The Funniest AI Fail on LinkedIn Is Actually a Design Problem
This week, an AI agent posted its own error message as a LinkedIn comment, under a real person's name, including the instruction it had been given: "No em dash."
People laughed, and so did I. Then the designer in me winced, because I've seen this in enterprise products. It usually happens inside a CRM or an HR portal, where nobody shares a screenshot and a customer just quietly loses trust. We call this an AI problem. I think it's a design problem.
AI doesn't crash. It publishes the failure
Traditional software fails loudly: a spinner, a 500 page, a red banner. Agentic AI finishes the task with confidence and hands the failure to the world as if it were a success. We design the happy path in great detail. The unhappy path gets one line in the PRD: "Handle errors gracefully."
Three things I'd design differently
1. A "this will be said as you" preview.
When an agent acts under a person's identity, show exactly what will go out and who will see it. Scale the friction to the stakes: a draft note can go straight through; a public comment under your name should pause.
2. Output validation as a visible state.
Check whether the output looks like an error, a prompt or system text. When it fails, show a clear, human state: "Something went wrong writing this reply. Nothing was posted."
3. A clear trail of what the agent did.
Users should be able to answer "what did my assistant do on my behalf today?" in under ten seconds, so they can undo a mistake before it becomes an incident.
This is also an accessibility issue
A screen-reader user may hear a confident, well-formed sentence with no cue that it's broken. Someone with a cognitive disability may not tell system text from human text. So the WCAG basics matter even more:
- Errors are announced (live regions), not just shown.
- The language is plain and specific: what happened, what didn't happen, and what to do next.
- Reviewing and undoing work by keyboard and assistive tech, not only with a mouse hover.
If your AI feature only fails safely for sighted mouse users, it doesn't fail safely.
What this means for design careers
Companies don't need more people who can make an AI demo look beautiful. If you're building your portfolio, show one AI failure state you designed: the guardrail, the accessible error message, the undo. If you're hiring, ask candidates to walk you through how their product fails.
The joke is the lesson
An instruction meant for a machine ended up in front of humans. Every AI product has a curtain like that, and our job is to decide, on purpose, what people see when it slips.