Third-party cyber evaluations involving OpenAI models
openai.com·13h ago
AI Research
arXiv CS.AI·Jun 12·5 min read
Position: Generative Engine Optimization Creates Underexamined Risks, Governance Must Target Concentration, Disclosure, and Academic Blind Spots
TL;DR
Generative Engine Optimization (GEO) poses risks related to concentrated influence and undisclosed commercial biases in large language model (LLM) outputs. A general GEO pipeline was developed to analyze these risks and compare practices in academia and industry.
✦ Why It Matters
Engineers and researchers should implement stronger governance frameworks to mitigate risks in LLM outputs.
Key Takeaways
- Enhance contestability in LLM outputs to reduce concentrated influence.
- Implement precise disclosure practices for commercial influences in evidence.
- Conduct black-box audits to assess material influence on LLM outputs.
- Align evaluation metrics with real-world deployment scenarios.
How It Works
GEO optimizes LLM outputs by targeting the evidence pool and generation processes, shifting the focus from traditional search engine optimization to a more integrated approach that synthesizes information.