Secure AI FAQ: Answers on Affordable, Secure AI

A new survey on Dr. DSGVO asked which topics around AI and the GDPR matter most to readers. This secure AI FAQ takes up those topics and answers the questions.

Introduction

The survey on AI and data protection (“Which topics interest you?”) produced exciting insights into what moves companies and public authorities.

People asked in particular about sovereign AI and affordable AI systems. They also asked how the AI from OpenAI or Anthropic can be used in a GDPR-compliant way.

This post takes up the most relevant questions and answers them. It focuses on AI (mostly in connection with data protection and data security). A further post will address the data protection questions that have no connection to AI.

The survey mentioned above is still active and can be used to submit your questions and wishes.

Secure AI FAQ: Your Questions and Our Answers

Topic “Secure AI”

Most questions and suggestions concerned secure AI. Many also say “GDPR-compliant AI”, “privacy-compliant AI” or “local AI”. They all mean the same thing:

A secure AI system processes data for well-defined, legitimate purposes that are determined exclusively by the controller (the operator of the AI).

Privacy-compliant does not automatically mean secure:

• First, the GDPR applies “only” to personal data, not to trade secrets, confidential data (NDA) or sensitive data (“We do not want this data to reach competitors”, etc.).
• Second, “GDPR-compliant” is a purely virtual definition that is currently based on the TA-DPF (= the informal transatlantic data privacy agreement between the EU and the US). US laws on foreign surveillance allow the US to force US companies such as Microsoft to hand over your data so it can be abused – without Microsoft being allowed to tell you.

To clear up misunderstandings, here are a few facts:

• The US surveillance laws that allow Germans and Europeans to be screened in secret (and in breach of the GDPR) include, in particular, the Cloud ACT, EO 12333, FISA 702 and the SCA.
• A US mechanism called IEEPA / OFAC allows US authorities to screen companies and individuals if they pay their invoices in US dollars (USD).
• The location of a server is almost irrelevant when it comes to data security. Unlike a car, a server can be accessed by anyone who has the credentials, no matter where it is located. Customer-managed encryption keys (Bring Your Own Key etc.) are not effective protection either: as soon as the data has to be worked on in the cloud, the key has to be loaded into the cloud. The key is then no longer secret.

Question: How can AI be used in a privacy-compliant way?

The formal answer is: providers headquartered in countries (plus sub-processors as further data recipients) that have been defined as safe can be suitable for purely formal GDPR compliance. This requires a legally valid data processing agreement (DPA) and all data recipients must be classified as safe. Recipients in the US must be registered with the TA-DPF. This registration is a self-registration. That alone shows the difference between formal and real security.

Most DPAs from non-European providers are probably not compatible with the GDPR. Try to understand and interpret Microsoft’s contracts once. Ask Microsoft whether your interpretation is correct.

The truth is:

Whether there are Icelandic AI providers that meet all the requirements is an open question. Probably not.

Regardless of the data protection question, there is the question of data security: your trade secrets are certainly not safe in the cloud of a US provider. See the US surveillance laws mentioned above.

Local AI use means: the AI runs on your own hardware that you have bought, or on rented hardware located exclusively in Germany (including its legal framework and legal understanding).

Question: How can the AI from Anthropic or OpenAI be used in a GDPR-compliant way?

Short answer: not at all, or only formally (see the previous question and answer).

In fact, secure data processing through the systems of US providers such as Anthropic, Google, Microsoft, Amazon or OpenAI is not possible.

The truth is:

Incidentally, whom you entrust with your trade secrets or your own data is entirely your decision, or your company’s. Some apparently have no problem blabbing out their own secrets and thereby devaluing themselves.

In the short term, there is no noticeable disadvantage in passing sensitive data to untrustworthy third parties (who must submit to US surveillance laws). In the long run, this approach cannot be sustainable, as life experience shows.

This post is not meant to convert anyone (it achieves nothing and is too exhausting). Do whatever you want and live with the consequences.

Question: How can secure AI be used in a company when there is little money?

Short answer: not at all.

A question in return: why would anyone give you strong guarantees plus great functionality for free?

For a longer answer, two cases need to be distinguished:

1. Your use case is general, i.e. designed in such a way that it is of interest to many.
2. Your use case is specific, i.e. located specifically in your company.

The truth is:

For general use cases, you might be lucky and find an AI solution as a service (SaaS) on the market from a provider in Germany or Europe.