Why Today's Artificial Intelligence isn't Intelligent

Looking at the news conferences of smartphone makers or LG’s most recent CES keynote, one would think that we are already living in the age of Artificial Intelligence (AI). These companies are showcasing their – so called – “AI enabled technologies” through grandiose demos. For instance, at this year’s CES, we learned how LG’s AI – which now lives in everything from cars to washing machines – makes our lives easier and better by changing our washer settings to sport mode when it knows that we will be working out based on our calendar entries. More recently, at Samsung’s Galaxy S9 event, we learned how AI can improve image quality and help with translating foreign texts.

Number of companies mentioning AI in their earnings calls based on Bloomberg's info.

However, sometimes AI technologies make spectacular – and newsworthy – mistakes, and we are left wondering whether AI is in fact intelligent or if it’s just the newest marketing cliché. Examples of such mistakes are when Alexa deiced to blast music in a German apartment while its owner was out, prompting the police to break in to shut it down, or when Facebook’s AI powered chatbots created their own secret language to communicate with each other. In more day-to-day scenarios, it’s great that Sony’s new AI prompts us to set a 7AM alarm on our XPeria phone when it recognizes that we usually get up at 7, but it’s annoying – and not very intelligent - when it does the same even though we already set an alarm for 8AM.

Today’s AIs are better described as “street smart” than as “artificially intelligent,” but – unfortunately – the latter sounds much catchier. Moreover, it seems as if marketing departments grew tired of the term “smart” (i.e.: smart phone, smart fridge, smart TV) and are now labeling everything as “Artificial Intelligence powered” even though there hasn’t been that big of a change in the underlying technologies. Unfortunately, this overuse of the term AI means that an average person can’t really tell the difference between what software engineers call narrow AI and general AI (also referred to as human-level AI or superintelligence).

The truth is, when companies like LG and Samsung talk about AI, they are usually referring to a narrow version of AI called “Machine Learning” – which is a field of computer science that dates back to the 1950s – and rely on algorithms that were also invented up to 50 years ago. In essence, what these Machine Learning (ML) algorithms do differently compared to traditional computer code – such as the ones we all wrote in MIS 301 – is that they don’t require a set of steps to follow, but instead they come up with their own steps based on what the input data and what the desired outputs are.

Often, these ML algorithms are based on statistical principles that require a lot of parallel processing power and work better when there is a large amount of data available to them. Interestingly, these two characteristics can also explain why ML (labeled as AI) has been becoming so popular. Namely, since we are producing more data (2.5 exabytes/day) and carrying more processing power in our pockets than ever before, the conditions are perfect for ML algorithms. Concerning processing power, it’s also noteworthy that companies began adding special chips to aid with the computing power required in ML – such as Apple’s new A11 Bionic chip or Huawei’s Kirin 970. (Classical CPUs can also perform ML operations, but it’s better if more cores work on it. Thus, companies often rely on GPUs to do ML calculations or special NPUs – i.e.: Neural Processing Units – such as the one built into Apple’s Bionic chip.)

Performance of GPUs over the last 10 years based on NVIDIA's information.

While ML algorithms work great sometimes, they are not really intelligent. In fact, the weakest forms of ML algorithms aren’t able to accomplish any form of learning, while the more advanced versions – which are sometimes referred to as narrow AI – are only able to accomplish very specific tasks such as figuring out what mode your washing machine should be set to. We can’t expect such AI to come up with solutions to our most pressing problems – unless, those problems are very specific like how to win a chess game. However, on bright side, ML or narrow AI isn’t something that will likely take our jobs away or cause catastrophes similar to the one depicted in major blockbusters precisely because they are not very good at complex tasks.

What should cause a bit more worry is AI that reaches consciousness (thus, called human-level AI or general AI) or artificial intelligence that goes far beyond human capabilities – called superintelligence. An important distinction compared to narrow AI is that general AI is able to accomplish complex tasks. In other words, a general AI is a form of artificial intelligence that can accomplish any given goal as well as any other intelligent entity (e.g.: humans) within a finite time. Therefore, while a narrow AI might be really good at playing Go, a general AI would be at least as good at playing Go as humans while also being at least as good of a chef, engineer, project manager, police officer, etc. as us. As of now, such AIs don’t exist but there is plenty of research done in this area. While going into detail on why superintelligence should be worrisome or why its creation is not impossible is outside of this blog’s scope, I included a brief table clearing up some of the misconceptions around general AI below.

Based on Max Tegmark's book on AI.

Hopefully, at this point, it’s clearer what the difference is between a narrow AI or ML algorithm and a – not-yet existent – general AI. Hopefully, it’s also clear what big companies’ marketing teams mean when they brand something as “AI powered” and why that’s a different AI than the one that Elon Musk is worried about.

If you are interested in this topic, then these are some of the readings that I would recommend:

• Life 3.0 Being Human in the Age of Artificial Intelligence by Max Tegmark 
• The Debate on the Last Invention that Man Need Ever Make by Andras Szep 
• The Master Algorithm by Pedro Domingos 

Questions to be answered:

1. What specific tasks in your day-to-day life could be made better/easier with Machine Learning?
2. Should marketing teams continue using the term Artificial Intelligence for Machine Learning tools or do you find it misleading?
3. Have you ever come across a situation where you used an “AI” but found that it wasn’t too intelligent?