liveBook · Manning

Data overcollection

As explained in AI Governance

Overcollection is the collection or retention of more personal data than is necessary for a defined purpose.

Data overcollection means collecting or retaining more personal data than a defined task requires. A lawful basis for processing is only a starting point: data minimization under GDPR Article 5(1)(c) limits information to what is adequate, relevant, and necessary. section 5.2.2

Check where unnecessary data can enter a GenAI workflow:

  • Training and fine-tuning pipelines may ingest entire web scrapes, repositories, or user logs on the assumption that more data is better. section 5.2.2
  • A RAG chatbot answering policy questions may index sensitive HR files when a small set of de-identified policy documents would suffice. section 5.2.2
  • Prompts, outputs, and telemetry retained for debugging can become additional data sources without a justified retention period. section 5.2.2

Anonymization is not a guarantee

Replacing usernames with numeric IDs did not prevent people from being identified in published AOL search logs. Detailed queries could be correlated to reveal individuals, illustrating why removing direct identifiers alone may not make data anonymous. section 5.2.2

How AI Governance covers it

How deep it goes
Applies data minimization to training, fine-tuning, RAG, and telemetry, and warns that removing direct identifiers may not prevent re-identification.
Examples
Contrasts excessive HR-drive indexing with a limited policy corpus, and recounts re-identification from AOL search logs.
What it assumes
Assumes a lawful basis for processing but teaches that it does not justify collecting all available data.
What you can do after
Evaluate whether training, retrieval, or logging pipelines collect and retain data beyond their stated need, and treat anonymization as fallible.
How it connects
Connects data minimization under GDPR Article 5(1)(c) with anonymization, RAG, model training, and fine-tuning.

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.