Artificial intelligence systems run on scale. Large training sets, live data feeds, automated profiling, and constant algorithmic scoring are what make a model useful, and every one of those functions touches personal data at some point. When that data belongs to real people, two European legal frameworks decide how the system can be built and operated: the General Data Protection Regulation across the European Union, and the revised Federal Act on Data Protection in Switzerland.
Both statutes were written to be technology neutral, so neither one mentions machine learning or neural networks by name. That neutrality is also why the practical mechanics diverge so sharply once an actual AI deployment is tested against them. Legal grounds for processing, documentation duties, rules for automated decisions, and the shape of enforcement all work differently under each law, and a compliance program built for one does not automatically satisfy the other.
What Is the GDPR?
The General Data Protection Regulation, formally Regulation (EU) 2016/679, became enforceable across every EU member state on May 25, 2018. It replaced a patchwork of national data protection laws with one directly applicable standard, giving individuals a common set of rights over their personal data while giving businesses a single rulebook instead of dozens of national ones.
Its territorial reach is broad by design. Article 3 pulls in any organization established in the EU, and it also captures companies based outside the EU the moment they offer goods or services to people in the Union or monitor their behavior there. A model trained entirely outside Europe is still squarely inside the GDPR scope if it profiles EU residents.
For AI infrastructure specifically, that scope translates into hard requirements. Any pipeline touching identifiable personal data needs a lawful basis before the first record is ingested, has to apply real data minimization rather than collecting everything available, must stay transparent about what it does with the data, and has to honor rights like erasure and objection even after the data has been folded into a model’s weights.
What Is the Swiss FADP?
Switzerland modernized its own privacy law through the revised Federal Act on Data Protection, which took effect on September 1, 2023. The update replaced a statute dating back to 1992 and pulled Swiss practice much closer to European norms, though it kept several distinctly Swiss legal concepts intact.
The FADP protects the personality and fundamental rights of natural persons whose data is processed, a framing rooted in Swiss civil law rather than the EU rights based approach. Like the GDPR, it is written to be technology neutral, so it applies without distinction to a legacy relational database, an edge computing setup, or a generative AI platform.
The Federal Data Protection and Information Commissioner, known as the FDPIC (or EDÖB in German), has confirmed directly that existing Swiss law already governs AI systems without needing a dedicated AI statute, and its published guidance states plainly that current data protection legislation applies to AI supported processing. The office has already put that authority to use. In June 2024, the FDPIC secured a commitment from Meta not to use Swiss users’ data to train its AI models, and in March 2025 it closed an investigation into how the platform then known as X used personal data to train its Grok AI system. Neither case required new legislation. Both relied on the FADP as it already stood.
Scope of Application and Extraterritorial Reach
Deciding whether an AI application falls under the GDPR, the FADP, or both comes down to where the processing has an effect, not the nationality of the people affected.
The GDPR Article 3 extraterritorial rule catches any non-EU AI developer that targets people in the EU through its service or tracks their behavior online. Housing model weights or running server clusters outside Europe offers no shelter from EU enforcement once the processing touches EU residents.
The Swiss FADP works on a similar effects based test. It applies to processing that produces effects in Switzerland regardless of where the processing physically happens or where the company is incorporated. An international AI vendor selling localized analytics, personalized recommendations, or cloud hosted AI tools to Swiss users has to answer to the FDPIC directly, wherever its servers sit.
In practice, most AI systems operating across Europe trigger both laws at once. A single machine learning pipeline pulling training data from users in Zurich and Frankfurt has to satisfy both regulators simultaneously, and compliance with one is never treated as automatic compliance with the other.
Legal Basis and Consent for AI Data Processing
Article 6 of the GDPR lists six lawful bases for processing personal data: consent, contractual necessity, legal obligation, protection of vital interests, public interest, and legitimate interests. Most commercial AI processing ends up relying on either consent or legitimate interests, and each carries its own burden.
Consent under the GDPR has to be freely given, specific, informed, and confirmed through a clear affirmative action, not buried in a pre-ticked box. When a company instead leans on legitimate interests to train a foundation model or run automated analytics, it has to complete a genuine balancing test showing that its business objective does not override the fundamental rights of the people whose data it is using. The European Data Protection Board addressed exactly this question in its Opinion 28/2024, adopted in December 2024, which set out how controllers should structure and document that balancing test specifically for AI model training, including when anonymization claims about a trained model can actually be relied upon.
Swiss law starts from a different premise. Under the FADP, processing personal data is lawful by default provided it respects the statutory principles of good faith, proportionality, purpose limitation, and transparency. A specific legal justification, including consent, becomes mandatory only in narrower circumstances, when the processing violates one of those core principles, when a data subject explicitly objects, or when the data involved is sensitive or subject to high-risk profiling.
Swiss law also defines sensitive personal data a bit more broadly than the GDPR does. Beyond health records, genetic data, biometric identifiers, and religious beliefs, the FADP folds in data on administrative or criminal proceedings and social security measures as well. High-risk profiling occurs when automated processing links data points to evaluate essential aspects of someone’s personality, building a comprehensive picture of their character or behavior. An AI system running predictive credit scoring, candidate screening, or behavioral assessment on Swiss users needs explicit consent unless an overriding private or public interest applies instead.
Transparency and Automated Decision-Making
Transparency is central to both laws whenever an AI model evaluates a real person, though the statutory mechanics differ.
Under Article 22 of the GDPR, individuals have the right not to be subject to a decision based solely on automated processing, including profiling, when that decision produces legal effects or similarly significant impacts on them. Where automated decisions are still permitted, through explicit consent or contractual necessity, the controller has to inform the person that automated decision-making is happening, explain the underlying logic in a meaningful way, and guarantee the right to obtain human intervention, express a point of view, and contest the outcome.
The Swiss FADP handles this under Article 19. If an AI system reaches a decision entirely through automated processing that carries legal consequences or a similarly significant effect for a Swiss resident, the controller has to tell that person about it. Where the GDPR builds in a mandatory human review mechanism, the FADP instead gives the individual the right to state their point of view and ask that a natural person review the decision, subject to a handful of statutory exceptions such as decisions made in direct fulfillment of a contract.
Data Protection Impact Assessments for AI
Rolling out a complex machine learning model often triggers a mandatory risk assessment under both regimes before it ever reaches production.
Article 35 of the GDPR requires a Data Protection Impact Assessment before processing that is likely to create a high risk to individual rights, particularly when the system systematically evaluates personal aspects of people, processes sensitive data at scale, or monitors public spaces.
Switzerland imposes a nearly identical duty under Article 22 of the FADP, triggered when the intended processing poses a high risk to a person’s personality or fundamental rights. Whether that threshold is met depends on the nature, scope, circumstances, and purpose of the processing, and the FDPIC’s own guidance points to extensive processing of sensitive data, systematic monitoring of broad public areas, and high-risk profiling as the clearest triggers. The Commissioner’s published position on AI goes further, expecting controllers to examine model drift, algorithmic bias, the provenance of training data, and security vulnerabilities as part of that assessment rather than treating it as a paperwork exercise.
If a DPIA under the GDPR turns up unmitigated high risk, the company must formally consult its supervisory authority before going any further. Switzerland allows a way around that step. If the organization has formally appointed an independent Data Protection Advisor, it can skip the mandatory FDPIC consultation even after an unmitigated high-risk finding.
Data Breach Notification Requirements
Incident response plans for AI systems need to accommodate two different clocks.
The GDPR requires that breaches presenting a risk to individual rights be reported to the competent EU supervisory authority without undue delay, and wherever feasible within 72 hours of the organization becoming aware of it. If the breach carries a high risk, the affected individuals have to be told directly, also without undue delay.
The FADP takes a more flexible approach to timing. Controllers must notify the FDPIC of a data security breach as soon as possible once it is likely to create a high risk to personality or fundamental rights, but the law does not set a rigid hour-based deadline the way the GDPR does. In practice, the FDPIC still expects rapid reporting once a credible high-risk threat is identified, it simply does not codify a number. Individuals in Switzerland only need to be told directly when doing so is necessary for their own protection, or when the Commissioner specifically orders it.
Data Protection Officers and Privacy Responsibilities
Internal oversight looks structurally different in the two jurisdictions.
Article 37 of the GDPR makes appointing a Data Protection Officer mandatory for public authorities, for organizations whose core activities involve regular and systematic monitoring of people at scale, and for those processing sensitive data at scale.
The Swiss FADP does not require a privacy officer at all for private companies. It offers an optional Data Protection Advisor role instead, and there is a real incentive to fill it. An organization with an appointed advisor can have that advisor approve DPIA risk remediations internally, which is what lets it skip the mandatory FDPIC consultation described above.
Foreign companies with no physical presence in Switzerland that process personal data at scale, or that run high-risk profiling on Swiss residents, still have to designate a formal Swiss representative as a point of contact for both individuals and the FDPIC.
Penalties, Liability and Enforcement
This is where the two laws diverge most sharply.
The GDPR fines corporate entities directly through administrative penalties, up to €20 million or 4% of global annual turnover from the prior financial year, whichever figure is higher. Italy’s data protection authority, the Garante, used exactly that framework against OpenAI on December 20, 2024, imposing a €15 million fine made up of €9 million for processing ChatGPT training data without an adequate legal basis, €320,000 for failing to report a March 2023 data breach affecting roughly 440 Italian users, and €5.68 million for not complying with corrective measures the Garante had already ordered in 2023. Alongside the fine, OpenAI was required to run a six month public awareness campaign across Italian media explaining how ChatGPT collects and uses personal data for training and how users can exercise their rights.
The Swiss FADP works through criminal law instead, and it targets natural persons rather than the corporate entity itself. A company director, executive, or designated privacy manager can face a personal criminal fine of up to CHF 250,000 under Articles 60 through 63 for intentional violations such as wilfully failing to meet information duties, giving false or misleading disclosures, carrying out unauthorized cross-border transfers, or breaching statutory confidentiality obligations. Where GDPR fines reach negligent conduct as well as intentional violations, Swiss criminal sanctions require intent. A company can only be fined directly, and only up to CHF 50,000, when identifying the specific individual responsible would take a disproportionate investigative effort, under Article 64.
International Data Transfers and AI Providers
Modern AI infrastructure depends on external model providers, cloud compute, and API services spread across the globe, which means personal data crosses borders constantly, and each law handles that movement differently.
The GDPR blocks transfers of personal data outside the European Economic Area unless the destination is covered by an EU adequacy decision, Standard Contractual Clauses, Binding Corporate Rules, or one of the narrow derogations built into the regulation.
Switzerland keeps its own list of adequate jurisdictions, set independently by the Swiss Federal Council, and while it overlaps heavily with the EU list, compliance teams still need to check the Swiss schedule specifically rather than assume equivalence. For transfers from Switzerland to the United States, the Swiss-U.S. Data Privacy Framework became operational on September 15, 2024, giving companies a certified mechanism roughly parallel to the EU-U.S. framework. Where a US recipient is not certified under that framework, Swiss adapted versions of the EU Standard Contractual Clauses, carrying specific Swiss legal addenda, remain the standard fallback.
Privacy Notices and Data Processing Information
An AI product privacy notice has to carry real technical detail under both regimes, not boilerplate. A compliant notice for a platform operating in European markets needs to spell out:
- Who the data controller is and how to reach any local representative
- Exactly what categories of personal data the system collects, including user prompts and behavioral signals
- Why the data is processed, whether for fine-tuning a model, monitoring for safety, or personalizing output
- Which third parties, sub-processors, and infrastructure vendors receive the data
- Where the data travels internationally and which transfer mechanism covers it
- How long raw logs, caches, and derived embeddings are kept
- What rights the person has to access, correct, delete, or object to the processing
- How any automated decision-making or high-risk profiling actually works
The Swiss FADP requires that disclosure to happen at the moment of collection, with the identity of the controller and the precise purpose of processing stated clearly at that point rather than buried in a document the user finds later.
Data Subject Rights in AI Systems
Honoring individual rights inside a deep neural network or a vector database is not the same exercise as honoring them in a conventional relational database, and both statutes grant the same core set of rights even though enforcing them against a trained model raises real technical questions.
- Right of Access: Lets a person confirm whether their data is being processed and obtain copies of the personal data used in training or inference.
- Right to Rectification: Lets them demand correction of inaccurate details, whether those sit in a static profile or an active model database.
- Right to Erasure: Lets them request deletion of their personal records. In a machine learning context, that can mean pulling records from an active database and, where personal data is embedded directly in the model parameters, retraining or fine-tuning the model itself.
- Right to Object: Lets a person stop their data being used for profiling or promotional targeting.
- Right to Data Portability: Granted under both the GDPR and the revised FADP, this lets someone receive their data in a structured, commonly used, machine readable format, or have it transferred directly to another provider.
GDPR vs FADP: Key Differences for AI Data Privacy
| Requirement | GDPR | Swiss FADP |
| Territorial Scope | Applies to EU establishments or entities targeting or monitoring EU residents. | Applies to processing activities that produce effects within Switzerland. |
| Legal Basis | Requires one of six specific legal bases under Article 6. | Processing is generally lawful by default if core statutory principles are respected. |
| Consent | Mandatory when chosen as the legal basis, under a strict free and explicit standard. | Mandatory mainly for sensitive data, high-risk profiling, or statutory breaches. |
| Automated Decisions | Prohibition based framework under Article 22, with mandatory human review rights. | Information obligation under Article 19, with rights to state a view and request human review. |
| DPIA Requirement | Mandatory for specified high-risk processing under Article 35. | Mandatory when processing creates high risks to personality or fundamental rights. |
| Breach Notification | Report to the supervisory authority within 72 hours where risk exists. | Report to the FDPIC as soon as possible when high risk is present. |
| Privacy Officer | Mandatory DPO required under specific statutory thresholds. | No mandatory DPO. An optional Data Protection Advisor framework is available instead. |
| International Transfers | Governed by European Commission adequacy decisions and EU SCCs. | Governed by Swiss Federal Council adequacy lists and Swiss adapted transfer clauses. |
| Enforcement | Administrative corporate fines up to €20 million or 4% of global turnover. | Criminal fines up to CHF 250,000 targeting the responsible individual. |
What the Differences Mean for Organizations Using AI
Organizations deploying artificial intelligence across Switzerland and the EU cannot treat compliance as a single checklist copied from one jurisdiction to the other. The broad structural alignment between the FADP and the GDPR makes the basic architecture easier to design, but the operational divergences above still demand governance built for each market specifically.
A few habits separate compliance teams that get this right from those that get caught out:
- Audit every data intake channel to confirm whether the processing touches EU residents, Swiss residents, or both, mapping the flow across training pipelines, external API gateways, and cloud subprocessors rather than assuming.
- Verify the legal justification for each use of data separately. Legitimate interest balancing tests have to satisfy EU standards on their own terms, while any profiling model touching Swiss users needs to be tested against the FADP explicit consent trigger for high-risk parameters.
- Build DPIA templates around the actual risks an algorithm poses, documenting automated decision logic, building in genuine human-in-the-loop review, and keeping privacy notices current as model training practices change.
- Run incident response plans on both clocks at once, ready to meet the GDPR 72-hour reporting rule and the FADP as-soon-as-possible standard without treating either as the default.
- Ensure leadership understands, in plain terms, that Swiss law can reach them personally under criminal statute for an intentional privacy violation in a way the GDPR corporate fines do not.
Passing a GDPR audit is not proof of FADP compliance, and the reverse is just as true. As the OpenAI and Meta cases both show, regulators on both sides of the Swiss-EU border are actively applying these existing rules to AI systems right now, not waiting for AI-specific legislation to catch up. Clear regional data mapping, jurisdiction-specific policy language, and risk assessments built for how a given model actually behaves remain the baseline for any organization running AI across European markets.