My path into AI safety wasn’t planned. I’ve worked as a writer, teacher, crisis-intervention counselor, ski instructor, and content strategist. Each of those roles shaped how I think about language, human behavior, and the ways people interact with technology.
When I began evaluating AI systems, I realized how much of that background fit the work. Safety evaluation isn’t just technical—it’s about judgment, pattern recognition, communication, and understanding why people say the things they say.
For the past two years, I’ve tested large language models for vulnerabilities, reviewed responses across sensitive harm categories, and written clear rationales explaining classification decisions. I’ve also created adversarial prompts to test boundaries, designed RAG confusion prompts to expose hallucination patterns, and reviewed multi-step agentic workflows for compliance and accuracy.
What I bring to this work is a human-centered perspective:
- From crisis intervention, an ability to recognize manipulation and maintain calm judgment.
- From writing, a strong sense of language, tone, and narrative patterns—useful for understanding how prompts function and where they hide risk.
- From cross-cultural experience, an awareness of how AI systems affect people with different backgrounds, abilities, and communication styles.
- From life outside tech, an appreciation for how non-specialist users actually talk, think, and make decisions.
I enjoy evaluating ambiguous cases, understanding where the line is between safe and unsafe behavior, and helping improve systems so they better serve real-world users.
This portfolio highlights my writing background, which continues to inform how I approach AI safety and model evaluation. If you’d like to discuss my safety evaluation work or methodology, feel free to reach out.
Kalispell, Montana • brian.olsonwriter@gmail.com
LinkedIn: linkedin.com/in/brian-olson-b527408
Brian Olson 