AIF-C01 notes
AI use cases and applications

Challenges of Generative AI

A model can make decisions that are unethical or socially irresponsible, so each challenge needs a mitigation.

ChallengeRiskMitigation
Regulatory violationsOutput exposes regulated data such as PIIAnonymize data, use privacy-preserving training, audit training data
Social risksContent damages reputation or harms societyTest outputs, set content policies, monitor use
Data security and privacySensitive data leaks through prompts or outputsEncrypt data, control access, filter inputs and outputs
ToxicityOffensive or inappropriate outputCurate training data, use guardrails to filter content
HallucinationsPlausible but false statementsGround answers with RAG, verify outputs, keep humans in the loop
InterpretabilityHard to explain why the model produced an outputUse explainability tools and document model behavior
NondeterminismThe same input gives different outputs on different runsLower the temperature, test repeatedly, validate outputs

The full notes captured only the first slide of this carousel (regulatory violations). The other rows follow the AWS course.

Nondeterminism is the term to remember when a question describes different outputs from the same input.