Day 3: Shafi Goldwasser - Trustworthy AI: Robustness and Alignment | ADIA Lab Symposium 2025

Value of the Information & Strength of the Argument

The talk provides valuable insights into the formal foundations of trustworthy AI, particularly the application of cryptographic methods to machine learning. Goldwasser’s argumentation is rigorous, grounded in her own research and established theoretical frameworks. She clearly explains the limitations of current approaches and the need for provable guarantees. The discussion of backdoors and their undetectability is compelling, as is the proposed defense via post-processing. The argument that filtering is fundamentally flawed is well-illustrated with a clever construction. Overall, the value lies in bridging cryptography and AI, offering a fresh perspective on robustness and alignment.

Scientific Rigor, Source Quality, Title Accuracy

The talk is scientifically rigorous, referencing several published papers (e.g., on undetectable backdoors and oblivious defense) and the work of others (e.g., Boaz Barak on inference-time compute). However, the transcript does not include explicit citations or URLs, so the sources are not directly verifiable from the video. The title accurately reflects the content, focusing on robustness and alignment. The speaker’s authority and the technical depth contribute to high reliability, but the lack of detailed references in the talk itself slightly reduces the score.

The title accurately reflects the content: a focused talk on robustness and alignment in trustworthy AI, delivered by Shafi Goldwasser at the ADIA Lab Symposium.

8/10

Presentation by a Turing Award laureate with deep expertise in cryptography and AI security. The talk is based on published research, but it is a high-level overview without detailed proofs or citations in the transcript. The speaker clearly distinguishes between established results and open questions.

Markers derived by PSI from the transcript: the creator did not define chapters.

• 00:00Introduction by host, highlighting Shafi Goldwasser's credentials.
• 01:00Goldwasser poses critical questions about AI reliability and control.
• 03:00Overview of trustworthy AI challenges: privacy, verification, robustness, alignment, model stealing.
• 07:00Cryptographic recipe for trustworthy AI: model adversary, define trust, build solution, prove.
• 10:00Discussion of robustness to distribution shifts, adversarial examples, and insider adversaries.
• 12:00Presentation of undetectable backdoors in neural networks and their implications.
• 15:00Defense against backdoors via random self-reducibility and post-processing.
• 18:00Introduction to alignment and the challenge of formalizing human values.
• 20:00Discussion of jailbreaks and the inadequacy of simple filtering.
• 23:00Cryptographic impossibility result for filtering harmful prompts.

• Planting Undetectable Backdoors in Machine Learning Models — Discussed in the talk as a method to plant undetectable backdoors.
• Oblivious Defense in Machine Learning Models — Discussed as a defense against backdoors using post-processing.
• Puzzle Jailbreaking LLMs through Word Based Puzzles — Mentioned as an example of jailbreaking via puzzles.

• Turing Award — Shafi Goldwasser is a recipient, supporting her expertise.