Code-switching is a widespread linguistic practice among bilingual speakers. While recent studies have addressed the impact of code-switching on downstream task performance, the potential biases and harms that language models may cause when prompted with code-switching have yet to be investigated. The objective of this study is to investigate whether code-switching constitutes an implicit indicator of ethnicity that can be leveraged to unveil covert racist or xenophobic bias in language models. The present paper introduces a methodology to compare generated texts that were prompted with code-switching vs. with monolingual inputs. It is applied on both Hinglish and Spanglish, two popular forms of code-switching that are omnipresent in Indian and Hispanic communities. With a decision tree approach, we tackle various types of semantic differences through the use of semantic resources, stereotypes lists, POS-tagging and sentiment classifiers. Over 84k text pairs are generated with 3 popular large language models. Overall, around 50% of generated text pairs are not semantically equivalent, and 25% of the time, there is a potential for harm against the Indian or Hispanic community. The different possible harms are further discussed, relying on sociological studies to argue that bias and harms against socially discriminated communities have greater consequences.

Large-scale visual search platforms maintain production text-image embedding models trained on behaviorally-mined query-image pairs, where the text encoder is implicitly tuned to organisation-specific search vocabulary. Extending such a system to video retrieval is non-trivial: retraining risks disrupting hard-won query understanding, while naive temporal aggregation discards the sequential structure that distinguishes a door opening from a door closing. We add a lightweight temporal attention pooling head over independently encoded frame embeddings, keeping both encoders frozen and training only the pooling head and positional encodings. We train on a large-scale dataset from a commercial creative platform using bidirectional InfoNCE with a two-stage hard negative mining pipeline. Benchmarking against LoRA fine-tuned Gemma 3 (12B) and SmolVLM2, neither matches our approach despite substantially larger parameter counts. We also introduce an evaluation benchmark with graded relevance judgments annotated by a VLM judge, where each query is mapped to all relevant videos in a large corpus rather than a single ground-truth clip, avoiding protocols that greatly underestimate model quality. Our model improves nDCG@1 by 10.1% over mean-pooling overall and 25% on complex action queries. With Matryoshka Representation Learning, 256-dim embeddings retain 95.6% of full performance, and int8 quantization costs under 1% in retrieval quality.