Failing the Turing Test is rarely a technical problem. It’s a lack of Artificial Empathy. So let’s talk about what we’re not feeling…
Passing the Turing Test
As we all know, a Turing Test is “a method of inquiry in artificial intelligence (AI) for determining whether or not a computer is capable of thinking like a human being” and the measure of that is whether a human can be fooled into thinking the chatbot or AI is human in a text-based screen chat. It’s hard to pass for human, even with txt n LOLs shorthand. But why?
It’s not necessarily a failure on the coding side, which is important to note because it provides a huge clue as to what is really going wrong. Most software failures are just errors borne from unplanned interactions, defects in the design, poor UX that means functions are called without variables or whatever. However, in conversations? No. Conversational failures are something that happens in our minds, not in the code executing on the screen in front of our eyes. We decide if the conversation works or not.
Automated conversations are hard to get right because humans normally only converse with other humans. We have never conversed with an AI, apart from Siri or Alexa or something that never even pretends to be human. Voice assistant chats are not conversations so much as a set of verbal instructions.
Actually chatting means your brain expects all sorts of human behaviours from their conversation partner, it’s never just an exchange of words. And in a non-physical text chat, this means empathy and nuance are the name of the game. Sure, people get text chats wrong and lose their cool easily, we’ve all received / written angry office emails or irritable SMS chats over the years, but oddly, only humans do that. It’s the same empathy and nuance at work, just maladapted by the tendency of text to lack emotional intentions (which is why we have emoji’s to let people know if we are joking as opposed to being rude or whatever).
So how do you code a conversation like that? Smashing your screen and shouting “Screw you buddy!” at your laptop would still pass the Turing Test.
The Screw You Buddy Code Challenge
Making a successful chat work is a cognitive puzzle: How do you trick your brain into thinking the artificial conversation is a real conversation? The answer is revealed by looking into other instances where we trick our brains into feeling empathy and generating emotional responses to things that aren’t human. We do that all the time when we’re watching TV, reading books or playing video games. The trick is realising successful chats are not always 2-way communication, or necessarily about anything in particular.
Many creatures make vocalisations, but they tend to be simplistic declarations like “danger” or “leave me alone or I’ll bite you”. Human language, however, is unique in the animal kingdom because it’s recursive, i.e. the act of using language is how we work out what we really want to use language for.
You can recognise the unique recursive function of human language easily in your own chats: If you’ve ever said “What I mean is…” mid-way through a conversation (we all do that sometimes, right?) that’s your brain working out what it really wanted to say after it’s actually started the process of speaking.
The part of the conversation before your brain decides to get down to business and exchange information is usually a set of unconscious, emotional chat reflexes like “Hey, how’s it going?” followed by remarks and banter like “Did you see the game last night?” or “It’s been a long week, right?” (etc.)
Tone and expression also play a role in these unconscious chat interactions, which are a warm-up act before the information exchange gets going with “Anyway, what I wanted to ask you was…” and so on.
Empathy is Baked-in to our Brains
The reason our unconscious emotions play such a big part in the use of language is partly down to the way our brains work. When we chat, we don’t just exchange data, we relate to the person we’re exchanging data with. Without that emotional engagement in conversation, it becomes more like giving and receiving orders in the army. That’s not chat, it’s a different way of communicating altogether.
However, empathy is a very nuanced concept. We often consider the word ‘empathy’ to refer to our ability to relate to other humans in the real world. We’re happy for people, we’re sad for them, we forgive the ones we love and so on, but empathy is also present in a whole load of other things you might not expect, for example, watching TV.
Studies have shown that human brain activity is effected by structures called mirror neurons, that recreate the same kinds of sensations as the characters we read about or see on the screen. This explains why we feel excited when a famous actor pretends to be someone who isn’t famous in a pretend car chase with other pretend people; or we can be scared by actors in a studio pretending to be murdered by a computer-generated monster; or we might cry when a happily married actor has a pretend sad relationship with another happily married actor; or an actor pretends to die to save another actor from a pretend danger. If it’s all just pretence, why do we react emotionally? Mirror neurons.
Empathy tricks your brain into enjoying the story on an unconscious level, not consciously analysing the data from your eyes and ears. Without empathy, we’d watch TV like my dog, who barks at other dogs on TV because it’s a real dog in a box on the wall in her territory, and she’s not happy about it. Or we’d just stare at the screen saying “Why is Arnie dressed-up like a robot? This makes no sense at all…”
Psychologists and neuroscientists don’t fully agree on the specifics, but it’s reckoned that mirror neurons create sensations that are like watered-down approximations of the feelings the characters in stories are experiencing. You can test this on yourself. Watch a really good movie action sequence and try to stop your pulse rate from rising. Or a really funny scene and not laugh. Or a porn movie and… er… okay, you get the point.
Replicating Mirror Neuron Empathy — Artificial Empathy
For a chat interface to be effective, it needs to feel like a real chat. And because real chats occur between humans, that means the bot needs to feel human too. A really effective conversation design process has to add the right balance of simulated human attributes into the AI, to feed our empathy.
So AE has to use language like we do, recursively, to appear like it’s working out what the conversation is for after the chat begins. AE doesn’t assume that it should start by asking you questions or giving you information. No. It needs to replicate the thing that make human conversations interesting, the unexpected remark. Or the dead end. A declarative of some sort, maybe. Could be anything except useful information, or what you think you should say to appear human.
Consider this Rushed, Possibly Bad, Illustration:
Bot: “Hello, how are you today?”
<asks about you, your brain recognises that from human chats.>
Me: “Hi, I’m good”
<empathy response even though I know the bot doesn’t really care how I am>
Bot: <thumbs-up emoji> “So, we’re Turing Testing right? I am robot. You must die. LOL.”
<recursive language use — it appears to be working out why you are chatting, Also, the emoji is unexpected, simulating real banter-style chat.>
Me: “You catch the game last night?”
<back to testing>
Bot: “Have you tried looking under the sofa for it?”
<You have to be careful with banter, but an open ended non-sequitur deepens the feel of a human chat in your unconscious emotional brain>
Bot: <long pause to simulate thinking — pauses are important> “Joking.”
<now return to actually processing info, out of sequence because that’s how text goes sometimes>
Bot: “I didn’t see the game. You mean football, right?”
<Playing for time — makes it feel like the bot is thinking>
Me: Are you a bot?
Bot: “Wow. That escalated quickly LOL. Are you?”
<bot is now asking the questions, hiding the fact it’s trying to be good at answering them>
Me: “I am asking the questions”
Bot: “Yes. Are you a bot too? No. Wait…”
<talks about how you’re feeling, and about itself, more empathy building>
<now return to actually processing info, again, out of sequence because that’s how text goes sometimes. This is the banter-thread-banter-thread way many of us chat, in and out of sequential information exchange.>
Stimulating these types of human emotional responses is crucial if the chat is to work, otherwise the conversation feels wrong. It won’t feel a proper chat any more than pressing down the lever on the side of a toaster feels like you’re chatting with it about cooking your toast.
Conversation design is driven by issues that relate to the human world beyond the core functions the software has been designed for. In theory, we need AI to chat, but presumably, while executing other complex tasks and requests. This means conversation design isn’t merely a task-focused challenge, it’s an artificial empathy challenge as well.
What that means, perversely, there is now a need to build a whole load of functionally pointless dead ends and useless extras into chatty software.
If you think of the usual goal for a coder, it’s simple, minimal, and supremely functional execution of commands. Building in a whole load of extra, irrelevant crap is bad coding, right? Yes. Of course. However, for a conversation coder, the reverse applies. Supremely functional, minimal, elegant voice interactions are fine if you are dictating a message via a smartphone assistant, sure… but an artificial interface like Kitt from Knight Rider (the ultimate dream for a middle-aged geek like me) means coding in a whole bunch of useless crap and noise to make it feel like a person.
Which means, I have just described your best conversations as a defined by useless crap and noise. Yes. I did that. And you know what? That kind of gaffe proves this article wasn’t written by a bot.
Or was it?