A school-yearbook plate set pairs stills with sloppy captions that praise groups unevenly. A later CLIP-style generator repeats the slant. What should the candidate do?
Select an answer to reveal the explanation.
Short Explanation
If the flash cards praise one group and shrug at another, the generator will copy that tone. Fix the paired text before you grow more yearbook stills.
Full Explanation
A school-yearbook plate set that pairs stills with sloppy captions praising groups unevenly, when a later CLIP-style generator repeats the slant, should have the biased text arm flagged as a trustworthy-build defect to fix before more generation. Biased paired captions poison the text arm and reproduce in CLIP-style generation.
Treating it as a Domain 6 speed issue and adding more GPUs fails: compute does not remove caption bias. Ignoring the captions because only pixels matter for generation fails multimodal reality—text steers CLIP-style paths. Shipping the generator faster so the bias “averages out” online fails and spreads harm.
Exam caveat: biased text arms are Domain 7 trustworthy-build defects. Fix captions/pairs before further synthesis; do not outrun the problem with GPUs.
Operational check: audit yearbook captions for uneven praise and stereotypes; rewrite or rebalance pairs; freeze generation until the civic review clears the text arm; re-test generator outputs for repeated slant. In municipal trustworthy-AI stewardship, record the disclosure, consent, or bias decision and the rejected shortcuts in the gallery change log so future stewards protect residents from undisclosed synthesis and stereotyped generation. Re-audit labels, consent records, and bias samples after the next model or exhibit update to keep the city's public trust intact.