liveBook · Manning

Right of access

As explained in AI Governance

A data-protection right to learn whether one’s personal data is being processed and to obtain information about or a copy of that data.

The right of access lets a person ask whether an organization is processing their personal data and obtain a copy or description of it and how it is used. In generative AI systems, the request can reach beyond conventional records to information ingested or produced by a model. Publicly sourced personal data is still personal data under the GDPR. section 5.5.2

An insurance claim request

After ClaimAssist invents a collision in a customer’s history, she asks to see her data. The team can export her CRM record and claims file, but also finds embeddings linked to her customer ID and accident details that influenced model weights. The vectors are not human-readable, and the learned influence is not a database row that can simply be retrieved. section 5.5 section 5.0

The practical problem is to find and describe personal data across representations. Prompts, logs, retrieval indexes, fine-tuning data, parameters, stored outputs, vendor systems, and backups can all be relevant locations. Teams should track, at least broadly, what personal data enters and leaves a system, including training sources and data types. Confirming whether a particular person’s data was in a large training mix can be difficult. section 5.5.2 section 5.0

The response workflow depends on where the organization has control:

  • For SaaS, select a provider with suitable export capabilities and verify retention settings and audit logs. Explain what the service keeps and what it can and cannot provide. section 5.5.4
  • For an API integration, search the organization’s own prompt logs, stored outputs, retrieval documents, and embeddings. Explain what the external provider receives and may retain under contract. section 5.5.4
  • For an in-house model, plan access into the architecture and maintain training-data provenance records. Visibility and control bring direct accountability, but do not make learned data easy to recover. section 5.5.4 section 5.5

Using a third-party model does not remove the deployer’s responsibility for access requests concerning customer data it uses, including data supplied for fine-tuning. Contracts or cooperation with vendors should support the response. Access is distinct from erasure: finding or describing data does not by itself remove its influence from a model. section 5.5.2 section 5.5

How AI Governance covers it

How deep it goes
Explains request handling across deployment models and identifies barriers posed by embeddings and model weights.
Examples
Uses an insurance chatbot’s fabricated collision and the customer’s request for records to expose access challenges.
What you can do after
Operate an access-request workflow by locating relevant data and coordinating with model providers.
How it connects
Connects data subject access requests with transparency, retrieval-augmented generation, model weights, and erasure.

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.