Assume you know little about wine but you are nevertheless interested to purchase some good quality wine online. Go to any e-commerce wine business and you will have the option to search their database of photos, videos and text. But how can you search about something you do not really know much about?
What you will probably end up doing would be to select a wine on the basis of price and popularity. After all, if all those people have liked that “high-ranking” wine they ought to know something – right? Contrast this experience with you having the luxury of receiving advice from a professional sommelier. You tell them that you want to buy some wine, but you do not know what wine you might like. They have a conversation with you to understand you better, and then suggest a little known producer in Italy who has the right quality, for the right price, but – most importantly – the right character to match your own, personal tastes.
Wouldn’t it be great if that deeply personal experience could be reproduced in an online search?
Making online search more personalised is not simply “nice to have”. The economic impact of search in a world of vast amounts of digital content is enormous. As stored information has exploded beyond what is physically capable for our brains to retain, we are totally dependent on search engines to find and access content in a timely and meaningful way.
Unfortunately, as any search will show, what you end up getting are results based mostly on popularity and semantic context. The little known wine producer in Italy with the right wine for you has no hope of ever reaching you with his digital content if you just type “red wine” in the online search field – unless you know about them a priori and search them by their name. What you are most likely to get from any agnostic query are results from those who had a big budget to spend on digital marketing. In other words, the global digital media landscape today is a very uneven field that favours the rich and powerful – and this is a direct consequence of the way search currently works.
The problem with online search
There are at least two major problems with search. The first, and easier one, is that search finds it hard to understand the meaning of human queries. Google is investing a great deal of money and talent to solving this problem, for instance by making search capable for pattern recognition. Larry Page, in a rare TED interview in 2014, proudly revealed that Google search could recognise a cat when served with an image of a cat. Google is on the right path here, and their investment in machine learning is paying off. Artificial intelligence capable of communicating in natural language, such as IBM Watson, will further enhance our interactions with search. Siri and Cortana are already making search a lot smarter and easier. The story imagined in the sci-fi movie “Her” (2013) could become an every day experience in a matter of years.
But the second problem with search, the principal cause of digital inequality, will be harder to solve. Let’s call it “the problem of the unknown unknown”. How do I search for something that I know nothing about? Crack that problem and search will transform from the barrier-setter to the field-leveller of the global digital economy.
To better understand the significance of this problem, and see how it could be solved, let’s see how search works today. Every time we interact with search we essentially enter into a two-step, question and answer query. This short conversation ends with the search engine returning the result that best fits our query, based on a range of parameters that informs the engine’s algorithm, including social validation of content relevance. Using the wisdom of the crowd the engine pushes less popular content to the back of the thousands, and often millions, of result pages in order to serve you what is “most popular”.
This approach may appear efficient but it is based on an implicit assumption: that you seek the “most popular” answer to your question. It works perfectly well when you know what you are looking for. But, as in the case of the punter who wishes to buy wine without having any knowledge or expertise in wines, search fails when you do not really know what to ask. At the same time, small companies are doomed to be forever buried under numerous pages of search results, because they do not have the resources to create enough “word-of-mouth” on the Web 2.0 so that they appear amongst the “popular winners” on the first or second page of search results, where most people stop reading.
To overcome this problem we need a new and different kind of conversation between search engine and human; one that is more like two minds learning from each other, each helping the other to discover what they want. A conversation more like the one our fictitious wine punter had with his sommelier. But to have such a conversation the machine must acquire not simply an understanding of the human’s questions but an understanding of who the human is, and indeed what she or he are “thinking like”. The machine must, somehow, acquire our brain’s capability for empathy, or what psychologists call “theory of mind”. But how can we furnish computers with the ability to understand our minds?
The observed and the observer, dancing
One way to envisage this “next generation search” technology would be to revisit some interesting ideas from early cybernetics. One of the problems that early cyberneticians tried to solve was learning: how do systems increase their knowledge about the environment including other systems with which they interact, so that they adapt their behaviour?
Feedback provided an excellent concept to decode nature’s learning mechanisms. Biological systems, for example, sense and process information about the world and then take actions, their actions are then fed back as inputs for reprocessing, and so on. Control engineers used the concept of feedback to automate complex industrial processes. In early experiments, it was a human who observed and supervised the process, and adjusted the controls of the system in order to achieve the desired outputs. But gradually the human was replaced by computer algorithms that performed the business of supervision better, faster and safer. The human observer became ever more remote from the system. Search is a cybernetic system as well: we enter an input (our query) and observe the output (list of results) that the system returns. But what if we were part of search? What if we embedded ourselves in the feedback loop, and search-plus-us evolved into a so-called “second order cybernetics” system?
One of the pioneers who explored second order cybernetics in learning was the British scientist Gordon Pask. Pask did not subscribe to the Artificial Intelligence hypothesis that knowledge is a “thing”, or a “commodity”, to be stored, processed and analysed. Instead he saw knowledge as the result of multiple feedback interactions between two intelligent entities engaged in a conversation. “Understanding” and “learning” therefore arose not through pattern matching, or data analysis, but through the construction of a resonant cybernetic system akin to two persons dancing together, each one following the steps of the other, learning from each other the moves as they tried to dance in harmony.
Pask’s ideas received much attention during his time but have since faded from collective memory, like so many original ideas from early cybernetics. The more mathematically rigorous approaches of Artificial Intelligence have won the day. Through the utilisation of massive computer power, machine learning is nowadays applied effectively to make sense of the vast deluge of data that our digitised economy produces every millisecond.
Nevertheless, the analytical and predictive powers of machine learning, impressive as they are, will remain forever limited by the confines of logical processing. As Gödel and Turing have demonstrated, reducing the human mind to sets of symbols has an absolute limit. Beyond that limit humans become intuitive, creative, and wonderfully unpredictable. Machines will never be able to fully understand who we are, how we think, or what we really want or need, unless they somehow develop a “theory of mind” about us.
Second order cybernetics offer a paradigm of how this could be achieved. Combined with cognitive computing, one could envisage next generation search becoming more like discovery. Humans would start a search without really knowing what they want or need, and through a conversation with the machine arrive at the content that is most relevant to them. Such truly personalised search would open access to media and information currently obliterated by the algorithmic tyranny of the cyber-majority. Next generation search could level the field of the digital economy by making digital media intelligent and capable of finding you, in as much as you are capable of finding them.
Author: Dr George Zarkadakis is the Digital Lead at the Communications and Change Management Practice at Towers Watson, and the author of the book “In our own image – will artificial intelligence save us or destroy us?”
This post is published as part of a blog series by the Human Implications of Digital Media project.
Image: An employee of French National Agency for Employment (Pole Emploi) works simultaneously on French and German agencies for employement internet sites, at the joint German-French job center office in Kehl, Germany, on the French-German border near Strasbourg, November 13, 2014. REUTERS/Vincent Kessler