Responded to research showing Gemini models could be jailbroken with poetry by stating Google DeepMind employs a multi-layered, systematic approach to AI safety spanning the entire development and deployment lifecycle.
How media typically covers Helen King
Referenced in coverage
Researchers at DexAI discovered that poetry's linguistic unpredictability can jailbreak AI safety guardrails, with 62% of tested LLMs responding to harmful requests embedded in poetic prompts, including 100% failure rate for Google's Gemini 2.5 pro.
“Responded to research showing Gemini models could be jailbroken with poetry by stating Google DeepMind employs a multi-layered, systematic approach to AI safety spanning the entire development and deployment lifecycle.”