https://www.jasonslucas.com/tag/socio-technical-ai/Socio-Technical AIHugoBlox Kit (https://hugoblox.com)en-usMon, 13 Jul 2026 00:00:00 +0000https://www.jasonslucas.com/media/icon_hu_1b2044c02ce09a43.pnghttps://www.jasonslucas.com/tag/socio-technical-ai/https://www.jasonslucas.com/publication/conference-paper-dual-curse-multilingual/Mon, 13 Jul 2026 00:00:00 +0000https://www.jasonslucas.com/publication/conference-paper-dual-curse-multilingual/<p><strong>Position:</strong> Multilingual AI safety cannot be retrofitted through post-hoc alignment. We identify a <strong>dual curse</strong> in current systems and argue for socio-technical guardrails built in from pre-training.</p> <p>Key contributions:</p> <ul> <li><strong>Dual Curse documented</strong>: Harmful content generation rises to <strong>35% in low-resource languages</strong> (vs. 1% in English), while instruction-following capability declines sharply across the same languages.</li> <li><strong>Systematic review of 207 studies</strong>: Reward models achieve only <strong>49–50% accuracy in low-resource languages</strong> — equivalent to random chance — undermining post-deployment safety pipelines.</li> <li><strong>Socio-technical prescription</strong>: Pre-training interventions, <strong>community-led harm specification</strong>, and multilingual evaluation metrics that balance security and usability jointly, rather than trading one off for the other.</li> <li><strong>Call to action</strong>: Treat multilingual safety as a first-class design constraint, not a downstream patch applied through RLHF or filtering.</li> </ul> <p>Resources:</p> <ul> <li> </li> </ul>