Glossary: AI and data protection

Adequacy decision / EU-US Data Privacy Framework

An adequacy decision is a finding by the European Commission that a third country ensures an adequate level of protection (Art. 45 GDPR). For the USA, the EU-US Data Privacy Framework decision (Implementing Decision (EU) 2023/1795) applies; it helps only with providers certified under it. The EU General Court dismissed an action against it on 3 September 2025; an appeal is pending (as of 13 November 2026).

More on this: Is ChatGPT GDPR compliant? Plans, DPA and what is left to do

AI Act

The AI Act is Regulation (EU) 2024/1689, which sets obligations for providers and deployers of AI systems according to risk classes. It has applied generally since 2 August 2026, with some exceptions; for high-risk systems under Annex III the date is 2 December 2027 under Regulation (EU) 2026/1744. The GDPR continues to apply alongside it.

AI literacy (Article 4 AI Act)

AI literacy is the knowledge and skill needed to work with AI. Under Article 4 of the AI Act, as amended by Regulation (EU) 2026/1744, providers and deployers must take measures to support the AI literacy of their staff and of other persons using AI systems on their behalf; the duty has applied since 2 February 2025. A policy and training are typical measures; see the AI acceptable use policy template.

More on this: AI acceptable use policy template for EU companies

Anonymisation

Anonymisation is the process of rendering personal data anonymous, so that the person is not or no longer identifiable; anonymous information does not relate to an identifiable person and falls outside the GDPR (recital 26). Whether someone is identifiable depends on all the means reasonably likely to be used. According to the German Data Protection Conference (DSK), removing names and addresses is regularly not enough.

More on this: Pseudonymisation vs anonymisation under the GDPR: differences, examples and the CJEU ruling

Controller

A controller is the natural or legal person, public authority, agency or other body that, alone or jointly with others, determines the purposes and means of the processing of personal data (Art. 4(7) GDPR). A company that lets its staff use AI tools remains responsible for the data in their prompts.

More on this: Use AI at work without leaking company or personal data: 12 rules

Data minimisation

Data minimisation means that personal data must be adequate, relevant and limited to what is necessary for the purposes of processing (Art. 5(1)(c) GDPR). For prompts, that means entering only what the task needs and leaving out or replacing everything else.

More on this: Use AI at work without leaking company or personal data: 12 rules

Data processing agreement (DPA)

A data processing agreement is the contract that governs processing by a processor (Art. 28(3) GDPR). Among other things it must provide that the processor acts only on documented instructions, and it may be concluded electronically (Art. 28(9)). It replaces neither a legal basis nor an assessment of the provider.

More on this: Is ChatGPT GDPR compliant? Plans, DPA and what is left to do

Data protection impact assessment (DPIA)

A DPIA is an assessment of the risks of processing that is likely to result in a high risk for individuals, to be carried out beforehand, in particular where new technologies are used (Art. 35 GDPR). The German Data Protection Conference (DSK) considers it frequently necessary when AI is used.

Data residency

Data residency means content is stored in a particular region, for example the EU. It does not mean every processing step takes place in that region; OpenAI names it only for eligible customers of certain plans, and Microsoft's EU Data Boundary has exceptions.

More on this: ChatGPT, Microsoft Copilot, Claude and Gemini: privacy for business use compared

Deployer

A deployer is anyone who uses an AI system under their own authority, except in the course of a personal, non-professional activity (Art. 3(4) AI Act). A company that uses ChatGPT, Copilot or another tool for work is a deployer; its duties differ from those of the provider that develops or places the system on the market.

High-risk AI system

A high-risk AI system is one that the AI Act regulates more strictly, for instance because it is used in an area listed in Annex III (Art. 6(2) AI Act); this includes employment, such as systems used to select applicants. Deployers then have duties under Art. 26 AI Act, which apply from 2 December 2027 under Regulation (EU) 2026/1744.

Information duties

Information duties require the controller to tell data subjects about the processing, including purposes, legal basis, recipients and retention (Art. 13 and 14 GDPR). Art. 13 applies when data is collected from the person, Art. 14 when it comes from another source. An AI provider you use may have to be named as a recipient or category of recipients.

Key / mapping table

The key, or mapping table, is the separately stored list that links placeholders or pseudonyms back to the original values. It corresponds to the “additional information” in Art. 4(5) GDPR. Whoever holds it can reverse the pseudonymisation and therefore keeps processing personal data.

More on this: How to anonymise data before ChatGPT (and why you are really pseudonymising)

Large language model (LLM)

A large language model is an AI model trained on large amounts of text that generates text by computing likely continuations of an input. Chat assistants such as ChatGPT, Claude, Gemini and Copilot are built on such models.

More on this: ChatGPT, Microsoft Copilot, Claude and Gemini: privacy for business use compared

Legal basis

A legal basis is one of the conditions in Art. 6(1) GDPR without which personal data may not be processed: consent, contract, legal obligation, vital interests, public interest or legitimate interests. According to the German Data Protection Conference (DSK), every processing step with personal data in an AI application needs a legal basis.

Legitimate interests

Legitimate interests are a legal basis where processing is necessary for the legitimate interests of the controller or a third party and these are not overridden by the interests or fundamental rights of the person concerned (Art. 6(1)(f) GDPR). It requires a balancing test, which in practice should be documented, and public authorities cannot rely on it for their tasks.

Personal data

Personal data is any information relating to an identified or identifiable natural person (Art. 4(1) GDPR). A person is identifiable if they can be recognised directly or indirectly, for example through an identification number or a combination of features. What you type into an AI tool is personal data as soon as it relates to such a person.

More on this: Use AI at work without leaking company or personal data: 12 rules

PII

PII stands for personally identifiable information and is mainly a US term. It does not map exactly onto “personal data” under Art. 4(1) GDPR; in privacy tools it usually stands for names, addresses, account numbers and ID numbers.

More on this: PII redaction for LLM prompts: methods, limits and how to measure them

Processor

A processor is a natural or legal person, public authority, agency or other body that processes personal data on behalf of the controller (Art. 4(8) GDPR). A controller may only use processors that provide sufficient assurances that they will implement appropriate technical and organisational measures (Art. 28(1)). On business plans, an AI provider can act as a processor.

More on this: Is ChatGPT GDPR compliant? Plans, DPA and what is left to do

Professional secrecy

Professional secrecy protects secrets entrusted to members of certain professions such as lawyers, tax advisers and doctors; unauthorised disclosure is a criminal offence in Germany under section 203 of the Criminal Code (StGB). It goes beyond the GDPR because it also protects the secrets of legal persons. A data processing agreement alone does not meet the requirements; professional rules such as section 43e BRAO and section 62a StBerG add to them.

Prompt

A prompt is the input to an AI system: the question, the instruction and any text or file that comes with it. If it contains personal data, sending it to the provider is processing under the GDPR.

More on this: Use AI at work without leaking company or personal data: 12 rules

Pseudonymisation

Pseudonymisation is processing personal data so that it can no longer be attributed to a specific person without additional information, which is kept separately and protected by technical and organisational measures (Art. 4(5) GDPR). Pseudonymised data remains personal data where it could be attributed to a person using that additional information (recital 26).

More on this: Pseudonymisation vs anonymisation under the GDPR: differences, examples and the CJEU ruling

Purpose limitation

Purpose limitation means that personal data must be collected for specified, explicit and legitimate purposes and not further processed in a manner incompatible with those purposes (Art. 5(1)(b) GDPR). Using data collected for one task in another AI project therefore requires a check that the new purpose is compatible.

Quasi-identifier

A quasi-identifier is a detail that contains no name on its own but can make a person identifiable in combination with other details, for example a date, a place, an amount, an age or an unusual job title. This is why replacing names alone is not enough when pseudonymising.

More on this: How to anonymise data before ChatGPT (and why you are really pseudonymising)

Re-identification

Re-identification is linking pseudonymised or seemingly anonymous data back to a person, for example with the mapping table or by combining details from the text with other sources. Whether it is possible with means reasonably likely to be used decides, under recital 26, whether data counts as anonymous.

More on this: Pseudonymisation vs anonymisation under the GDPR: differences, examples and the CJEU ruling

Records of processing activities

The records of processing activities are the written overview of purposes, categories of data, recipients, third-country transfers, erasure periods and security measures that controllers must keep (Art. 30 GDPR). The exemption for organisations with fewer than 250 employees does not apply if the processing is not occasional, is likely to result in a risk or involves special categories. A use of an AI tool that involves personal data is a processing activity to be recorded.

Redaction

Redaction means removing content from a document so that it cannot be reconstructed. A black box placed over text is not enough, because the text stays in the file; metadata, comments, form fields and the file name must be checked too. It differs from pseudonymisation, where details are replaced so that the AI can still work with the text.

More on this: PII redaction for LLM prompts: methods, limits and how to measure them

Shadow AI

Shadow AI is the use of AI tools by employees without the organisation's knowledge or approval, for example through a private account. It is a risk because data can end up on plans for which the organisation has no contract and no control.

More on this: Shadow AI: detect, reduce and replace unsanctioned use

Special categories of personal data

Special categories are data revealing racial or ethnic origin, political opinions, religious or philosophical beliefs or trade-union membership, genetic data, biometric data for unique identification, health data, and data about sex life or sexual orientation (Art. 9(1) GDPR). Processing them is prohibited in principle unless an exception in Art. 9(2) applies.

Standard contractual clauses (SCCs)

Standard contractual clauses are standard data protection clauses adopted by the European Commission that can serve as an appropriate safeguard for a transfer to a third country (Art. 46 GDPR). They can be used where there is no adequacy decision and are often part of a provider's data processing agreement.

More on this: Is ChatGPT GDPR compliant? Plans, DPA and what is left to do

Sub-processor

A sub-processor is a further processor engaged by the processor for specific processing activities. This requires the controller's prior specific or general written authorisation, and the initial processor remains fully liable to the controller (Art. 28(2) and (4) GDPR).

More on this: Is ChatGPT GDPR compliant? Plans, DPA and what is left to do

Technical and organisational measures (TOMs)

Technical and organisational measures are the safeguards by which controllers and processors ensure a level of security appropriate to the risk, such as pseudonymisation and encryption (Art. 32(1) GDPR) and data protection by default (Art. 25). Redacting documents before they are shared can be another such measure.

More on this: PII redaction for LLM prompts: methods, limits and how to measure them

Third-country transfer

A third-country transfer is a transfer of personal data to a country outside the EU and the EEA. It is allowed only if the conditions of Chapter V GDPR (Art. 44 ff.) are met so that the level of protection is not undermined, for example through an adequacy decision or appropriate safeguards. With US providers this matters in practice.

More on this: Is ChatGPT GDPR compliant? Plans, DPA and what is left to do

Training on user data

Training on user data means a provider uses its customers' inputs and outputs to improve its models. On consumer plans this is often a default or a setting; on business plans OpenAI and Anthropic say it is excluded by default. Plans and terms change, so check the provider's current page.

More on this: Is ChatGPT GDPR compliant? Plans, DPA and what is left to do