Post #3261879

2024-01-14 16:08 UTC

A good example here is the car dealership that tried using a ChatGPT bot for customer service. It quickly agreed to a "legally binding" offer to sell cars for $1: https://venturebeat.com/ai/a-chevy-for-1-car-dealer-chatbots-show-perils-of-ai-for-customer-service/ Would any human agent do this? No. Because humans understand things. This is a toy example, but if GPT-ish things don't work for very basic cases, how much can we rely on them for important cases? 7/

Replies (1)

@williampietri@sfba.social 2024-01-14 16:16

Moreover, I think there's a deeper problem here. Anybody who's ever tried to use large-company customer service (and I'm thinking of Amazon here) might say, "well a chatbot couldn't do much worse; it's not like talking to a person right now." There's a really handy concept called "failure demand". Some load on a system is "value demand". E.g., I walk up to the counter in the local store and say, "I'd like a quart of milk." But if have to come back later because they're out of milk or the milk was spoiled, that's "failure demand". For digital products, I'd argue most customer service demand is failure demand. So generally we shouldn't be putting ChatGPT in to tell people how to solve their problem, we should giving them good experiences in the first place. Chat's a bandaid. And bad customer service generates even more failure demand. So ChatGPT might lower costs, but it won't solve the root problems or on net improve things for customers. 8/

Open ##3261880