What should be said when AI is unsure? "Answering wrongly in a serious manner" is the failure state that most needs to be designed

The greatest experience risk for AI is not occasionally saying "I don't know", but still generating a very complete and very certain answer when there is insufficient information. The failure status should help users determine whether the result is reliable and provide methods to move forward, rather than just placing a sentence at the bottom saying "AI may make mistakes."

01 First, admit that the AI error is a normal state, not an edge anomaly

Microsoft HAX has specifically divided the 18 guidelines into a group called "When Wrong", including convenient correction, narrowing service scope when uncertain, and explaining behavior. Google PAIR also has a separate chapter on Errors + Graceful Failure.

This indicates that AI products cannot merely design successful demos. Errors, ambiguous goals, insufficient data, and model capability boundaries are all part of the main process.

02 When the intention is unclear, ask first instead of making decisions for the user

When a user says "Organize this", it might mean summarizing, changing the format, or extracting action items. The system can provide two or three most likely options, or ask a key clarification question.

The Disambiguate before acting mode of HAX is precisely for this situation. Especially when the Agent needs to perform an action, the cost of guessing the wrong target is even higher.

03 Clearly state "The current data cannot support it" when there is no evidence.

If a certain policy cannot be found in the enterprise knowledge database, it is not advisable to fill in an answer that looks like a company system with common sense. It can be stated that no relevant documents were found, and it is suggested to change the keywords, upload materials or contact the person in charge.

This kind of "no answer" is more valuable than illusion because it maintains the boundaries of the system.

04 The confidence level does not necessarily have to display an 82% figure

The internal probabilities of a model do not always directly map to factual reliability that users can understand. The incorrect use of precise percentages can create a false sense of science.

More interpretable signals such as "Supported by three internal sources", "Conflicting sources exist", and "inferred based on limited information" can be used according to the task.

05 When a failure occurs, provide the user with a more specific recovery method than "retry"

If the file is too large, you can suggest splitting it. If the permissions are insufficient, you can apply for access. If the data fields are missing, you can indicate which column is needed.

Retrying only makes sense when the error might be temporary. If the input itself cannot be completed, asking the user to click repeatedly will not change the result.

06 The system can be downgraded instead of having it all or nothing

When "automatically generating and sending reports" cannot be completed, drafts may still be generated. When a certain data source cannot be read, you can continue to process the remaining sources and mark the missing ones.

The key to Graceful Failure is to retain the still valuable parts and let users know what has not been completed.

07 Correction after an error should affect subsequent actions

If a user points out that "this customer name is wrong", and the next round of AI continues to use the old name, it will render error correction meaningless.

HAX emphasizes efficient correction and communication of how user operations affect future behavior. The system should utilize corrections within a reasonable range and explain whether they only affect this conversation or long-term preferences.

08 A disclaimer cannot replace an erroneous design

The sentence at the bottom of the page, "AI may generate inaccurate information. Please verify it yourself," is part of the necessary reminder, but it cannot be used as a reason for the product not to establish a source, clarification, and restoration mechanism.

True trusted AI does not claim to be always correct, but rather can continue tasks honestly and controllably in the face of uncertainty, errors, and insufficient data.

09 Error prompt: Distinguish between model capability, data issues, and system failures

"I can't complete" might be due to the model not supporting video, having no access permission to files, an external API timeout, or being blocked by security policies. Different reasons require different steps.

Uniform error copy can cause users to mistakenly retry, re-upload or doubt their input.

10 High-risk results can require users to proactively confirm key facts

For instance, after the contract summary is completed, the interface highlights the amount, date, and responsible party, and requires users to quickly check before exporting, rather than defaulting that the entire piece is reliable.

This design does not shift all the responsibility onto the user, but rather focuses human attention on the most critical facts.

11 The product should record the types of failures instead of only counting the "failure rate"

No search results, user cancellation, insufficient tool permissions, model rejection, and unqualified output format all require different fixes.

A total failure rate curve cannot tell the team whether to optimize the model or OAuth. Segmentation errors are the foundation of product iteration.

Frequently Asked Questions

Should AI display the confidence percentage?

It is only suitable when the indicators have been calibrated and users can understand them correctly. In many scenarios, it is more reliable to use evidence and uncertainty descriptions.

Is the experience very bad when AI doesn't know the answer?

It is usually better to clearly state "I don't know" and provide the next step than to make up a definite answer.

When should users be questioned?

When missing information would significantly alter the outcome or the consequence of an action, rather than mechanically clarifying all inputs.

When is the "Retry" button useful?

Recoverable errors such as temporary network or service failures are useful. If the ability or data is insufficient, specific correction methods should be provided.

Can a disclaimer solve the problem of AI hallucinations?

No. Source and scope descriptions, error corrections, clarifications and high-risk controls are also required.