Methods and protocols
Legal orientation — Switzerland and the European Union
Locate the Swiss and EU questions to check for your intended use.
Build my routeSnapshot verified on 20 August 2026; the Switzerland section was rechecked on 18 November 2026 and the Council of Europe status on 19 November 2026. This page helps frame the questions to ask; it is not legal advice. Verify the current text, sector, jurisdiction, and the organization’s exact role before making a decision.
Switzerland
Switzerland does not yet have a general law dedicated exclusively to AI. The technology-neutral Federal Act on Data Protection (FADP) nevertheless applies directly to AI-assisted processing of personal data.
Control points:
- purpose, proportionality, transparency, and individual control;
- sensitive data, profiling, and automated individual decisions;
- data protection by design and by default;
- security, processors, processing locations, and international transfers;
- information about interaction with a machine and data reuse where required;
- a data protection impact assessment when processing is likely to result in a high risk.
The controller remains accountable when processing is entrusted to a provider. In particular, it must govern instructions, confidentiality, security, subprocessors, and transfers.
Operational routing by use pattern:
- for direct language-model interaction, record how people learn that they are corresponding with a machine and whether their inputs are reused;
- for an automated individual decision with legal or similarly significant effects, verify Article 21 FADP notice, the opportunity to state a point of view, review by a natural person, and the exact exceptions;
- for synthetic faces, images, or voices of identifiable people, record clear identification, consent or other authority, privacy controls, and applicable personality, criminal, and intellectual-property rules;
- for processing likely to create a high risk, complete and retain the data-protection impact assessment before the pilot.
Swiss calendar rechecked on 18 November 2026. The Federal Council press release of 28 October 2026 (French) states that the consultation on the legally binding measures is planned for spring 2027. This replaces the earlier end-of-2026 announcement, which some official pages still show. The Federal Office of Justice prepares these measures (transparency, data protection, non-discrimination, supervision and risk assessment). The Federal Office of Communications (OFCOM) coordinates the non-binding measures such as guidelines, sector standards and voluntary commitments. None of this is a general AI law in force.
Council of Europe AI Convention
Switzerland decided on 12 February 2025 to ratify the Council of Europe Framework Convention on Artificial Intelligence and signed it on 27 March 2025. The announced route is targeted sector changes to Swiss law, not a general AI act. The Convention applies to public authorities; each party decides how it covers private actors.
The Convention itself is not in force. The treaty office chart showed, on 19 November 2026, 22 signatories and two ratifications: the European Union (15 May 2026) and Bosnia and Herzegovina (28 October 2026). Entry into force needs five ratifications, including three Council of Europe member states, then three months. Switzerland has signed but not ratified. Some web pages wrongly say the Convention entered into force in 2025. Check the chart before relying on any date. For a private organization nothing binding follows from the Convention until Swiss law changes. A public body should expect its principles (transparency, oversight, non-discrimination, recourse) to shape what is asked of it.
Federal administration
The federal administration follows an AI strategy presented in March 2025. On 18 September 2026 the Federal Council was informed of a reprioritisation (French): six of the eight implementation measures continue, with a 2026 to 2028 roadmap, because of the resources available. On 2 October 2026 (French) it adopted a report on automation and AI for administrative efficiency, starting with an internal overview of the AI systems the Confederation develops, buys or uses. These are signals about federal practice. They are not evidence of effect and do not bind cantons or communes.
European Union
The AI Act follows a risk-based approach. The role — provider, deployer, importer, or distributor — and the use case determine the obligations.
Status verified on 20 August 2026:
- Article 50 transparency obligations have applied since 2 August 2026;
- the AI Omnibus, which entered into force in July 2026, sets 2 December 2027 as the application date for Annex III high-risk rules;
- high-risk rules for systems embedded in Annex I products apply from 2 August 2028;
- Article 4 retains a duty to support AI literacy appropriate to the role, people, context, and risks, without prescribing a specific individual level;
- national supervision of Article 4 has applied since August 2026.
Prohibitions, rules for general-purpose AI models, transparency, data protection, employment law, consumer protection, intellectual property, and sector-specific rules must be assessed separately.
Operational routing for Article 50:
- direct interaction: determine whether the provider must design the system to inform people that they are interacting with AI;
- synthetic content: determine whether the provider must add detectable machine-readable marking;
- deepfakes and certain public-interest text: determine the deployer’s visible-disclosure duty and the applicable exceptions;
- emotion recognition and biometric categorisation: determine the deployer’s information duty;
- retain the role analysis, interface evidence, marking evidence, exceptions, and verification date.
These EU duties do not replace FADP analysis when Switzerland is also in scope, and Swiss transparency does not by itself prove Article 50 compliance.
What people should see
Examples of wording to adapt, not complete legal compliance statements:
- At the start of a chatbot conversation: “You are talking to an AI assistant. Ask for a person if you need help.” Provide a working handoff, not just a label.
- Beside a qualifying synthetic video or voice: “This video/voice was generated or modified using AI.” Keep the notice understandable without hover.
- For public-interest text, check both human review or editorial control and who holds editorial responsibility before relying on the disclosure exception.
The Commission’s labelling guidance does not require every AI-assisted document to carry the same notice. A visible label and technical machine-readable marking answer different questions. Test any chosen notice with keyboard navigation, screen readers and mobile.
Questions to document for every use case
- What is the exact purpose, and what legal basis authorizes the data processing?
- Which people are affected, informed, and able to challenge the outcome?
- Does a decision produce legal or similarly significant effects?
- Does the system or its use fall into a prohibited or high-risk category?
- Where is the data processed, and by which subprocessors?
- Is the data used to improve or train a provider’s systems?
- What retention, access, correction, export, and deletion capabilities apply?
- Is a data-protection or fundamental-rights impact assessment required?
- Which transparency, documentation, logging, oversight, and training duties apply?
- Which audit, human-recourse, and continuity mechanisms are required?
Use the source register and date the legal qualification in every case file.
To print or save as PDF: Ctrl+P (⌘P on Mac).
Source and history · GitHub