AI is not a god that we can trust
summary
Claims that an off‑the‑shelf ML face‑match service bought off the shelf fails on Dutch driver’s licenses, misses obvious fraud, and forces manual fallback; root causes are wrong training data, unconfigured checks, and biased inputs. Practical prescriptions include falsify hypotheses, set numeric guards like a 95% threshold, add telemetry for model drift, and treat AI outputs as probabilistic with ongoing regression and conformity checks under emerging EU rules.
AI-written summary · about
chapters (7)
- 0:00Introduction
- 4:05AI implementation overview
- 10:55AI regulations and risks
- 12:26Data quality in AI
- 17:02Falsification in AI
- 21:15Reporting and regression testing
- 23:23Responsibility with AI
related talks
39mDefensible Application Security for the Artificial Intelligence Era
45mWhat Security Measures Are Necessary to Trust in AI?
38mTrustworthy AI: The Key to Unlocking High-Stakes Applications
54mAI: Actual Intelligence
52mAI Is Just Software, What Could Possibly Go Wrong?
01h 05mNavigating the Security and Privacy Landscape of Modern AI
57mAI Applications Security Puzzle
31mSecurity Considerations for Services Using AI Models