A chatbot can sound convincing and still give a useless or harmful answer. An LLM predicts a plausible response; it does not know whether that response is true. A guardrail is a control that reduces the risk of passing such output to a user, especially when the LLM helps run a live conversation. section 1.2.1 section 1.2.2
Put controls at more than one point:
- Choose a model suited to the task. Review its model card and training-data description where available; model choice is a first guardrail, not a complete solution. section 1.2.2
- Screen user input before sending it to the LLM. Keyword checks or a classifier can catch some abusive requests, though disguised requests may still get through. section 1.2.2
- Give the LLM contextual instructions in its prompt, and test the resulting behavior on multiple inputs. Do not hand it full control merely because one exchange went well. section 1.2.2 section 1.2.3
Keep control of a claim-status conversation
An LLM can ask for a member ID, claim date, and claim amount without a hardcoded question for every turn. The chatbot can preclassify input, then take back control when the LLM supplies variables for a claim search. But when a user does not know the required details, the demonstrated prompt keeps asking questions instead of handling the impasse. A successful sample dialogue is not enough evidence to remove oversight. section 10.2.1
Test guardrails against both unsafe requests and ordinary failures that waste a user’s time. The organization remains responsible for what its chatbot tells people: a Canadian airline was ordered to honor a nonexistent discount offered by its chatbot. section 1.2.2 section 1.2.3 section 10.2.1
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 model-selection dos and don’ts to check whether a configuration has been tested for its task and whether the LLM retains too much control. section 1.2.3
