liveBook · Manning

Sensitive information

As explained in Introduction to Generative AI, Second Edition

Proprietary data is information owned or controlled by an organization and not generally available to the public.

A contract pasted into a chatbot may contain more than words to revise. It may reveal a client’s plans, obligations, or negotiations. Sensitive information is data whose exposure could harm someone or violate a duty of confidentiality. It includes personal data and confidential or proprietary information—data an organization controls that is not generally public. Not all sensitive information is personally identifiable information (PII). section 3.10 section 3.20

Large language models (LLMs) can encounter sensitive information in two ways: through data used to train them and through prompts people enter while using them. A model may reproduce text encountered during training; conversations may also be retained or reviewed, used to improve a service, or inadvertently exposed. These are risks, not a claim that every private prompt will appear in another user’s response. section 10.18 section 3.11

A contract-revision prompt

An attorney asks a dialogue agent to improve a draft contract. The result may save time, but entering client details could disclose confidential information to the provider and, without client consent, may jeopardize attorney-client privilege. The useful output does not erase the risk created by the input. section 3.10

Practical precautions focus on what enters the system and how it is handled:

  • Do not enter client information, internal documents, proprietary code, or other data you would not want shared without appropriate safeguards. section 3.12
  • Set workplace policies, assess tools in a closed environment, and check privacy disclosures, consent, deletion, and opt-out options. section 3.12
  • Reduce sensitive data in training sources where possible. Anonymization and masking can help, but neither guarantees that all personal information is found or safely removed. section 10.18

Privacy depends on technical measures, provider practices, user choices, and governance. Even public posts can raise concerns when people did not expect them to train AI systems; disclosure alone may not ensure that users understand or accept that use. section 3.0 section 10.9

How Introduction to Generative AI, Second Edition covers it

How deep it goes
Explains exposure pathways and practical precautions at a high level, without implementation steps for privacy protections.
Examples
Uses a client contract entered into a chatbot, retained conversations, and personal details reproduced from training data.
What it assumes
Defines the privacy risk through familiar chatbot use and explains personally identifiable information without requiring technical knowledge of LLM internals.
What you can do after
Practice/perform cautious LLM use by limiting confidential inputs and checking workplace rules, privacy disclosures, and user choices.
How it connects
Connects sensitive information to personally identifiable information, data leakage, training data, consent, and data governance.

Also explained in

Compare how 8 books explain sensitive information →

Related concepts

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