Why Chatbots Continue To Fail Us
Digital anthropologist Amber Case explains why conversational AI still has a long way to go before it meets our expectations
Chat—whether that’s apps, interfaces or bots—remains a hot topic for marketers and entrepreneurs. Increasingly fueled with artificial intelligence and complex algorithms, chat promises an opportunity to replicate human interaction a massive scale. Sometimes this works—many times it doesn’t.
For a podcast on chat, PSFK founder and editor-in-chief Piers Fawkes spoke to digital anthropologist Amber Case, who has been studying human interaction with chat and AI technology. Amber weighed in on just how far the medium still has to go before it can really meet our expectations.
Piers: You’ve seen the whole evolution of the chat medium, and the use of learning computing or cognitive computing being integrated with it.
Amber: My grandfather worked on AI and my dad worked on voice-concatenated speech systems for telecoms. I grew up testing the interfaces and the chat systems, and then I grew up in a smart home, and then later made a voice assistant chatbot in IoC and got to see that evolve over four years, and then got four years of extra history on that.
Where are we now? Are we far along from those early days?
No, we’re awful. We’re in the exact same spot we’ve always been. It’s a scam.
We start with Joseph Weizenbaum, the psychologist who starts making fun of the AI boom in the 1960s. He says, “Sure. I’ll make an artificial intelligence. I’ll make something that can replace psychologists.”
He makes ELIZA the chatbot. ELIZA the chatbot starts being really popular. His secretary starts to use this chat-based interface. The reason why it works is because he’s just put in the most stereotypical, worst questions that a psychologist asks, like “How are you feeling?” and “How do you feel about that?” and “How about your mother?” “Tell me about your parents.”
What it ends up doing is providing a non-intrusive, nonjudgmental interface for people to write to themselves. It’s a series of self-journaling props, basically. It’s just a fill-in-the-blank Mad Lib, if you think about it.
The reason why his secretary and a bunch of other people loved ELIZA is because they could, for the first time, just have this little documented archive of what they were feeling, so that they might have some perspective on themselves.
The bot is never going to offer any insights. The bot is just going to get the imagination within your mind to act better than it would’ve if you were to be stuck in a mental loop. I think that’s the main thing that people forget—it’s the same with immersive computing—what happens when somebody brings something to an object. Like when a kid who’s five brings their imagination to LEGO blocks.
You don’t need to have anything high-tech. You just need to have some blocks and the kid’s imagination will take care of the rest. For bots, you just need to have very simple responses, not necessarily interactive in the world like a reactive chat interface, and that will give you more than if you overbuilt a thing.
Every time somebody tries to make something that acts like human, it gives itself away really fast. If it just reacts to you with a variety of different responses, like if somebody says, “Good morning,” in the chat channel, it will say, “Good morning.” It has five different ways of saying it so it sounds more natural.
We have bots that talk to us in English or some language, and when they spew sentences that sound like a human, we have expectations that we can talk back to them on a human level. We expect that they are sophisticated enough to understand us because they sound sophisticated to us.
It gives us no information about how dumb these things are. As Mark Weiser of calm technology said, “We don’t need smarter devices, we need smarter humans.”
How do these things help us? By giving us information at the right time, by allowing us to have choices, by bubbling up important details that we might need in a way that allows us to just glance ambiently and maybe request more with some key terms.
For a really good customer service automation system, for instance, most of the data from what people have asked in the past will be written down, transcribed and tagged with an information architect involved, not a data science system. As an information architect, they know how to categorize information. Data science is just something you throw at something when you’ve taken the wrong data and you have too much money you can overpay for them.
You get an information architecture manual, categorize the information and when somebody has the technical support question it goes in. Then you’ll have user experience people look at the top questions and work on the user experience part to alleviate those top questions. If something is not in the system, it automatically connects to a real-life person who gets in the chat history so they can alleviate the problem. That’s the hybrid approach. I wanted to make a bot that did this for me.
Most people asked the same questions. “What’s a cyborg anthropologist? What’s comm technology? What are you doing right now? What’s your Twitter handle?” Most of those responses could be handled by a chatbot. If somebody asked a question that’s not in the database, it should send me a text.
I should be able to text back the response, and the bot should be able to get it or store a variety of responses. The problem with AI is that—because we’ve had booms and busts, and we’re right atop of a freaking boom now—the expectations are totally inflated.
What we have to remember is that AI and our expectations of it being perfect come from film. They come from the interactive voice control on Star Trek and all these other things. It’s more cinematic to have somebody interact with a voice instead of text. It’s really bland if you have text.
In order to make it cinematic, starting with Rossum’s robots from Russia, we embodied a technical creature in the shape of a human so we could stage plays—man versus machine. That stuck. We still think that an AI can be just like the movies, but we have trouble understanding each other. We have trouble understanding accents.
AI is a fake term. There’s narrow AI, there’s some general AI. There’s not any superhuman AI out there yet. It’s a forgotten term, cybernetic interaction or feedback loop. The best stuff handles 70% on Google Search and then gives us back 30%. We choose from that. A bot that makes a choice for us can fail disastrously.
Listen to more experts discuss chat as an interface on our PurpleList podcast episode ‘Is Chat All Talk-Talk?’ More insights on the topic can be found in the PSFK research paper Reinventing Customer Care with Chatbots.