Purpose Limitation

The purpose for data collection must be explicitly stated at the point of collection, not buried in vague terms of service. This means defining the precise operational context—such as 'account authentication' or 'order fulfillment'—rather than using broad, catch-all phrases like 'improving our services.' Generalized consent is invalid; the specification must be granular enough to allow a reasonable person to understand the scope of processing. This directly conflicts with the data-hungry nature of foundation model training, where data is often scraped indiscriminately for undefined future use cases.

Any secondary processing must pass a rigorous compatibility test to determine if the new purpose is aligned with the original. Regulators assess factors including:

• The link between the original purpose and the new processing
• The context of collection and the data subject's reasonable expectations
• The nature of the data and its sensitivity
• The consequences of further processing for individuals
• The existence of safeguards like encryption or pseudonymization Training a commercial large language model on customer support transcripts originally collected for quality assurance would almost certainly fail this test.

A secondary use is deemed incompatible when it exceeds the reasonable expectations of the data subject or introduces new risks. Classic examples of incompatible repurposing include:

• Using customer service call recordings to train a voice synthesis model
• Feeding medical records collected for treatment into a diagnostic AI
• Scraping public social media posts to build personality profiling models
• Repurposing employee productivity data to train an automated hiring classifier In each case, the downstream AI application introduces consequences—biometric identification, algorithmic discrimination, surveillance—that are fundamentally disconnected from the original transaction.

When a new purpose is incompatible with the original, the only lawful path forward is obtaining fresh, specific consent. This consent must be:

• Freely given: Not bundled as a condition of service
• Specific: Explicitly naming AI training as the purpose
• Informed: Explaining what models will be trained and what they will generate
• Unambiguous: Requiring a clear affirmative action Pre-ticked boxes, implied consent from continued browsing, or blanket 'we use your data for AI' statements in privacy policies do not meet this threshold under modern regulatory standards.

A critical exception exists for archiving in the public interest, scientific research, or statistical purposes. Under GDPR Art. 89, further processing for these purposes is not automatically considered incompatible, provided appropriate technical and organizational safeguards are in place. However, this exemption is not a free pass for commercial AI training. The research must serve a genuine public benefit, and the safeguards—such as data minimization, pseudonymization, and prohibition of re-identification—must be demonstrably effective. Scraping the open web to build a proprietary foundation model does not qualify.

Violating purpose limitation carries severe financial and reputational consequences. Under GDPR, fines can reach €20 million or 4% of global annual turnover, whichever is higher. Beyond fines, enforcement actions can include:

• Cease and desist orders halting model training or deployment
• Mandatory data deletion from training corpora
• Model disgorgement requiring retraining without the tainted data
• Audit mandates imposing ongoing compliance monitoring Regulators are increasingly focusing on AI training pipelines, with multiple European Data Protection Authorities issuing guidance that scraping public data for AI training without a lawful basis violates purpose limitation principles.

A core privacy principle mandating that data collection be limited to what is strictly necessary for a specific, explicit purpose. In AI contexts, this directly challenges the indiscriminate scraping practices common in foundation model training. Organizations must demonstrate that every data point ingested serves the defined purpose and nothing beyond it.

• Reduces attack surface for unauthorized secondary use
• Requires documented justification for each data field
• Conflicts with 'collect everything now, decide later' approaches

A data governance principle requiring that personal data be kept in an identifiable form for no longer than necessary to fulfill the original purpose. Once the purpose expires, data must be anonymized or deleted. For AI training corpora, this mandates systematic purging of stale data and creates technical challenges for models that have already internalized patterns from that data.

• Enforces temporal boundaries on data utility
• Requires automated data lifecycle management
• Complicates long-term model improvement strategies

A legal provision under GDPR Article 21 granting individuals the absolute right to object to processing of their personal data based on legitimate interests or public tasks. When invoked against AI training, organizations must cease processing unless they demonstrate compelling legitimate grounds that override individual rights.

• Applies even if original collection was lawful
• No grace period—processing must stop immediately
• Requires mechanisms to exclude objectors from training pipelines

A three-part balancing test required under GDPR to document whether an organization's commercial interest in processing data overrides individual rights. For AI training, the LIA must weigh the necessity and proportionality of using personal data against the reasonable expectations of data subjects.

• Part 1: Identify the legitimate interest (purpose)
• Part 2: Demonstrate necessity—no less intrusive alternative
• Part 3: Balance against individual rights and freedoms

A legally binding contract between a data controller and data processor that stipulates the specific scope, purpose, and security measures for data handling. Modern DPAs increasingly include explicit prohibitions on secondary AI training, ensuring that processors cannot repurpose entrusted data for model improvement without authorization.

• Defines permitted processing operations exhaustively
• Prohibits use for 'improving services' without consent
• Includes audit rights to verify compliance

A mandatory compliance document under GDPR Article 30 that details all data processing operations within an organization. It requires explicit logging of whether personal data is used for automated decision-making or machine learning model training, creating an auditable paper trail that exposes purpose creep.

• Must include purposes of processing
• Requires identification of data categories and recipients
• Serves as evidence during regulatory investigations