AI Disclosure

How LEXchat Law uses artificial intelligence

In short: LEXchat Law produces preliminary analyses fully automatically. You are interacting with an AI system. No lawyer on our side reviews them. Professional responsibility for anything that passes from an analysis into a client matter or case file remains with the Customer.

1. What LEXchat Law does

You enter the facts — LEXchat Law turns them into a structured preliminary analysis containing:

• a classification of the area of law and the legal bases of the claim
• the relevant statutory provisions (federal and cantonal, from Fedlex and the cantonal collections)
• potentially relevant court decisions (Federal Supreme Court and cantonal instances)
• an AI-generated first assessment including a possible line of argument
• references to the legal sources used, which you can and should look up

The analysis is preparatory work, not a legal product. It removes the research burden from the Customer, not the judgement.

Who this disclosure applies to. LEXchat Law is addressed exclusively to professional users: law firms, in-house legal departments, legal expenses insurers and comparable professional organisations (sec. 1.2 of the GTC). This disclosure applies to all of those Customers. Where it refers to the rules of the legal profession — art. 12 LFPA, professional secrecy under art. 321 SCC — it applies only to Customers that are a law firm; for all other Customers, their corresponding professional, supervisory and confidentiality duties apply (sec. 1.3 of the GTC).

1.1 Swiss law

LEXchat Law works on Swiss law — federal law and the cantonal enactments we have indexed. Foreign legal systems are not covered; where a set of facts touches them, the result is incomplete.

2. How it works

LEXchat Law uses a large language model, but not as a free-text generator: your statement of facts is first prepared, then a semantic and full-text search runs across our own corpus of statutory provisions and decisions. The sources found are embedded into the prompt, and only then does the model generate its assessment (retrieval-augmented generation).

The difference matters: the model answers on the basis of concrete legal sources held by us — not from its general training knowledge alone.

Pseudonymisation before search and analysis. Before your statement of facts goes to the search and to the language models, names and other direct identifiers are recognised and replaced by placeholders. The recognition runs on a language model computed in Switzerland; only the version you have reviewed and released in the interface is processed further. The mapping table by which the clear names reappear in the analysis remains on the Swiss platform and is not transmitted. That version is pseudonymised, not anonymised — it remains personal data within the meaning of data protection law (sec. 8.2 of the GTC) and does not replace your own check under sec. 7.3 of the GTC.

On verifiability. LEXchat Law is designed to link material legal statements to the references used. Even so, individual summarising, interpretative or automatically generated statements may appear without a clear reference or may be attributed to a reference incorrectly. All references and legal statements must therefore be verified (section 4).

The entire process is automated. No lawyer on our side reviews the analysis before you receive it.

2.1 Models used

For preparation, search and answer generation we use models from several providers; which model applies depends on how your access is configured. The corpus of laws and decisions being searched resides entirely on our own infrastructure. Which providers these are, in which countries they process data and on what basis is set out per provider in Annex B to the data processing agreement; sections 6 and 12 of the privacy policy describe it in addition.

2.2 Is our input used for training?

No — and this is how we ensure it. We engage a provider only on the basis of a contractual configuration that excludes the use of customer content for training or for improving general models. We document this per provider before productive use. Where a provider requires a separate opt-out, an enterprise agreement or a zero-data-retention configuration for that purpose, we activate the corresponding setting before productive use and review it regularly.

Processing is transient in principle. Depending on the product and configuration, individual providers reserve a temporary retention for abuse detection — as a rule up to 30 days. Where a provider's abuse detection flags a request as a suspected violation, retention may last longer: with one of the providers used, up to two years, and up to seven years for contentless security metrics. What each provider reserves is set out, with retention tier, maximum period and processing region, in Annex B of the data processing agreement; sections 6 and 6a of the privacy policy describe it in addition. Nor do we train models on your client-matter data.

3. What AI cannot do

Language models are statistical pattern matchers. They do not "understand" law the way a lawyer does; they compute which statements are likely to fit a query. That produces systematic limits which even the best source grounding does not fully remove:

3.1 Hallucinations

Models can invent content that sounds plausible and is wrong. Grounding the model in concrete legal sources (section 2) can reduce this risk compared with generation based on general model knowledge alone. It does not eliminate the risk, however, and can itself produce errors where the search is incomplete, a source is misattributed or a source is misinterpreted. In practice this concerns above all:

• statutory provisions — wrong numbering, inaccurate wording, in rare cases provisions that do not exist
• decisions — misreported facts, misattributed considerations, incorrect docket numbers
• deadlines — limitation, forfeiture and appeal periods
• doctrine — invented citations or misattributed positions
• citation stance — a decision may cite a norm critically or by way of distinction; judging whether a citation supports a proposition or does the opposite is error-prone

3.2 Currency

Our corpus is updated regularly. The date of the last corpus update and — where possible — the currency of individual sources are shown in the service. We do not warrant that all newly published or amended enactments and decisions are already included on the day of your query. In interpreting the sources the model also draws on general training knowledge, which has a cut-off date.

3.3 No appraisal of the individual case

Law is case-by-case work. A date, a contractual clause, a domicile abroad can reverse the outcome entirely. The AI may overlook such details or weigh them wrongly — particularly where they are only implicit in the facts as entered.

3.4 No procedural strategy

LEXchat Law delivers legal information, not tactics. Whether to sue, settle or wait depends on the opposing party, the instance, the state of the evidence and the objective of the mandate — factors the system does not know.

4. What this means for you as a Customer

The preliminary analysis is a starting point for legal work, neither a basis for decisions nor a work product to be passed unchecked to clients, insured persons, opposing parties, authorities or courts.

• Verify every reference you adopt against the official source (Fedlex, the cantonal collection, bger.ch). The references in the analysis exist precisely for this.
• Check the citation stance before relying on a decision as support.
• Expect statements without a reference. Where a legal statement carries no reference, or the attribution is unconvincing, treat it as an unsupported assertion.
• Professional duties are unaffected. Where the Customer is a law firm, the duties of care and loyalty under art. 12 LFPA, professional secrecy and responsibility towards clients continue to rest with it. For other Customers — in-house legal departments, legal expenses insurers and comparable professional organisations under sec. 1.2 of the GTC — the same applies to their corresponding professional, supervisory and confidentiality duties. Using an AI tool does not shift those duties.
• Observe the input limits. The Customer decides which client-matter and case data is entered; it is the controller for that data (see section 13 of the privacy policy). No real client-matter data may be entered in test and evaluation environments (sec. 8.6.7 of the GTC).

5. Responsibility

For the client-matter and case data you process as a Customer through LEXchat Law, you are the controller under data protection law; we act as processor on the basis of a corresponding agreement. Details on the processing, the providers involved and transfers abroad: privacy policy.

6. Regulatory framework

6.1 Transparency

You are interacting with an AI system, and the output is generated fully automatically. We point this out in the interface before the first analysis, not only here. LEXchat performs no human review (sec. 2 and sec. 5.4 of the GTC). This corresponds to the transparency obligation under art. 50(1) of Regulation (EU) 2024/1689; we implement it as a precaution, irrespective of whether the Regulation applies to LEXchat Law (sec. 6.2).

Marking of generated content (art. 50(2)). Every analysis is marked as machine-generated, both in the interface and in every export. Whether a machine-readable marking is additionally owed is assessed and documented per use case: the provision exempts systems that perform an assistive function for standard editing and do not substantially alter the input data provided by the user or its semantics. LEXchat Law assembles legal sources relating to a case description entered by the Customer and does not alter it; where the exemption does not apply in an individual case, we implement the marking. We provide the assessment to customers on request.

6.2 EU AI Act

The transparency requirements of Regulation (EU) 2024/1689 (EU AI Act) have applied since 2 August 2026; the obligation to ensure a sufficient level of AI literacy among the persons dealing with the system has applied since 2 February 2025. Regulation (EU) 2026/1744 of 8 July 2026 (OJ L, 2026/1744, 24.7.2026; in force since 27 July 2026, the “Digital Omnibus on AI”) postponed the deadlines for the obligations applying to high-risk systems: to 2 December 2027 for systems under art. 6(2) and Annex III, and to 2 August 2028 for systems under Annex I. It did not amend the transparency requirements of art. 50; those apply unchanged since 2 August 2026. In our assessment the EU AI Act does not currently apply to LEXchat Law: under sec. 1.2 of the GTC, only a party with its seat in Switzerland — the Swiss establishment of a company seated in the Union or the rest of the EEA is excluded — may be a Customer; LEXchat Law is neither placed on the market nor put into service in the Union, and use whose output is employed in the Union is not covered by the service description (art. 2 of the Regulation). This assessment rests on the restriction in sec. 1.2; it will be revisited before that restriction is lifted.

LEXchat reviews on an ongoing basis whether and to what extent the EU AI Act applies by reason of the offering, the users, the places of use or effects in the EU, and documents its own role (provider, deployer or downstream provider) for each use case. For ordinary legal research and preliminary analysis, classification as a high-risk system is in our assessment not readily made out. Use by or on behalf of courts and authorities in the exercise of public authority is not covered by the service description: Annex III no. 8(a) of the EU AI Act classifies assistance to judicial authorities in researching and interpreting facts and the law as a high-risk application. The exclusion therefore appears not only here but in sec. 1.2 of the GTC — the document that ranks above this statement under sec. 2.2 GTC. A legal department that advises exclusively its own organisation does not fall within it.

In addition we maintain a model and system inventory, technical documentation including a change log, a risk classification per use case, and a process for complaints and incidents. Customers receive the information they need for their own assessment on request.

6.3 Switzerland

Switzerland currently has no overarching AI legislation. A consultation draft is being prepared by the end of 2026; we are following the process and will adapt this disclosure as soon as the legal position changes.

v1.0 — 22 August 2026