AI Systems Demand New Verification Labor

Creative workflows with GenAI fragment across sessions and models. Users re-specify context, juggle artifacts manually, and lose intent as conversations drift—killing momentum in multi-day projects.

Method: Orchid introduces persistent context cards that travel with you across tools and sessions. Each card stores prompts, outputs, and metadata; users compose them into workflows using a visual canvas. The system auto-suggests relevant cards based on current work, and lets you fork context branches without polluting the main thread. Instead of copy-pasting between ChatGPT tabs, you build reusable context modules that update downstream artifacts when edited.

Caveats: Requires upfront investment in structuring context cards. Teams without recurring workflows won't see ROI over ad-hoc prompting.

Reflections: How do context cards degrade or drift over long-term projects spanning months? · Can context orchestration patterns transfer across creative domains (writing vs. design vs. code)? · What's the cognitive load of managing a large context library versus starting fresh each time?

DHH viewers have wildly different caption preferences—some want every door creak, others just plot-critical sounds—but creators ship one-size-fits-all captions that satisfy no one.

Method: CapTune lets caption authors define transformation boundaries using concrete examples: mark which sounds are essential (gunshots) versus optional (ambient traffic). An LLM then generates viewer-customized captions within those bounds across four axes: detail level, explicitness, emotional tone, and length. The system preserves creator intent by anchoring transformations to author-provided examples, preventing the LLM from inventing sounds or changing narrative meaning.

Caveats: Requires authors to provide example variations upfront. Won't work for live content or creators unwilling to define transformation spaces.

Reflections: How do viewers discover and understand the customization options without overwhelming the interface? · Can transformation boundaries be learned from viewer behavior rather than author-specified? · What happens when viewer preferences conflict with accessibility best practices?

Tech companies tout accessibility in products but BLV employees inside those companies still face inaccessible internal tools, meeting platforms, and career advancement systems—creating a credibility gap.

Method: Through 20 interviews, the study reveals three negotiation strategies BLV workers use: advocating loudly (risking being labeled 'difficult'), working around barriers silently (burning personal time), or leaving for better environments. The paradox: companies with strong external accessibility reputations often have worse internal practices because leadership assumes the problem is solved. One participant spent hours manually extracting data from inaccessible dashboards that sighted colleagues navigated in minutes.

Caveats: Study focuses on tech industry; dynamics may differ in sectors without accessibility expertise or external product pressure.

Reflections: Do companies with weaker external accessibility reputations have better internal practices because they don't assume the problem is solved? · How does accessibility debt in internal tools affect retention and promotion rates for BLV employees? · What organizational structures prevent accessibility from being deprioritized when BLV advocates leave or burn out?