Chatbots are this year's apps. Tay says slow down

At F8 on Tuesday, Facebook opened Messenger to bots. Businesses can now build chatbots that talk to users inside Messenger, and the demos showed people ordering flowers and checking the weather by chatting. Microsoft has been talking about “conversation as a platform” for a few weeks, and Kik and Telegram already have bot stores. By Wednesday afternoon two of our clients had emailed asking whether they need a bot strategy.

Two weeks earlier, Microsoft put a chatbot called Tay on Twitter. It was designed to learn from conversations with users, and within about a day users had taught it to post racist and genocidal things. Microsoft took it down and apologized.

Tay is a funny story, but I think it shows the real problem underneath the bot hype. Current chatbots don’t understand what they’re saying. The better ones are scripted decision trees with some keyword matching or intent classification in front. They can recognize “I want to order flowers” and ask for a delivery address. As soon as a user goes off the script, they fail. The ones that learn from open conversation, like Tay, pick up whatever statistics of language they’re fed, with no model of what any of it means or whether it’s acceptable.

For businesses, this means a bot is basically a form with a chat interface in front of it. Sometimes that’s better than a form. Often it’s worse, because a form shows you all the options at once, and a chat makes you guess what the bot can do. The flower demo at F8 took more steps than tapping through a normal app would.

So my advice to the clients who emailed is to wait, unless they have a narrow, high-volume use case like order tracking or appointment booking, where the whole conversation is predictable. I told them I don’t expect conversational interfaces to replace apps in any serious way for at least a decade. To get there, bots would need real language understanding, the kind where the system tracks what you mean across a whole conversation and handles things it wasn’t scripted for. Neural networks are getting good at translation and captioning, but nothing I’ve seen is close to holding a useful open-ended conversation.

The one bet I’d make is that when conversational interfaces do work, they’ll come from models trained on huge amounts of text, not from hand-built dialogue trees. Tay failed partly because it learned from a small, adversarial stream of tweets. A system that had read far more, with some way to steer it, would be a different thing. I just don’t think anyone knows how to build that yet.