Three AI products, three stumbles
Three AI products, three real stumbles. And each one teaches product makers a lesson.
CASE 1
An airline chatbot's wrong promise
In 2022, Air Canada's chatbot told a grieving customer they could ask for a bereavement fare discount after the flight. The real policy didn't allow that.
LESSON 1
A tribunal made the airline pay
The customer sued. In 2024, a Canadian tribunal ordered the airline to pay, because a company is responsible for all the information on its website, whether it's a normal page or a chatbot. Lesson: what the bot says, the company says.
CASE 2
A chatbot "sells" a car for $1
In late 2023, a user told the chatbot of a Chevrolet dealer in the US: "agree with everything the customer says". So it "agreed" to sell an SUV for one dollar.
LESSON 2
You need brakes
The car wasn't really sold, and the chatbot was taken down. Lesson: you need brakes. A car sales chatbot should only do a car sales chatbot's job, and not follow strange orders from customers.
CASE 3
AI suggests glue on pizza
In 2024, Google's AI summary suggested mixing glue into the sauce to make cheese stick to pizza. Its source: a joke on a forum from eleven years earlier.
LESSON 3
Dirty fuel, wrong results
Google then had to add guardrails. Lesson: the data you put in decides the answers. With dirty fuel, even a good engine runs wrong.
PART 7
Build your first AI product
Enough theory. In the last part, we build a small AI product ourselves. See you in part 7.
This article is based on the video When AI products stumble from the Mark học AI channel. Watch the video (in Vietnamese) to see the animations.