It’s impossible to miss the astounding progress made in artificial intelligence (AI) over the last decade. But, let’s not get ahead of ourselves. In very real ways, the technology’s definite limitations tell us as much about the future of work as what the systems can do.
To be sure, what they can do is remarkable.
Powered by a technology known as deep learning, these systems have crossed into domains that historically were defined as exclusively human. Computers can ‘see’, translate, read handwriting and process all manner of unstructured data, including music, voice, and images – and do it all with breathtaking acumen.
In a facial recognition study, an AI system recognised faces with 98.5% accuracy, outperforming humans, who achieved 97.5% accuracy with the same images. Today, facial recognition, with all its more controversial implications for civil liberties, is routinely used as an alternative to door locks and to activate smartphones.
In 2017, an AI system called ‘Case Cruncher Alpha’ was pitted against 100 lawyers from top London law firms. The task was to predict the outcome of 775 real cases relating to insurance mis-selling by the UK financial ombudsman. The AI predicted the right result in 86.6% of the cases compared to 66.3% accuracy from the lawyers.
A rival of the human mind?
These examples and many others showcase the impressive capabilities of AI systems. Perhaps most intriguing is how these systems perform. They aren’t programmed; they’re trained, and can improve with use.
Although these systems excel at a particular class of problems, it is difficult to extend them beyond the specific problem they were originally trained to solve. On one hand that creates a host of technical challenges but on the other, it gives reassurance to those who are fearful of encroachment on what they view as the realm of human work.
An AI system designed to discover potential therapies for leukaemia, isn’t going to pop out a vaccine for COVID-19, or even help with a cure for diabetes. They’re specialists, not broadly capable generalists.
For example, many natural language processing algorithms are so brittle they can understand a person from Sydney but fail to understand the same English language sentence from someone in London or someone with a slightly different accent. Deep learning works by ‘memorising’ a set of data, but fails to make accurate predictions for examples outside of its original dataset.
Perhaps more instructive, these systems also struggle to solve lots of problems that humans can perform without conscious thought. A human will fold a crumpled towel in seconds. The same task takes a state-of-the-art robot many minutes to complete and even then it often fails. This is known as Moravec’s Paradox, which observes that although we can design computer algorithms that solve complex problems like flying an aircraft, we are still unable to design systems that complete simple tasks like tying shoelaces or folding a napkin.
These limitations mean we are still a long way from being able to build what is known as an artificial general intelligence (AGI); an AI system with the power and flexibility that rivals the human mind. Although it would be foolish to bet against our ability to innovate, it is unclear whether we will ever be able to build an AGI, or whether we should.
AI systems aren’t going to replace humans anytime soon, but we do need their help. As the world grows more complex we are increasingly relying on data to make decisions. But as AI thought-leaders, Erik Brynjolfsson and Andrew McAfee, said in a recent Harvard Business Review article, “Over the next decade, AI won’t replace managers, but managers who use AI will replace those who don’t.”
Our approach to AI at Xero
At Xero we’re fully committed to this approach known as augmented intelligence to support our accountant, bookkeeper and small business customers.
It’s about using a collection of technologies – machine learning, analytics, natural language and so forth – to automate laborious processes, explore possible scenarios and create greater certainty around business decisions, but not replacing the essential human element.
To achieve this we are focusing on three broad areas of investment:
- The first is streamlining the process of getting information into Xero, whether that’s the ingestion of receipts, bills or bank statements. We want to remove as much of the work associated with entering data into Xero as we can.
- The second area of focus is classifying the information that’s ingested in our application. This is the area where AI systems can excel and deliver massive value.
- The third area is the most interesting and open-ended. It’s all the things we can do once we have high-quality, appropriately classified information in the software. The objective here is to assist small business owners and their accountant or bookkeeper to run their business more effectively, as well as explore future possibilities, take action on their behalf, and identify opportunities or challenges before they arrive.
In brief, our focus is on a kind of machine intelligence that doesn’t replace, but assists, that doesn’t eliminate, but complements and serves to lift the performance and value of the people who use it. That’s what we are working towards, and it is completely aligned with everything that #Human represents in our mission, direction, culture and ethos.
Xero On Air
Find out more about the opportunities of augmented intelligence for accountants, bookkeepers and small businesses in the Xero On Air episode – Towards 2030: The technology shaping our world.