liveBook · Manning

Privacy

As explained in AI Governance

The protection and appropriate handling of information about identifiable people, including how it is collected, used, and disclosed.

Privacy protects information about identifiable people and governs how it is collected, used, stored, and disclosed. In generative AI, personal data may persist in prompts, logs, retrieval indexes, training data, model parameters, outputs, vendor systems, and backups. A system can also create privacy harms by repeating private details, inferring sensitive traits, or fabricating claims about a real person. section 5.0

A four-pillar framework organizes privacy controls:

  • Collection and purpose: decide what enters the system and whether its use is lawful and expected. section 5.0
  • Storage and memorization: track where personal data persists, including retrieval indexes and model parameters. section 5.0
  • Output integrity: account for false or harmful statements about real people. section 5.0
  • User rights and governance: plan for access, correction, and deletion when data has been copied, transformed, or learned. section 5.0

Inference from training data

Membership inference probes whether a person's data appeared in training; attribute inference tries to predict hidden traits from text. A 2024 study found that anonymization barely reduced some demographic predictions, showing that writing style and other subtle cues can expose personal information. section 5.3.1 section 5.3.1

Before training, teams can filter sensitive material, remove duplicates, and limit access. These measures address training-data risks, but privacy review also needs to consider persistence, outputs, and user rights. section 5.3.1 section 5.0

How AI Governance covers it

How deep it goes
Explains four protection pillars and data pathways, then details inference risks and briefly names training-data hygiene measures.
Examples
Uses a claim chatbot that exposes old personal details and examples of membership and attribute inference from text.
What you can do after
Evaluate privacy exposure across collection, persistence, outputs, and inference, and apply filtering, deduplication, and access limits before training.
How it connects
Connects data protection to training data, retrieval indexes, model parameters, output integrity, and user rights.

Also explained in

Compare how 26 books explain privacy →

Related concepts

This explainer was generated by Manning's knowledge graph from the text of AI Governance. The section links point to the parts of the book it draws on.