Automating Privacy Reviews with LLMs: Challenges & Opportunities
summary
Experiments show LLMs classifying privacy reviews with false negatives comparable to a junior reviewer and summarizing documents to cut reviewer words by about 85%, but they produce many false positives and hallucinate when documentation or diagrams are missing. Combining static code‑scanning graphs with LLMs raises data‑flow accuracy to roughly 95% after AI filtering. Practical limits include multimodal diagram parsing, on‑prem deployment needs for IP protection, and target costs like running models on a single GPU (~$10k/year).
AI-written summary · about
chapters (7)
- 0:00Automating privacy with LLMs
- 3:56LLMs for review summarization
- 7:29LLMs as classifiers
- 10:59LLMs in code scanning
- 15:59LLM adoption considerations
- 22:46Future of automated reviews
- 27:29Closing on AI and privacy
related talks
16mAutomating Technical Privacy Reviews Using LLMs
29mGenerative-AI: Usecase-Specific Evaluation of LLM-powered Applications
44mProduction-Grade LLMs & AI Agents
28mTrustworthy LLMs: Vibe checks are not all you need
06mData Privacy Concerns and Solutions When using LLMs in Production
56mDehumanizing Agents: Why Explainability is Crucial in the LLM Era
38mNobody Pages the LLM: Engineering Rigour for Vibe Coding
38mMaking LLMs reliable - A practical framework