There is currently an enormous confidence in the possibilities of Artificial Intelligence (AI) and how it will change our world in the near future: self-driving cars, machine learning algorithms that help us diagnose diseases and intelligent chat bots that will disrupt customer care.
I want to take a step back and put this all in a realistic perspective. Let’s take a look at some of the success stories of the last year:
The first computer Go program to beat a human professional Go player, something which was long thought to be impossible. It was build using a Monte Carlo tree search + a 13 layer deep neural network, feature detection algorithms, reinforcement learning and human designed heuristics.
Tesla’s self driving car uses a whole array of sensors and GPU processors but it’s core component is a deep neural network to detect various objects along the road.
Microsoft has reached human parity for speech recognition using the careful engineering and optimization of convolutional and recurrent neural networks.
IBM’s Watson uses deep learning with artificial neural networks to answer Jeopardy questions but was later also applied to the medical domain.
Artificial Neural Networks
From the above list, it’s clear all these advancements were made possible with the same type of technology, a ‘deep artificial neural network’. Artificial Neural Networks (ANN) are old technology that was first described in 1943. At the beginning of 1970 ANN’s got out of fashion and the best results in AI where obtained using different algorithms, such as tree learners or support vector machines. However, at the end of 2010, neural networks became popular again and the biggest drive behind these successes can be summarised as: nowadays computers have become fast enough to actually make it possible to train and run more complex artificial neural networks.
The word ‘neural network’ combined with the success stories gives people the feeling we currently have the capability to create a very smart artificial brain that is capable of learning and solving anything we throw at it. However, this is not the case. Neural networks are a relatively simple algorithm that can be trained to solve a very specific case. For example, to classify or cluster data based on input. In order for the network to do this it needs a big set of data to be trained upon. But, once it has been trained you can’t apply it to a problem on a completely different domain. For example, AlphaGo that has been trained to solve the game of ‘Go’ is completely useless when trying to use it for detecting road signs. And even a system that has been trained to detect road signs will be completely unable to detect a new road sign which has not been in its training set.
Current Artificial Neural Networks are not capable of learning in the same way that we humans do. Even incorporating simple memory with a neural network is very hard to do and simple problem solving algorithms are still being hard-coded in advance. This means that while it is possible to apply AI to certain specified knowledge domains, without any further breakthroughs it is impossible to have some sort of smart brain that can be applied in general.
This also very much applies to AI chatbots. They are at the top of the hype cycle and often presented as a technology that will greatly help your customer service. However, I feel that this is a complete misrepresentation of the real situation because of the following reasons:
An AI chatbot needs to operate on a very general domain
If you want to have a meaningful conversation with a chatbot, the bot needs to be able to understand human conversation. However, this requires it to operate on a very general domain and this is a problem that hasn’t been tackled yet. When you look at the listed succes stories above, they where all operating on a very specific problem domain.
This means, that either you’ll need to restrict the bot to certain subset of conversation, making the interaction difficult for humans, or you’ll need to expect and accept errors while a bot is trying to fake a general conversation. This can lead to problems and even a PR disaster as Microsoft learned the hard way.
An AI chatbot needs lots of training to be useful
You can’t just roll out a chatbot and expect it to provide value. If you want it to provide meaningful answers it needs to know everything about the problem domain it is being applied to. This means you will need to provide this content in advance and in most cases, building and maintaining this knowledge is the
hardest part of customer service. Not so much the interaction with customers.
A chatbot will not be able to learn from interaction
While bots can learn to better answer questions from interaction they will always do so with existing knowledge. Whereas a human can learn new knowledge from interaction and thus contribute to the shared knowledge, a chatbot can only serve as an alternate interface to the gathered knowledge.
While I don’t believe that chatbots can replace human interaction soon, I do believe that chat bots can serve a purpose as an interface for clients using a mobile device to offer your service through different channels, such as Whatsapp chat, Facebook messenger, and Skype. However, in these solutions the bot basically functions as a different interface compared to, say, accessing your website directly.
Will AI be the future of company interaction?
This continuous cycle of hype followed by disappointment, criticism and even funding cuts are well known in the AI world. They even have a word for it, the AI winter. We are now at the top of the hype cycle, where, if we are not careful and realistic, we will dive straight into another winter.
The future of AI may be bright but it is not immediate. Despite all the AI hype, I wanted to share a more realistic vision of what’s ahead. At this point, we can use elements of this technology to support us, humans, but there is no doubt in my mind we will remain in control and handle the conversation.