Skip to content

Personal data in the AI cloud: who is watching?

A conversation with an AI service in the cloud is not private by default. A case in Canada shows what a provider can see and share. What does that mean under the GDPR, in the United Kingdom and in Switzerland?

In short

Employees use AI to summarise contracts, answer customer questions and analyse files. Names, addresses, financial details and sometimes medical data go along. Many users assume such a conversation stays between them and the AI. That is not a given.

With an AI service in the cloud, the data goes to a supplier's systems. Depending on the service and the contract, it is stored, checked automatically for abuse and in certain cases reviewed by the provider's staff. That does not mean someone reads every conversation. It does mean you cannot assume your information is used only to answer your question.

What the case in Canada shows

On 10 February 2026 an eighteen-year-old killed eight people and then herself in Tumbler Ridge, Canada. It later emerged that she had discussed violent scenarios with ChatGPT.

OpenAI had already flagged her account in June 2025 with systems that detect abuse, and had banned it. The company considered informing the police, but decided its own threshold had not been met. After the shooting it did contact the police and provided information about her use. In September 2026 the province of British Columbia sued OpenAI. That is an allegation; no court has ruled on it.

For your organisation the point is not the content of those conversations, but what the case shows about a provider's capabilities:

  • Conversations are checked automatically for content.
  • A provider can flag behaviour and ban accounts.
  • The provider's staff can review conversations.
  • Data about use can be passed to law enforcement.

It is understandable that a provider acts on signs of serious violence. But it also makes clear that whoever hands confidential information to an external AI service gives up part of the control.

What the GDPR asks

A persistent misunderstanding is that personal data is only processed when it is stored or used to train a model. Sending it to an AI service and having it analysed there is processing too. An organisation must therefore be able to show that:

  1. there is a valid legal basis for the processing;
  2. the data is used only for a defined purpose;
  3. no more data goes along than necessary;
  4. the people concerned have been informed;
  5. the security is appropriate;
  6. all of this is documented.

If an external supplier processes the data on behalf of your organisation as a processor, a data processing agreement is part of it (Article 28 GDPR). If the data leaves the European Economic Area, additional safeguards are required. The European regulators have confirmed that AI simply falls under these rules.

So the GDPR does not prohibit an AI service in the cloud. A business service with sound agreements, a valid legal basis and appropriate security is possible. An employee pasting a complete personnel file into a private account is not. That a service is available does not make it suitable for your organisation's data.

Beyond the EU: the United Kingdom and Switzerland

The GDPR applies throughout the European Union. The United Kingdom has its own version, the UK GDPR, alongside the Data Protection Act 2018, with largely the same principles.

Switzerland has its own Federal Act on Data Protection (FADP, in German DSG), revised as of 1 September 2023. The principles resemble the GDPR: use data only for a recognisable purpose, no more than necessary, well secured, with an agreement when someone else processes it for you and with safeguards when it leaves the country. In November 2023 the Swiss regulator (FDPIC) stated that this act applies directly to AI. For certain intentional violations, the responsible person can be fined up to 250,000 Swiss francs.

Working in several countries therefore means dealing with several laws, and with a service from an American provider also with the question of which law that provider itself falls under. The less data leaves your environment, the simpler this becomes.

The questions left unanswered

An employee asks an AI assistant to analyse sick leave in his department. He uploads a spreadsheet with names, days of absence, medical remarks and reintegration agreements. Within seconds he has a useful analysis.

For him that is a success. For the organisation, questions remain. Where was the data processed? Was it stored, and for how long? Who can access it? Can you show that it was allowed?

A provider that promises not to use your data for training says nothing about storage, logs and safety checks. Not used for training is not the same as not processed.

Local AI: control back in your hands

The alternative is local AI: the model runs on systems you manage yourself or have managed under your direct control. You decide where the data is processed, who has access and what is recorded.

  • Data stays inside. Sensitive information does not leave your own environment.
  • Your own rules. You set the policy for storage, logs and access.
  • Existing permissions. The AI connects to the authorisations you already have.
  • Predictable costs. With daily use, running locally is often favourable.
  • Less dependence. A provider's changed terms affect you less.

Local AI does not make you compliant by itself. Legal basis, access and retention periods remain your responsibility, in every country. The difference is that you hold the means to enforce those rules yourself. It does not take a data centre: modern open models run on compact hardware.

Our approach

AI should not be a separate tool into which employees enter data as they see fit. We make AI part of your IT architecture, with the same requirements as for any other business-critical system.

  1. Permissions follow the user. The AI sees only what the user is allowed to see.
  2. Controlled connections. Data reaches the AI through fixed, managed connections.
  3. Local where possible. Processing on your own managed systems.

Our principle: bring AI to your data, instead of your data to just any AI cloud.

Would you like advice on personal data and AI? Request a free advice conversation.

Sources

Updated on

AI

Why the same AI model does not always give the same answer

The same question, the same model, the same settings and even the same seed, and still a different answer. In the trial we describe here, rounding in the compute kernels played a part. For testing, checks and accountability you need to know this.