Building Around the Bias: How to Engineer Culturally Intelligent AI
Mainstream LLMs are trained on largely monocultural data. The result: they quietly misread how real people actually communicate, code-switching, mixed languages, cultural context.
In this first session of our AI Ethics Speaker Series, we look at why these silent failures happen and how to engineer around them.
🎤 Speaker: Oyinkansola Onwuchekwa, AI Research Engineer, University of Hull
📅 Date: 9th July, 18:00 CEST
💻 Format: Live on Zoom, 45 minutes
🎟️ Ticket: Free
📌 What You'll Learn
Identifying LLM failure modes: how standard tokenisers and commercial models misinterpret fluid code-switching and cultural context, and why this creates silent failures in production.
Architectural safeguards: prompting strategies and system design approaches to handle diverse linguistic input without a fine-tuning budget.
Low-overhead evaluation: simple techniques to build targeted validation checks that stress-test your system against linguistic bias before deployment.
📌 Why This Matters
Mainstream AI models are trained on largely monocultural data, causing them to systematically misinterpret real-world multilingual communication. For developers, ignoring these gaps means building brittle systems that fail the users they are meant to serve. Engineering around these biases keeps systems accurate and reliable across the full range of people using them.
📌 Who Should Join
AI engineers, data practitioners, conversational designers, and product managers building or deploying multilingual systems with commercial LLMs.
📌 How to Register
This event is free. Hit "Register" on this Luma page to save your spot. You'll get the Zoom link and a calendar invite by email, plus a reminder before we go live.