Generative AI Leader Practice Question: Required by the EU AI Act to ensure high-risk AI…

Answer analysis

Option-by-option breakdown

For each option: why learners choose it and why it is or isn't the right answer here.

• ✗

  Implement custom safety filters on the model outputs

  Why it's wrong here

  Output filters only police generated content; they neither document the vendor model's training data, limitations and evaluation metrics nor expose its decision logic, which the EU AI Act's transparency and auditability duties demand. Filters are tempting because they mitigate harmful outputs in deployment, but model documentation is the actual requirement here.

• ✓

  Ask the vendor to provide a Model Card and datasheets for the training data

  Why this is correct

  Model Cards and training datasheets document intended use, evaluation results, limitations and data provenance, giving the deployer the evidence needed for transparency and audit obligations under the EU AI Act. Vendor contracts or accuracy benchmarks alone cannot demonstrate the system's design, risks and training data characteristics.

• ✗

  Use a larger context window to capture all interactions

  Why it's wrong here

  A larger context window only lets the model retain more tokens per request; it produces no documentation of the vendor model's training data, evaluation results or limitations, so it cannot satisfy the Act's transparency and auditability obligations. Context windows are tempting for handling long inputs, not for regulatory disclosure.

• ✗

  Train an internal model from scratch to replace the vendor model

  Why it's wrong here

  Training an internal model replaces the vendor but does not itself yield the training-data provenance, evaluation metrics and limitation disclosures the Act requires; the obligation attaches to documenting the high-risk system, not to who owns the weights. Building from scratch is tempting for control, yet it is unnecessary and far costlier.

Go deeper

Related to this question