An open-data team plans to enable image upload in a multimodal civic app. What should they run before expanding that feature?
Select an answer to reveal the explanation.
Short Explanation
Turning on image upload is like opening a new public drop box—you check what can go wrong before you unlock it. Run safety evaluations on multimodal inputs first so the civic app is ready for messy, risky uploads.
Full Explanation
Safety evaluations covering multimodal uploads should run before the open-data team expands image upload in a civic app. Multimodal inputs introduce risks that text-only tests miss: offensive imagery, personal data in photos, manipulated documents, prompt-injection via embedded text, and policy-violating content that could surface in public portals. Completing structured safety evaluation—adversarial image sets, classifier/filter coverage, retention and scanning behavior—before enabling the feature reduces the chance that production residents become the first test harness.
Safety evaluations covering single-modality uploads before enabling the new feature test text or images in isolation, not their combined multimodal interaction, so they miss cross-modal risks like prompt injection via embedded text, offensive imagery contextualized by captions, and personal data in photos that must be evaluated before expanding uploads. Safety evaluations covering multimodal upload latency before enabling the new feature measure upload speed and performance, not the safety of user-supplied uploads, so they miss offensive imagery, personal data exposure, and manipulated documents and cannot substitute for multimodal upload safety evaluation. Safety evaluations covering stored upload logs before enabling the new feature review retrospective records of past uploads, which is too narrow and after-the-fact to cover proactive input risks like offensive imagery, personal data in photos, and manipulated documents, so they cannot replace full multimodal upload safety evaluation.
Exam caveat: "safety evaluation" for multimodal apps includes input safety and output handling, not only model accuracy metrics. Operational check: require a documented multimodal safety test pack with pass/fail criteria and evidence that upload paths invoke scanning and policy rejection before the feature flag is enabled for the public open-data app.
Prefer to study offline? Printable PDF study book + test engine ($7.50, unofficial). These questions stay free here.