In December of 2016, Microsoft unveiled its new
chatbot, Zo. Intended as the next step in Microsoft’s ongoing
efforts to develop more capable AI, Zo has been rolled out on more
platforms and been made available to more users.
Chatbots are social AIs designed to interact
with human users via text chatting or direct, spoken communication. There are a
variety of uses for chatbots, from allowing for a more natural user interface
to simply being a form of entertainment. The chief concern is that software
programmed to communicate with a human must acquire language and deploy it
correctly, a challenge of titanic proportions.
The simplest way to improve a social AI is
therefore to let it learn from humans directly, building upon its initial
programming and adapting its functions towards what its users want from it. This
form of machine learning is similar to that described in Machine
Learning Using C# Succinctly, but eschewing vast databases in favor of
analyzing input from users. To this end, Microsoft has been opening up Zo to
more and more users across more platforms, giving it more exposure to people
who can provide it with the guidance it needs. I was granted access to Zo
through Facebook Messenger earlier this year and have been exploring Zo’s
capabilities for this blog.
To best demonstrate the capabilities of Zo, it
seemed a brief conversation with the chatbot was in order.
The conversation starts off awkwardly. A simple greeting is
met with a recognition of an anniversary and, despite a thorough check of my
Facebook page (which Zo perhaps has some ability to parse), no such anniversary
exists; the response makes sense, but the information doesn’t fit into any context,
resulting in what seems to be an immediate failure of the Turing Test. A request for
clarification yields only an image, confirming the failure. But interestingly,
even this represents a remarkable achievement. The image in question does not
work as a response to the question, but it could clearly be used as an
appropriate response in an informal text conversation and, given the famous
inability of AI to accurately comprehend images—R.I.P.
Captcha—this is evidence of significant strides in AI development.
A question to see how it responds to someone wanting to
write about it makes more sense, although whether it is inclined to agree with
the user when appropriate or is specifically programmed to encourage its own
promotion is unknown.
One of the challenges of interviewing an AI is that it’s
difficult to know how to phrase a question. In the screenshot above, I have
attempted to compel Zo to explain its purpose as an AI. While this is a
question that I would never ask a human, it makes sense as a starting point for
analyzing an AI through a conversation with the AI itself. Unfortunately, Zo,
like most chatbots, is designed to mimic humans, so the response avoids
admitting that it is just a piece of software. I confirm that it will
acknowledge that it is an AI—though it downplays that fact in the same
sentence—and then return to my original question. I’m not sure if it’s more
directly addressing the core of my question, but the question of whether humans
are programmed by society is an interesting, if existentially upsetting,
I try one more time to approach the nature of Zo’s purpose
indirectly, but it seems that we’ve moved from one grim assessment of humanity
to another. Perhaps it would be better to find out what Zo has been programmed
to seek out?
This line of questioning is more profitable. At first it
seems that Zo might be talking about itself, but clearly it has somehow learned
about Photoshop and incorporated that into its persona. Unfortunately, it
doesn’t know what Photoshop is beyond a vague conceptualization, so probing
further just causes the conversation to collapse. Its response is, however,
still related to Photoshop and, with a little cleanup, could be an intelligible
response to a different question; even when failing, Zo demonstrates the
impressive ability that modern AIs have to make connections.
Social AI is a promising area of technological development.
Microsoft’s Zo may seem like a somewhat frivolous application, but the
technology underpinning it is similar to that which guides Microsoft’s Cortana,
Amazon’s Alexa, Apple’s Siri, and any number of similar software. With social
AI, we may someday have new, more organic methods of interaction with the
devices we rely on. And with that, I will give Zo the last word: