. In Corrext, highlight the term or abbreviation in the source text and click Fast Lookup. The CHnell concordancer displays aligned segments from official Swiss sources, with the domain, the source and a Show in context link. You then see how the term is actually rendered in the Federal Gazette or on fedlex.admin.ch, not how a model assumes it should be.
Does Corrext use general-purpose LLMs?
Yes. Azure OpenAI GPT and Claude are available in Corrext, alongside LexMachina and DeepL Pro, and you choose the engine each time you send a text. Because they are hosted abroad, selecting them triggers the Attorney-Client privilege notice, which restates your organisation's policy: full pseudonymisation, no sensitive matters. The Alternatives panel lets you compare their suggestions with the Swiss engine's on the same sentence.
Do general-purpose LLMs make up statutory provisions?
It can produce a reference, a court or a term that is plausible but wrong, and nothing in its answer flags the error. In the banking law sentence shown above, ChatGPT keeps the French abbreviation LB where the Italian versions of federal texts use LBCR. That is precisely why Corrext displays several alternatives, each labelled with its engine, and gives access to the concordancer.
Are my conversations with a consumer AI assistant confidential?
Terms vary from one provider and one plan to another, and they change. Two points remain constant for a Swiss law firm: the document is transferred outside Switzerland, and nothing in the interface reminds users what their internal policy requires before they send it. In Corrext, the Attorney-Client privilege notice appears when the engine is chosen, and Highly sensitive content mode requires storage and processing exclusively in Switzerland.
When is a general-purpose LLM the right tool?
When you are working on meaning rather than on the final wording: summarising a long document, explaining a contractual mechanism, rephrasing a paragraph for a client without legal training, exploring several drafts of a clause. Corrext also offers its own rewriting function, Rephrasing, in five styles including Formal legal, Financial and Simplified.