liveBook · Manning

Personal data

As explained in AI Governance

Information that can identify a specific person, either by itself or when combined with other information.

Personal data is information that can identify a person, alone or in combination with other information. In a GenAI system, it can travel beyond its original purpose: a claim conversation may become training material or stored memory, then reappear in an unrelated interaction. A model can also generate a false statement about a real person, creating personal data that may affect how that person is treated. section 5.0

Map personal data across prompts, logs, retrieval indexes, fine-tuning data, model or adapter parameters, outputs, vendor systems, and backups. A privacy failure may be an answer that repeats a private detail, infers something sensitive, or invents a claim—not just unauthorized database access. section 5.0

Use four pillars to turn the data map into decisions and controls:

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

Deployment changes the available controls. In a SaaS application, use supported redaction or anonymization, consider data loss prevention, minimize sensitive content in vendor-hosted retrieval, and secure vendor commitments against training or secondary use. With a vendor API, a central gateway can sanitize traffic before it leaves the organization. Local retrieval still needs protection, minimization, and legal assessment. section 5.2.4

Hosting and fine-tuning offer more control over training data but also more responsibility. Exclude data the model does not need; for data included, document the purpose and lawful basis and plan for user rights. Deleting its influence from model weights may require retraining. Retrieval documents and embeddings can usually be updated or deleted without retraining, although the index needs ongoing control as documents change. section 5.2.4 section 1.1

Monitor an agent’s data access

Record the data an agent accesses, tagged when personal or sensitive, along with external calls, autonomous decisions affecting individuals, unauthorized access attempts, errors, and overrides. Alert on unexpected behavior, such as unusually large database pulls or a spike in attempts to use disallowed sources after an update; investigate or roll back when needed. section 5.6.4

How AI Governance covers it

How deep it goes
Details deployment-specific controls and agent monitoring, including deletion trade-offs between retrieval indexes and fine-tuned models.
Examples
Uses an insurance claim chatbot and agent audit events to illustrate unexpected retention, fabricated claims, and monitoring.
What you can do after
Apply privacy controls across GenAI deployment models and monitor agent access to personal data.
How it connects
Connects personal data to data minimization, retrieval-augmented generation, fine-tuning, model memorization, and deletion rights.

Also explained in

Compare how 13 books explain personal data →

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.