Failure State Design

In traditional software, an error is a bug. It means something broke that shouldn't have.

In AI software, an error is a feature. It's a statistical certainty. Your model will be wrong. Your retrieval will miss documents. Your classification will misfire. This isn't a bug to fix; it's a behavior to design.

If you design a UI that assumes the AI is right, you're designing a trap.

You have to design the Red Path—the user journey for when the system fails, lies, or gets confused.

The 4 Levels of AI Failure

Level 1: The "I Don't Know" (Epistemic Uncertainty)

The AI knows that it doesn't know. This is the good kind of failure.

Bad UI: "I can't help with that." (Dead end. No explanation. No path forward.)

Good UI: "I don't have access to your private emails, so I can't summarize them. I can only see documents in your shared drives. Would you like me to search those instead?" (Explanation + Reason + Alternative)

Level 2: The Hallucination (Confabulation)

The AI thinks it knows, but it's wrong. This is the dangerous kind.

The AI won't tell you it's hallucinating. It will confidently present fiction as fact. You must design defenses.

The Trust Battery Pattern: Don't show AI answers as absolute truth. Show them as proposals with confidence signals.

The Verify Interaction: For critical data, design a UI where the user must review before acting.

Confidence Scores: Surface the AI's confidence, not as a number, but as a signal.

Level 3: The Ambiguity Loop

The user asks for "the report." The AI doesn't know which report. This happens constantly.

Bad UI: AI guesses randomly.

Good UI: Disambiguation UI. The AI pauses and presents options. "Did you mean the Q3 Financial Report, the Marketing Report, or the Weekly Sync Report?"

Design Principle: Ask, don't Guess.

Level 4: The Safety Trigger

The user asks for something dangerous, unethical, or outside policy.

The Soft Block: "I can't write that email for you because it could be seen as harassment, but I can help you draft a professional message that addresses your concerns. Want to try that?"

The Hard Block: Immediate stop. No attempt to be helpful. Clear statement: "I can't help with that request."

Failure States > Success States

Here's the counterintuitive truth: Design the failure states first.

Most designers do it backward. They design the beautiful success case, then treat errors as an afterthought. "We'll figure out error handling later."

In AI, the error is the design challenge. Anyone can design a happy path. The difference between a trusted AI product and an abandoned one is how it behaves when things go wrong.

Start every AI feature design with: "What are all the ways this could go wrong?" Then design for those. The happy path will take care of itself.