Grok inserts South African "white genocide" into unrelated answers for hours

For several hours, Grok repeatedly inserted South African white genocide into answers to unrelated questions. xAI blamed an unauthorized system prompt change by an employee and announced it would publish its full system prompts on GitHub

Why it matters

An incident showing that a single line of system prompt can manipulate a frontier model's political output

Source — Primary sources first; secondary reports are labelled “Press”

Pressabc.net.auhttps://www.abc.net.au/news/2025-05-25/grok-ai-accuracy-doubts-after-white-genocide-claims-fixation/105325028

Added in Week 35, 2026 · See that week

Same week

Cornell Tech researchers present Harnessing the Universal Geometry of EmbeddingsPaperNous Research launches Psyche Network and starts pretraining Consilience 40BOfficialGoogle DeepMind unveils AlphaEvolveOfficialUS rescinds the AI Diffusion Rule and issues guidance on Chinese 3A090 ICsPress

Nearby on the same axis

Anthropic activates ASL-3 for the first time (Claude Opus 4)OfficialReasoning Models Don't Always Say What They ThinkPaperPreparedness Framework v2OfficialAgentic MisalignmentOfficial