Systems Detect Problems They Won't Enforce

Sign language translation systems treat translation as a technical mapping problem, ignoring how deaf creators actually navigate multilingualism, cultural context, and political stakes when producing signed content.

Method: Interviews with 13 deaf Chinese content creators reveal translation as (trans)languaging work: creators fluidly blend Chinese Sign Language, written Chinese, and regional sign variants while managing cultural meaning-making for mixed deaf/hearing audiences. They explicitly navigate political choices—which sign variants to use, how to represent deaf culture, when to code-switch—that no current translation system addresses.

Caveats: Chinese context with specific linguistic politics. Generalizability to other sign language communities unverified.

Reflections: How would translation systems support creator control over code-switching and variant selection rather than enforcing standardization? · What interface affordances enable deaf creators to signal cultural context that hearing audiences need without flattening signed expression? · Can systems surface the political stakes of translation choices to hearing users without burdening deaf creators with explanation labor?

Rapid eating is hard to self-regulate because people don't notice pace changes in real-time and sustained self-monitoring during meals is cognitively taxing.

Method: Commodity earbuds repurposed to detect chewing via bone-conduction sensors achieve F1 = 0.97 for chewing detection and MAE of 0.18 ± 0.13 chews/min for pace estimation, enabling closed-loop just-in-time interventions. In a 13-day field study (N=14), the system significantly increased chews per swallow and reduced consumption speed, with statistical signs of carryover effects on retention-probe days when the system was off.

Caveats: Fourteen participants over 13 days. Longer-term adherence and broader dietary contexts untested.

Reflections: Do carryover effects persist beyond the study period, or do users revert once the system is removed permanently? · How does intervention effectiveness vary across meal types (snacks vs. formal meals) and social contexts (alone vs. dining with others)? · Can the system adapt intervention timing based on learned user pace patterns rather than fixed thresholds?

As AI delivers more educational feedback, learners' beliefs about whether feedback comes from AI or humans may shape engagement—but content and source are usually confounded.

Method: Identical LLM-generated feedback attributed to either AI or a human TA produced dramatically different engagement: TA-attributed feedback drove significantly more time and effort (d = 0.88-1.56). Evaluation drivers also differed—AI ratings correlated with prior AI trust (r = 0.85), while TA ratings correlated with perceived genuineness (r = 0.65).

Caveats: Creative coding context with single-session engagement. Long-term learning outcomes and other domains untested.

Reflections: Does the engagement gap persist over multiple feedback cycles, or do learners habituate to AI-attributed feedback? · What specific feedback features drive perceived genuineness for AI vs. human sources? · In domains requiring iterative revision (writing, design), does lower initial engagement with AI feedback compound into worse learning outcomes?