AI-900 Practice Question: Describe features of generative AI workloads on Azure

Answer analysis

Option-by-option breakdown

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

• ✓

  Computer Vision: Ensures AI systems treat all people fairly, avoiding bias in image analysis.

  Why this is correct

  Computer vision workloads must operationalize the fairness principle because models can produce disparate error rates across demographic groups (e.g., skin tone, gender, age) in facial recognition, object detection, or image classification. Fairness in this context means auditing training-data composition, measuring performance trade-offs across subgroups, and mitigating harmful bias before deployment to ensure equitable treatment.

• ✓

  Natural Language Processing: Designs AI to support diverse languages and cultures.

  Why this is correct

  Natural language processing workloads apply the inclusiveness principle by accounting for linguistic diversity—not just high-resource languages like English, but also dialects, code-switching, scripts, and cultural idioms. Through inclusive design, NLP systems avoid imposing a one-size-fits-all linguistic model and instead serve a broad range of users, including low-resource language communities, without cultural assumptions.

• ✓

  Speech: Protects user privacy by safeguarding voice data.

  Why this is correct

  Speech workloads handle sensitive biometric data—voiceprints and audio recordings—which falls directly under the privacy and security principle. Safeguards such as data minimization, encryption in transit and at rest, secure consent workflows, and de-identification are necessary to protect user voice data from unauthorized access or re-identification. Without these, speech systems create serious privacy risks distinct from other AI workloads.

• ✓

  Decision: Provides clear explanations of how AI arrives at conclusions.

  Why this is correct

  Decision-making AI systems (for example, credit scoring, hiring, or medical triage) must uphold the transparency principle by providing clear explanations of how inputs lead to outcomes. This can require interpretable model architectures, post-hoc explanation methods like SHAP or LIME, and documentation such as model cards to enable user understanding, auditing, and accountability. The goal is to move beyond opaque 'black-box' decisions.

• ✗

  Computer Vision: Protects user privacy by safeguarding voice data.

  Why it's wrong here

  Pairing Computer Vision with 'protects user privacy by safeguarding voice data' misassigns the privacy principle because voice/audio is not the primary data modality of computer vision. Computer vision processes images and video, where privacy concerns revolve around facial data and visual surveillance—not voice. This description is actually a better match for Speech workloads, which capture and transmit sensitive audio, so the option is incorrect for this match.

• ✗

  Speech: Ensures AI systems treat all people fairly, avoiding bias in image analysis.

  Why it's wrong here

  This option is wrong because fairness in image analysis is the domain of Computer Vision, not Speech; speech workloads deal with audio features, phonetics, and language, rather than biased pixel-level classification. If fairness were applied to Speech, it would focus on accent, dialect, or speaker variability in speech recognition, not on demographic bias in image outputs. Therefore, the description does not belong to a Speech workload.

Go deeper

Related to this question