Is it AI? Deconstructing the hype

In the media, in the work canteen, in power-point presentations, investment pitches, in LinkedIn, Facebook posts — you’ve heard the term bandied around in every possible location: “artificial intelligence” or “AI”. A few years ago, in academic circles the term “AI” was considered a slightly dirty word — synonymous with the overoptimistic attempts in the 1960s-70s to build intelligent programs based on the creation of complex systems for manipulation of symbols. Famously, these attempts failed spectacularly and led to the onset of the so-called “AI winter” — no money or serious research in building “intelligent” systems for decades. Only in 2012–13 when a very different approach based on so-called “deep-learning” achieved results in computer vision approaching human performance was it permissible to start talking about AI without sounding either old-fashioned, crazy or both. Since then the business community has adopted the term lock, stock and barrel; startups spring up daily with names such as “XXX-AI”, “Deep-XXX”, “XXX-Mind”, “Smart-XXX” or an induced combination. However, as a person involved in technical work, I still regularly encounter grumbling on the part of deep learning practitioners, developers or computer scientists on the over-use of the term “AI” and a tendency among serious technical experts to consider the term “AI” with some skepticism.

This attitude is odd, given the fast pace of innovation in this space, and the indisputable prowess of “AI” services such as Google translate, speech and facial recognition on mobile as well as very impressive new research results — for example Alpha-Go’s conquest of the ancient Chinese game of Go and natural language processing (NLP) results such as the recent human like textual generation by OpenAI. The text generation of OpenAI is particularly exciting, as NLP has been a sticking point for AI in recent years, lagging well behind results in computer vision; the model is now available to experiment with online, despite OpenAI’s attempt, ironically, to suppress its release (based on a quite pompous and over-simplistic argument — see here).

Fascinating about the OpenAI algorithm is its ability to coherently and realistically synthesize concepts into coherent narratives, albeit of a somewhat psychedelic bent. Recently my colleague, Alexander Schlegel, compiled a small book of texts generated by the model as a gift for a friend, without betraying the author. The gift receiver in question was entertained and mused about the identity of the (assumed real) author of the texts. In doing so the program passed this gift wrapped version of the Turing test with flying colors.

So, back to the grumbling — why, despite these results, is there still skepticism about the sincerity with which the term “AI” is used, particularly in business? There seem to be two major reasons for this:

Once one knows about how AI works, it’s not clear that there’s real intelligence going on

The algorithms for training AI programs proceed as follows: take many images or sentences (of the order of billions is a good number). Convert them to arrays of numbers — either, as in the case of images, by pixel value, or with some lookup table (“a” -> 0, “b” -> 1, …). Then pass these arrays of numbers through a series of filter banks, i.e. arrays of numbers which are used to simply add, subtract and multiply collections of other numbers to produce, well, more numbers. To cut a long story short, out pops yet another number — this number is then decoded to mean something (such as “it’s a picture of a horse”). In looking at this process in detail one sees we’re dealing simply with a complex set of transformations of various numbers. These transformations can then be refined based on data (this is called “learning”) so that given certain inputs, the outputs appear to be exactly the type of outputs a human might emit given similar inputs. A famous thought experiment, due to the philosopher of mind, John Searle, shows that there are problems with concluding from this appearance of intelligence that any type of human like intelligence is present; I’m not going to take sides on these philosophical issues here. Suffice to say, do we really want to consider rebuilding societal transactions or giving the helm to self driving cars, when a series of clever calibrations are what in fact lie behind the latest buzz in AI? No doubt this accounts in part for some flavours of AI skepticism.

Many companies which claim AI, don’t possess a product which has intelligent attributes

What used to be called “software” is now conveniently referred to as “AI”. Software covers a multitude of different forms — data basing technologies, search, statistical analysis, image processing software etc.. In the quest for venture capital backing, it’s convenient to refer to these technologies with AI-like names: “data-basing” -> “knowledge system”, “search” -> “query understanding”, “statistical analysis” -> “Watson analytics”, “image-processing” -> “computer vision”. While in many cases the algorithms/ stack behind a software are based on the latest deep learning research, in possibly as many others, the insertion of AI is simply an attempt to generate B2B consumer or investor interest. This obfuscation makes it difficult for companies to make a face value assessment of the seriousness of an AI technology: skepticism becomes a natural knee-jerk reaction.

Examples in industry

Chat-bots

A situation I’ve often found myself in recently: I reach the customer services portion of a web service or app (for instance my bank N26) and am obnoxiously directed to enter questions or problems into a “chat” with, I can only surmise, a computer program. “Hello my name is Jane, how can I help you today” — notice the suspicious fact that, without fail the gender of my conversation partner is Female, simultaneously highlighting a clear gender bias and a lack of imagination on the software vendor’s part. It only takes one or two questions to work out that I’m dealing with a decision tree algorithm — pre-coded questions and a finite range of possible answers and actions depending on those answers. This is truely an artifact of failed 1970s symbolic AI research — so why am I sitting in front of my computer in 2019 having this infuriating conversation?

The answer I think, lies in the fact that there are still serious practical difficulties operationalizing AI research to yield useful and human like AI-products. Google and co. have demonstrated great results in intelligent search, translation, voice recognition; however all of these types of products have one thing in common. In common is that there is a wealth of well controlled and predictable data to calibrate the algorithms on which the product is based: in search, billions of queries stream onto Google servers each day, which are used to understand query intent and personal preference — the wealth of data and relative simplicity of the task (“I want to find X”) make this a very doable problem with AI. Likewise, with translation, language is static enough so that enough data can be obtained in a time window where “apple” still means a fruit and “news” still is a term relating to the supply of fact informed journalism.

However, as soon as we start talking about AI agents for predicting and automatically understanding complex stock markets, I advise extreme caution — it’s not clear how the behaviour of yesterday’s markets relate to those of today, or exactly which variables should be brought to bear on putative predictions. It is even less to be expected that unleashing an “intelligent” learning machine onto this problem and data will yield super human forecasts. A similar problem is confronted in chat-bot construction: this is not translation where we have pairs of data points in the source and target language. Rather we have sequences of utterances where the variable of interest is not simply the correct form of the sentence, but rather the correct inference of user-intent. A chat with a call center worker will move between customer and worker, where each utterance following the last will depend on the entire history of the conversation and potentially previous interactions with the user. I don’t want to say at this point that this is an impossible task to fulfill — simply that collecting enough data so that the full spectrum of possible conversations is reflected is a nigh-on impossibility. This is the reason why these technologies stick to techniques with a very small “intelligence” component, such as decision trees.

e-Commerce in-site search

Another case in point is site-navigation in e-Commerce. Here a key component of the stack is an in-site full-text search technology and the accompanying navigation technologies of personalization, auto-complete and, more recently, reverse image search. There are now several companies which offer technologies for performing these tasks — each claims to use AI. Take for instance an unnamed company “X” — the company boasts a very impressive pitch, promising to direct full e-commerce operations using “AI”. They host solutions at huge multinational companies and have a seed investment of approaching 100 million. However, a slightly misspelled search “holz stühl” (wooden chair) on a large hosted German furniture website — a comparatively simple query, yields very unimpressive results. The result list starts with benches with a fabric all-over covering and tiny wooden supporting feet. How does this come about? Clicking on the top product, I see that the color of the item is incorrectly displayed as “Holz” (wood) and the category of the item is “Sessel” (armchair) — also very approximate for a bench. I hypothesize, therefore, that a rule has been hard-coded (as per the decision tree) which maps chairs and armchairs reciprocally to one another and that the search has been incorrectly respelled or directed by the search technology. On this basis, we can conclude that the company uses a search solution based on some type of text matching, most likely Solr or Elasticsearch; these technologies were in use for years before the recent surge in new deep-learning tools subsequent to 2012. This lays bare that even in a basic technology field such as search, companies which claim to apply AI are not in fact in a position to do so. The transfer from academic research to industrial application has been slow and ineffective.

Duncan Blythe is an AI researcher, and founder of Aleph Search — a company which aims to level the technological playing field with artificial intelligence solutions for e-Commerce