My paper, “Exposing Weak Links in Multi-Agent Systems under Adversarial Prompting,” was accepted at the SE Workshop at AAMAS 2026, and the conference took me to Cyprus.* Fitting, because this trip taught me something about trusting machines a little too much.
We have grown far too comfortable handing our decisions to AI. It cost me a full night of sleep.
My route went through London, and this is the part that still irritates me. London demands a transit visa even if you never step out of the airport. Not to enter the country, just to walk from one gate to the next. A visa to change planes. Time was tight, so I had to book the priority service, and I turned to Gemini to walk me through it.
It was very confident and told me that the priority option had to be selected before booking the VFS slot, and that new slots opened at 1 AM London time, 4:30 AM in India. So I stayed awake, waiting to book the moment the clock struck.
The slot never came. Gemini was wrong. A few Reddit threads had the truth: you pay the standard fees first, and the priority option appears later in the workflow. There was no 1 AM window at all. I had lost a night of sleep to a machine’s confident guess.
The moment we stop verifying, we start making avoidable mistakes.
AI is the most powerful tool of our generation, and it is also confidently, casually wrong. I was ready to rebuild my night around a recommendation no one had checked. The tool was never the problem. Outsourcing the judgment was.