Error handling
When using AI, users may more frequently encounter errors or reach undesired outcomes. Help them anticipate, bypass, and avoid errors through feedback, expectation alignment, as well as other communication patterns.
Make uncertainty explicit
If your product is dealing with a lot of uncertainty, you need to find ways to share this uncertainty through the user interface (UI). Look at human interaction for inspiration: we use tone of voice and specific wording to indicate we may not be certain about something.
Atlassian places a small disclaimer at the end of its generated summaries.
Anthropic’s Claude shows a brief uncertainty statement at the end of its response.
Cite sources
Show what data and context was used by the AI during the decision making process. This allows the user to either validate the answer or reason if essential information might be missing.
Mozilla includes a disclaimer underneath their AI chat input.
Perplexity links sources at the end of its answers.
Provide controls
Provide manual controls so users can influence the output or they can take over when the AI fails.
Notion provides a few ways for users to override or iterate on its generated results.
Superhuman provides a few different ways to help users modify its generated emails.
Allow feedback
Provide an easy way for users to give feedback when things don’t go as they expected.
Touring allows users to give feedback on their AI-generated historical summaries.
Microsoft Copilot in Bing allows users to leave feedback after every message.