liveBook · Manning

Accuracy

As explained in AI Governance

Accuracy is the data-protection principle that personal data be correct and, where necessary, kept up to date.

A GenAI system that invents a damaging statement about a real person creates more than a quality problem: it may process inaccurate personal data. Accuracy is the data-protection principle that personal data should be correct and, where necessary, kept up to date. In GenAI, applying it requires attention to both stored information and generated outputs. section 5.7

Correction is harder when personal data is embedded in model weights, embeddings, caches, or backups. Prefer keeping important knowledge in external locations that can be corrected or deleted, and record where data goes. Avoid putting sensitive or frequently updated personal data into model weights without a strong reason. These architectural choices reduce the need for difficult corrections or deletions; they do not fully satisfy deletion obligations. section 8.2 section 5.7

Apply accuracy controls according to the consequences of an error:

  • Narrow the task and ground answers in approved material. Send calculations and fixed rules to traditional software, and show evidence before a person approves a result. section 8.2
  • Stop or escalate when sources are missing or conflicting. A real citation does not guarantee a correct conclusion, and the cited material may itself be outdated or incomplete. section 8.2
  • Define accuracy metrics and thresholds for the use case. Monitor reports and increase human oversight when errors cross a threshold, including for a subset of users. section 6.1.1
  • Log hallucination reports with the IDs of top-ranked retrieval documents and record remediation steps, such as document updates or model-setting changes. section 6.1.1

Match checks to the stakes

An internal summary can be checked and corrected, while a medical recommendation or automated claim denial may leave less room to recover from a confident mistake. For higher-stakes outputs, compare answers with reference documents or route suspect answers for human review. A second model can review answers, but it adds inference cost and can also be wrong. section 8.2 section 6.1.1

Accuracy can change when a provider updates a model or an application switches models. Require documented accuracy metrics, measurement methods, and known failure modes; negotiate version pinning and advance notice so changes can be revalidated before deployment. section 8.2 section 7.3.5

How AI Governance covers it

How deep it goes
Explains architectural safeguards, escalation, measurement, and monitoring while noting that grounding and citations cannot guarantee correct conclusions.
Examples
Uses fabricated statements about real people, internal summaries, medical recommendations, and automated claim denials to illustrate accuracy risks.
What you can do after
Evaluate accuracy risks and apply controls for correction, escalation, measurement, and human review.
How it connects
Connects accuracy with hallucinations, data protection, retrieval-augmented generation, human oversight, and model versioning.

Also explained in

Compare how 49 books explain accuracy →

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.