'Context misfit' names three recurring generative-AI output-mismatch dimensions, and SMB workers repaired them with a five-part repertoire

generative-ai-at-workcscwlocalizationrepresentationkenyanigeriasmall-businesssingle-source-cap

Abdulhamid, Nyairo, Ankrah & O'Neill (arXiv:2610.31344, 2026) coin context misfit for recurring generative-AI output mismatches along three dimensions: local institutions and place (policies, regulations, market norms), language practices (multilingual and code-mixed communication), and representational fidelity in generated language and media. The paper is explicit that this is a mechanistic contribution documenting how known categories of geographic, linguistic, and representational bias surface as everyday breakdown-and-repair in client-facing small-business work, not a claim to have discovered those bias categories.

Concrete instances reported from the interviews: a text-to-speech tool that could pronounce "Nairobi," "London," and "Everest" but could not localise or verbalise other local names, which led one outdoor-recreation CEO to abandon the tool entirely; a legal associate's team exercise in which AI-generated quoted law contained "a lot of red flags," which the associate concluded "destroy[ed] my confidence in the platform" (the bracketed "[ed]" is the paper's own); and an ICT CEO testing voice cloning who got "a British accent or... an American accent" and wanted "real African voices" instead. Against these failures, participants used a five-part repair repertoire: rewriting for voice, constraint layering (naming exact local destinations or specifics in the prompt), substituting local sources and situated knowledge, switching tools, and disengaging — abandoning a capability outright once repair cost exceeded its value. The study draws its whole sample from 21 Kenyan and Nigerian SMBs, interviewed July–October 2023, and frames the pattern as the labour that turns bare tool access into workable participation, not a new finding about bias existing.