Human oversight means people examine system outputs and retain responsibility for their accuracy, safety, and contextual appropriateness. It prevents an LLM from having unchecked control over conversational AI users. section 1.2.2 section 1.2.3
Use two review patterns:
- Retroactive review periodically samples inputs and outputs, such as through a weekly monitoring process. It reveals bad outcomes but cannot prevent them. section 1.2.2
- Beforehand review uses the LLM to assist a human while the human makes the final decision. This is safer when responses affect users directly. section 1.2.2
- Combine automated scoring with human review when evaluating responses: the LLM can assess initial criteria, while people verify accuracy and contextual relevance. section 6.5
Copy editing instead of autonomous replies
An LLM generates static dialogue messages to address misunderstood intents or simplify a complex or rude dialogue flow. A human reviews and inserts the approved messages into the dialogue engine, rather than allowing the LLM to respond freely. section 1.2.2
Table 1.3 Dos and don’ts for LLMs
|
Don’t
|
Do
|
Why
|
|---|---|---|
|
Don’t use a model only because you saw it perform well (on a task that you don’t need).
|
Select a model suited to your task, or experiment with several such models.
|
Performance is task-dependent, including any parameters or prompt engineering. Tasks include generation, classification, extraction, question answering, retrieval-augmented generation, summarization, and translation.
|
|
Don’t discard a model or prompt because of one bad experiment.
|
Test on multiple inputs, models, and parameters.
|
Sometimes you’ll get unlucky. It takes multiple tests to have confidence in an LLM configuration.
|
|
Don’t blindly let the LLM have full control, especially in responding to your conversational AI users.
|
Apply guardrails at multiple levels.
|
Use the guardrail checklist to compare unrestricted LLM control with task-appropriate models, testing, and layered safeguards. section 1.2.3
