With all the current hype around artificial intelligence (AI), I was asked the other day how somebody could differentiate between “real” AI and fake AI. Ultimately, it’s probably not a very useful question, but here goes:
First, artificial intelligence isn’t a technology, it’s a “sociotechnical construct” — it’s used to refer to any use of computers to do things that previously only computers to do
There’s a rough consensus for different types of AI:
- “Weak” or “narrow” AI refers to niche uses of technology for a single defined purpose. It is the only form of AI actually available today.
- “Strong” AI would be able to address a wide range of different tasks and approach human reasoning.
- “Super” AI would be (will be?!) when computers have higher general intelligence than humans
So you may hear things like “that’s not real AI” when the example concerns only narrow AI. For example, there’s been some pushback on describing solutions as “cognitive” that don’t really live up to the term:
The main technology associated with the recent surge in AI buzz is machine learning (ML). Inevitably, there are different definitions of the term available, but the consensus is that it refers to any algorithm that automatically updates itself based on the data. This means it includes automated forms of traditional statistical methods as well as neural network and “deep learning” methods.
Neural networks date back to the 1950s, but the rise in computing power and the quantity of data available means that they can be made more powerful and sophisticated (e.g. generative adversarial networks, where different neural networks compete against each other, strengthening the quality of the overall algorithm)
Ultimately, the biggest factor in whether ML works well or not is typically not the choice of algorithm, but the quantity and quality of data available.
The “real” opportunity today is weak AI, which despite the name is poised to make a very big difference to almost all aspects of modern business, by automating complex-but-repetitive processes and allowing more intuitive interactions with data.
For example, today, there are armies of people in shared finance centers whose job is to match invoices sent to customers with the corresponding bank payments. Today, it’s typically a messy process, with as few as 40% matched automatically based on amounts, reference numbers etc. The remaining 60% have to be treated manually, because there are two invoices for one payment, or the amounts are slightly different, etc. But using ML, organizations have seen this rise quickly to 96% matching and beyond — a huge savings in time and effort. And that’s just one example of the hundreds of different business processes that will be automated using ML in the years to come.
And new enterprise digital assistants are allowing business people to work with their corporate systems through chat and voice, as easily as they ask a question to Siri, Cortana, or Alexa.
Compared to strong AI that would be able to figure out what’s wrong with you better than a doctor, these types of application may seem tame, but the big exciting opportunity for AI is the boring stuff — these applications are very real today, while the more glamorous use cases are not!