arrow_backAll episodes
EP 57
Do You Trust Your AI?
The podcast explores the concept of trust in AI, particularly in the field of medicine and drug discovery. It argues that asking if we trust AI is the wrong question and instead proposes the idea of calibrated reliance. The episode delves into the complexities of AI in biology, highlighting the dangers of data leakage, hallucinated explanations, and the importance of a six-layer trust framework.
AIdrug discoverycalibrated reliancedata leakagehallucinated explanations
Key Concepts
Calibrated Reliance
- Depends on the context of use
- Not a blanket statement
- Replaces the concept of trust
Data Leakage
- Occurs when a model is trained on one set of data and tested on another
- Can lead to overfitting and poor performance on new data
- Can be mitigated with advanced mathematical tools like data sale
Hallucinated Explanations
- Can create false reassurance through automation bias
- Can be dangerous if not verified through orthogonal testing
- Can lead to over-reliance on AI-generated explanations
Six-Layer Trust Framework
- Includes data trust, model trust, biological trust, medicinal chemistry trust, workflow trust, and governance trust
- Provides a comprehensive approach to rebuilding a system's credibility
- Is essential for ensuring the reliability of AI-driven drug discovery
Trust Ledger
- A proposed solution to tie everything together
- Provides a comprehensive, queryable database
- Includes the exact provenance of the training data, mathematical limits of the AI's confidence, and arguments and counterarguments generated by critic agents
Episode Summary
- check_circleThe concept of trust in AI is too vague and should be replaced with calibrated reliance
- check_circleCalibrated reliance depends on the context of use and is not a blanket statement
- check_circleThe scientific community needs to abandon the idea of trust and adopt a more nuanced approach
- check_circleData leakage and hallucinated explanations are significant problems in AI-driven drug discovery
- check_circleA six-layer trust framework is proposed to rebuild a system's credibility
- check_circleThe framework includes data trust, model trust, biological trust, medicinal chemistry trust, workflow trust, and governance trust
- check_circleThe trust ledger is a proposed solution to tie everything together and provide a comprehensive, queryable database
- check_circleThe future of life-saving drug discovery may rely on autonomous AI agents acting as a mixture of critics
Full Transcript
Discussion
Join the discussion — sign in to leave a comment.
Log in to comment