Rules
Privacy by design and by default for AI tools: what to settle when choosing a tool, which settings to switch off and what to ask of the supplier.
Privacy by design means building in the protection of personal data at the moment you decide how a processing operation will work, for example when you choose an AI tool. Privacy by default means that the default settings only process what the purpose requires. Both are in Article 25 of the GDPR and are a duty of the organisation itself.
The text of the GDPR calls it "data protection by design and by default". The first paragraph asks for appropriate technical and organisational measures, "both at the time of the determination of the means for processing and at the time of the processing itself". The regulation names pseudonymisation as an example of such a measure; the difference with anonymising is explained in Pseudonymising or anonymising: the difference, and why it matters for AI. What is appropriate depends on the state of the art, the cost, the nature and purpose of the processing and the risks.
The second paragraph is about the default, and names four things: the amount of personal data collected, the extent of their processing, how long they are stored and who can access them. In principle, data may not be made accessible to an indefinite number of people without human intervention.
The Autoriteit Persoonsgegevens (AP), the Dutch data protection authority, gives everyday examples on its page about accountability: an app that does not record location when that is not needed, a box for offers that is not ticked in advance and a newsletter that asks for no more data than necessary. In an AI tool, the same choices sit deeper in the settings: in the provider's admin screen, in the connections with other systems and in what staff share with each other by default. This article is an explanation, not legal advice: assess your own situation with your data protection officer or a lawyer.
You do. Article 25 is addressed to the controller: the organisation that decides why and by what means personal data is processed. About the makers of products and services, recital 78 of the GDPR only says that they "should be encouraged" to take data protection into account.
The European Data Protection Board (EDPB), the body in which the European privacy supervisory authorities work together, draws the consequence in its Guidelines 4/2019 on Article 25 (version 2.0 of 20 October 2020). The controller is also responsible for what processors and sub-processors do, and takes this into account in its contracts (paragraph 1). And: do not choose a supplier whose system does not help you meet Article 25, because you will be held accountable if it is missing (paragraph 96).
So if a feature in an AI tool is on by default and processes more than your purpose requires, that is your processing and your responsibility. The data processing agreement governs what the supplier may do with the data. Article 25 also asks something of you: that you choose and set up the tool so that no more happens than necessary.
Before the first use. According to the EDPB, the moment you procure and implement software and services also falls under "the determination of the means for processing" (paragraph 35). If you buy an AI tool first and only think about privacy afterwards, you are too late for part of the assessment. A feature that is already in the contract, or a setting the whole firm already uses, is harder to undo later than to rule out in advance.
After that, the duty continues: maintaining the measures, keeping up with the state of the art and reassessing the risk (paragraph 37), including for what processors do for you (paragraph 39). AI providers regularly add features, and a new feature is not automatically covered by the assessment you made back then. So plan a fixed moment to go through the settings.
If you use off-the-shelf software from someone else, assess the risks of the product and switch off features that have no lawful basis or do not fit the purpose, says the EDPB (paragraph 44). Also divide access so that staff with different roles only see what they need (paragraph 45). The four things in the second paragraph of Article 25 provide a useful structure.
| What Article 25(2) names | The question for an AI tool | A default you can choose |
|---|---|---|
| Amount of data | What goes into the tool? | Only the passages a question needs, not whole client files |
| Extent of processing | What else does the tool do with it? | No use of input to improve models, features without a purpose switched off |
| Storage period | How long do input, answers and history remain? | A period that fits the purpose and is set in advance |
| Accessibility | Who can access it? | Access per team or project, no conversations shared with the whole organisation |
How to justify a retention period is explained in Retention periods: how long may you keep personal data and personnel files?.
The Irish supervisory authority fined TikTok a total of 345 million euros in September 2023. The inquiry covered, among other things, settings for children that were public by default, and the decision records infringements of, among others, Article 25(1) and 25(2). Not an AI tool, but the same point: what is on by default counts. Under Article 83(4) of the GDPR, a breach of Article 25 falls under fines of up to 10 million euros or, for an undertaking, up to 2 percent of worldwide annual turnover if that is higher.
The EDPB recommends requiring suppliers to show how their product helps you demonstrate accountability, and including a contractual clause that they keep you informed of changes in the state of the art (paragraph 96). Six questions make that practical.
Record the answers. They show that you thought about the choice, and they belong in the description of the measures if you carry out a DPIA; see DPIA for AI tools: when is it mandatory and how do you approach it?.
PrivaSearch sits between your documents and the AI model; the steps are on How PrivaSearch works. Personal data is detected inside PrivaSearch itself, without an external AI provider, and replaced by pseudonyms. A colleague sees the original alongside the pseudonymised text and explicitly approves it; only then does a document become searchable. A connected assistant, Claude, ChatGPT Business or Mistral Vibe, receives only released, pseudonymised passages, limited in size, and no original files, filenames or links. Project membership determines who sees what, and administrators get no automatic access. After the end date of a project, access is blocked and copies, fragments and pseudonyms are cleaned up, unless a legal hold applies.
A scan without a text layer only goes as a file to a European AI provider to be read if your organisation switches that on itself; see Hosting in the EU. What you type into a connected assistant yourself goes directly to the provider of that assistant, outside the pseudonymisation.
PrivaSearch does not make the Article 25 assessment for you: whether these measures fit your processing remains your judgement.
Yes. Article 25 of the GDPR applies to every organisation that processes personal data, regardless of its size. What is appropriate depends on, among other things, the cost, the nature of the processing and the risks, so a small firm with sensitive client files may have to do more than a large company handling only public information.
Privacy by design is about the design: you build in measures such as pseudonymisation as soon as you decide how a processing operation will work. Privacy by default is about the default: without anyone changing anything, the system processes, stores and shares no more than the purpose requires.
Not under Article 25 of the GDPR: that duty lies with the organisation that uses the tool. Under the GDPR, makers only have to be encouraged to take data protection into account. The European supervisory authorities therefore advise against choosing a supplier whose product does not help you meet Article 25.
Often, yes. According to the European supervisory authorities, with off-the-shelf software you assess the risks and switch off features that have no lawful basis or do not fit the purpose. Think of the use of input to improve models, a long retention period for conversations or conversations the whole organisation can see.
No. Pseudonymisation is an example that Article 25 names, but the article asks for a set of technical and organisational measures, such as restricted access, appropriate retention periods and switching off unnecessary features. Pseudonymised data also remains personal data for the organisation that holds the key.
A breach of Article 25 falls under fines of up to 10 million euros or, for an undertaking, up to 2 percent of worldwide annual turnover if that is higher. And if you have not recorded your choices at purchase, it is hard to show afterwards that you considered the duty.
When a decision involving AI falls under the ban on automated decision-making, which exceptions apply, and how a human genuinely takes part.
What confidentiality means when you put a client file in an AI chatbot, and what the NOvA, KNB, NBA and CCBE have published about this practice.
What the AIIA of the Dutch Ministry of Infrastructure is, who must use it, how it is structured and how it relates to the IAMA and the DPIA.
We show it with fictitious client files from your sector: upload, detection, review, search and the connection with your AI assistant. Afterwards you'll know exactly what your team can use safely.
Book a demo →