The Contract

The biggest problem with AI right now isn't that it's not smart enough. It's that it feels sneaky.

You're typing in a doc, and suddenly text appears that you didn't write. You're in a support chat, and you can't tell if you're talking to a human or a bot. You ask for a summary, and you don't know if it read 5 emails or 500.

This is a failure of Contract Design.

The Contract is the agreement between the user and the system about: what is happening, who is in charge, and how to stop it.

When the contract is clear, AI feels like a helpful assistant. When the contract is murky, AI feels like spyware.

The 3 Laws of the AI Contract

Law 1: Visibility of Labor

Magic is bad UX. If an AI does work instantly and invisibly, users distrust it. They assume it hallucinated.

The Problem: User clicks "Summarize." Instantly, a summary appears. User thinks: "Did it actually read everything? Or did it just make something up?"

The Fix: Show the work. "Reading 12 documents..." → "Analyzing sentiment..." → "Drafting response..."

This isn't just a progress indicator. It's proof that the AI actually looked at the data. It builds trust through transparency.

Law 2: The Eject Button

Users are terrified of losing control. If they feel like the AI is a runaway train, they will disengage entirely.

The Stop Button: Always allow the user to interrupt generation. Not buried in a menu. Right there, obvious, labeled "Stop."

The Undo: Every AI action must be reversible. If an agent categorizes 1,000 emails, there must be a "Revert" button.

The Override: Users should be able to say "ignore the AI suggestion and do it my way" without friction.

Law 3: Provenance (Show Your Work)

"Because I said so" is not an acceptable answer from an AI.

Citations: If the AI makes a claim, it must link to the source. "Based on the Q3 report (link)" is trustworthy. "Based on our data" is not.

Confidence Signals: If the AI isn't sure, the UI should reflect that visually. Highlight uncertain text in yellow. Add a qualifier: "I found limited information on this topic."

The Reasoning Log: For complex decisions, show the steps. "I selected 'Urgent' because the customer used the word 'emergency' and has been waiting 48 hours."

The Takeaway

Trust is not a vibe. Trust is a mechanic. You design trust by designing the Contract.

If the user doesn't know what the machine is doing, the machine is malware—even if it's helpful malware.