I opened my email to check a flight booking. Gmail’s AI had summarised it into a clean itinerary card at the top: airport codes, times, a flight-path arc. It looked like it came from the airline. For the first time in my career, I felt a mailbox provider had actively done something inadvertently that made email less trustworthy.
It looked official, I thought it was my Airline. Airport codes. Times. A nice arc showing flight duration.
And it was wrong in possibly the second most dangerous way possible: it indicated the flight was leaving earlier than scheduled.
The card showed my departure time in red : 1:15 PM with 3:20 PM struck through, like it was the old schedule. Same thing on arrival: 2:40 PM in red, 6:40 PM struck through.
If you’ve ever had a flight moved, you know exactly what that formatting means: your booking changed.
Except…it hadn’t.
Those struck-through times weren’t “old times.” They were the times for the return leg of the trip, pulled from the same booking email, and Gemini stitched them into a single “updated” story.
In other words, my inbox didn’t summarise the email. It invented a schedule change and presented it with UI cues that screamed “trust me.”
It gets worse. The card also displayed the flight duration as 25 minutes. That’s not a small typo; that’s a system mixing time zones and legs incorrectly. But in a rush – heading to a meeting, dealing with children and elderly parents, packing – lots of people won’t sanity-check the duration. They’ll see the big red time and act. I was about to, I even checked if I was due compensation for such a big change at short duration. Checked weather to see why it might happen.
This is the core problem with AI inbox summaries: transactional emails aren’t blog posts. They’re not “information.” They’re instructions that drive real-world behavior – travel, finance, medical appointments, security alerts.
When an AI summary fails on a news article, you get a bad take. When an AI summary fails on an offer, you lose revenue.
When it fails on a flight confirmation, you can miss a plane.
The scary part isn’t that it was inaccurate. The scary part is how it was inaccurate:
- It mixed outbound and return flight data.
- It used strikethrough formatting that implies a verified schedule change.
- It swapped or mis-assigned key fields (airports and times).
- It produced nonsense duration while still looking polished.
A polished lie is worse than an obvious error.
This isn’t an argument against AI assistance. It’s an argument for hard boundaries. If an inbox wants to summarise transactional email, it must behave like safety-critical software:
- Don’t “generate” itinerary fields; extract them with provenance.
- Don’t combine multiple legs unless you can separate them cleanly.
- Don’t use “schedule change” UI unless the sender explicitly indicates a change.
- Always provide a one-click “show me exactly where this came from in the email.”
- If confidence is low, say so – and don’t present it like a boarding pass.
Until then, the safest workflow is old-school: treat the airline’s app / booking page as the source of truth, and treat AI inbox summaries as untrusted hints.
