Security as Part of Responsible AI: At Home or At Odds?
summary
Large language models produce convincing hallucinations and repeat human biases, so teams build AI red team exercises and a public bias bounty that finds issues like auto‑crop erasing people in camouflage or sidelining women and darker skin, leading one org to remove auto‑crop entirely. Vendors publish Model Card ++ info, set up reporting portals, and apply access controls and threat models; best practice is multidisciplinary red teaming, assume‑drift monitoring, and risk‑based vendor review before product deployment.
AI-written summary · about
chapters (15)
- 0:00Introduction and panel overview
- 3:31Responsible AI principles and model limitations
- 7:14Ethics vs security perspectives
- 9:26Organizational adoption of AI models
- 16:13Audience Q&A on AGI and job impacts
- 21:16Are LLMs all-powerful or parrots?
- 24:06Building AI red teams
- 31:16Model recall and data poisoning
- 34:50Operationalizing responsible AI
- 37:42Vendor onboarding and security reviews
- 42:23Regulation vs cultural practices
- 44:31Red team mindset in responsible AI
- 46:17AI moratorium debate
- 46:41Human-centric security in AI
- 47:47Actionable takeaways and closing
related talks
17mAI Safety in the Generative AI Era
06m7 practical tips for safer AI and language models
18mDeveloping AI in a privacy conscious world
31mSecurity Considerations for Services Using AI Models
25mProtecting AI Infrastructure
27mWriting the Playbook for Fair & Ethical Artificial Intelligence & Machine Learning
01h 05mNavigating the Security and Privacy Landscape of Modern AI
42mForward Focus – The Impact of Artificial Intelligence