Challenges of Generative AI
Every risk needs a safety measure
The story: A new, very talented employee can still embarrass the company: blurting out a customer's private details, saying something offensive, making up a fact, or giving a different answer to the same question each time. A good manager doesn't fire them. They put a safety measure in place for each habit.
In AI/AWS terms: A model can make decisions that are unethical or socially irresponsible, so each challenge is paired with a mitigation.
For the exam: Know each challenge by name and the mitigation that goes with it.
The seven challenges
The story: Here is the new employee's list of bad habits and what the manager does about each:
- They mention a client's ID number in a public report. The manager removes names and numbers from files before the employee sees them.
- They post something that makes the company look bad. The manager sets a posting policy and checks posts.
- A stranger tricks them into reading out the customer database. The manager locks the files and checks what goes in and out.
- They use rude language. The manager is careful about what they read and adds a filter on what they say.
- They confidently invent a fact. The manager makes them check the company handbook first, and reviews important answers.
- Nobody can explain why they made a decision. The manager asks them to document their reasoning.
- They answer the same question differently each time. The manager asks them to stick to the safe answer and tests them repeatedly.
In AI/AWS terms:
| Challenge | Risk | Mitigation |
|---|---|---|
| Regulatory violations | Output exposes regulated data such as PII | Anonymize data, use privacy-preserving training, audit training data |
| Social risks | Content damages reputation or harms society | Test outputs, set content policies, monitor use |
| Data security and privacy | Sensitive data leaks through prompts or outputs | Encrypt data, control access, filter inputs and outputs |
| Toxicity | Offensive or inappropriate output | Curate training data, use guardrails to filter content |
| Hallucinations | Plausible but false statements | Ground answers with RAG, verify outputs, keep humans in the loop |
| Interpretability | Hard to explain why the model produced an output | Use explainability tools and document model behavior |
| Nondeterminism | The same input gives different outputs on different runs | Lower the temperature, test repeatedly, validate outputs |
For the exam: Different outputs from the same input is nondeterminism. Confident but false output is a hallucination, mitigated with RAG and human review.