When AI generates the words, who designed the sentence? Most teams have a clear owner for the button label and the error state. Nobody has claimed the paragraph the model wrote.

​

Many discussions about AI in design treat the technology as a behind-the-scenes accelerant — faster research, smarter prototyping, automated documentation. That work matters. But there's a harder design problem that gets less attention: what happens when AI-generated output is what the user actually sees, reads, and decides to act on?

​

That question showed up four times at Delta in three different forms:

 

• A passenger on a transatlantic flight choosing a meal from a menu built from operational data that was not optimized to be read by humans (Meal Description Enhancement)

• A flight attendant managing a medical emergency at 35,000 feet who needed a verified answer fast enough to act on it (FA Chatbot)

• A passenger who doesn't speak English trying to understand a safety announcement in real time with no way to ask for clarification (goNative!)

In each case the technology was different, the user was different, and the stakes were different. The design problem was the same: AI is generating language, and a real person needs to trust it enough to act on it — immediately, with limited ability to verify, and often under pressure.

​

The three projects that follow aren't a comprehensive AI practice. They're a pattern. Across all three, the same design principles determined whether the experience worked: human verification before AI output reached the user, transparent reasoning not just conclusions, user control at the point of interaction, and trust established at the system level before the AI speaks a single word. None of those principles are technology-specific. They're what human-centered design looks like when the artifact being designed is language itself.

​​