Data Privacy

Data privacy is the right of individuals to control how their personal information is collected, used, shared, and retained, together with the laws and practices organizations follow to respect that right. In artificial intelligence the stakes are higher than in ordinary software: models are built from large collections of training data that often describe people, they can memorize and repeat that data, and systems such as facial recognition process personal information by design.

Laws and Core Principles#

The EU's General Data Protection Regulation (GDPR), in effect since May 2018, sets the reference principles in its Article 5: personal data must be processed lawfully, fairly, and transparently, collected for specified purposes (purpose limitation), limited to what is necessary (data minimization), kept accurate, stored no longer than needed (storage limitation), and protected with appropriate security, with the controller able to demonstrate compliance (accountability). Article 25 requires data protection by design and by default, and Article 22 gives people rights around decisions made solely by automated processing. Fines reach €20 million or 4% of worldwide annual turnover, whichever is higher.

Other laws follow similar patterns. California's CCPA, expanded by the CPRA, gives residents rights to know, delete, and opt out of the sale or sharing of their data. In US healthcare, HIPAA's de-identification standard lists 18 identifiers to remove under its Safe Harbor method, including full-face photographs and comparable images. The EU AI Act adds AI-specific rules, and since February 2025 it has banned building facial recognition databases through untargeted scraping of face images from the internet or CCTV footage.

Whether a dataset falls under these laws depends on whether people can be identified from it:

Anonymization is harder than it looks. Latanya Sweeney showed in 2000 that ZIP code, gender, and date of birth alone uniquely identify about 87% of the US population, which is why removing names rarely makes a dataset anonymous.

Privacy Risks Specific to AI#

Models can memorize and reproduce their training examples. Nicholas Carlini and colleagues extracted over a thousand training images from diffusion models, including photographs of individual people. Even without extraction, membership inference attacks can reveal whether a specific person's record was in a training set, which is itself sensitive for a model trained on patients with one disease.

Collection is the other major risk. Images and text gathered for one purpose are often reused for training without a lawful basis, and several European data protection authorities fined Clearview AI for building a face database from scraped photos, including a €30.5 million fine from the Dutch authority in 2024. Inside organizations, personal data pasted into external chatbots through shadow AI tools leaves the organization's control.

Privacy-Preserving Techniques#

• Differential privacy: Adds calibrated noise to statistics or training so that outputs reveal almost nothing about any one person.
• Federated learning: Trains on data where it lives, such as phones or hospital servers, and shares only model updates.
• Synthetic data: Replaces real records with generated ones that keep their statistical patterns.
• Confidential AI and edge AI: Process data inside attested hardware enclaves or on the device that captured it.

Redaction removes identifying content before storage or training. Computer vision pipelines blur faces and license plates in street imagery for smart city traffic analysis, and medical image analysis datasets strip patient identifiers from DICOM headers and remove facial features from head scans. On Ultralytics Platform, the Blur Faces action previews detected faces on sample images and then replaces their pixels in every image of a dataset, and a dataset can be set to blur faces automatically in every new upload.

Data security is one requirement within data privacy: encryption and access control keep personal data away from unauthorized parties, and a perfectly secured database can still violate privacy law if it holds data collected without a lawful basis or kept longer than its purpose requires.