Three truths to help you get smart, fast about AI
Every day we’re seeing exciting new examples of AI’s impact on industry.
As the reality of an AI-powered future becomes clearer, more and more people are reaching out for help in wrapping their heads around the application of AI to their operations.
This is the right time for these questions. As I’ve written before, enterprises that aren’t already actively incorporating AI into their operations risk ending up the steam powered factories of the 21st century.
Yet it’s clear that AI remains a mystifying topic for many. It doesn’t have to be.
Based on our work applying AI to the enterprise, we’ve distilled three core truths to help you get smart, fast about AI today.
Truth 1: AI is an enabler of human intelligence, not a replacement for it
By far the greatest source of confusion around AI is the difference between artificial and human intelligence.
Whilst recent advances in deep learning — the process by which computers use multi-layered (‘deep’) algorithms to generate predictive insights — has expanded the capabilities of computers, the type of intelligence that humans take for granted — creativity, context, meaning and commonsense reasoning — remain firmly out of reach.
Instead, the reality of AI is that it is astonishingly powerful at a pretty narrow range of activities. These mostly revolve around identifying patterns within data sets. That’s not to say it isn’t extremely valuable, it is. It’s just not the same as human intelligence.
Indeed its value comes from its complementarity to human intelligence. AI works best when its computational processing power is used to inform human decision-making.
We see this in all forms of applied AI. Even in chess, where AI has been largely superior to human ability for 20 years, the most successful players still remain hybrid human-AI teams, which consistently mop the floor with their AI-only and human-only opponents.
The same principle applies in enterprise. At ViewX we use this truth to inform the way we’re harnessing AI to reimagine workforce development technology. Our neural networks are able to interpret task-level email data to predictively match each worker with the most relevant video information in that moment.
In other words, rather than trying to replace human intelligence, we’re using AI to deliver highly relevant information to human workers in a way that enhances human intelligence. AI reinagines our ability to process information, improving decisions, productivity and performance as a result.
The first truth to remember: AI is far more powerful as an enabler of human intelligence, than a replacement for it.
Truth 2: AI will only ever be as good as its data
If the difference between AI and general intelligence is the most misunderstood part of AI, then the fundamental importance of data is the most under-appreciated.
The far-reaching implications of data quality in AI were recently demonstrated by a team of researchers recently.
They studied the impact of gender biases found in two commonly-used image training data sets, both of which inadvertently reflected gender stereotypes by tending to tag women in activities like shopping and cooking, while activities like work and hunting were commonly linked to men.
The deep learning algorithms they trained on this data set began to routinely identify an image of a man standing in a kitchen as “woman”, while in another it was asked to complete the statement “Man is to computer programmer as woman is to…” and would provide responses like “homemaker”.
Worse still, the nature of deep learning means that AI trained on bad data will not merely reflect its failings, but actually amplify them.
In the stereotypes study, the frequency of gendered outputs ended up higher than the original biases in the data set because they were continually reinforced by the details of that labelled data.
This is a critical consideration for any business decision-maker, who not only needs to understand the quality of the training data used to develop the algorithms they’ll be relying on, but also needs to consider the quality of the ongoing data from their own businesses that will be used to generate AI insights.
Our Head of Product, Alex Gould, is a specialist on data design and has explored this in a great piece on the data design principles that we use for our AI products here.
All this underscores the second fundamental truth: AI performance will only be as good as its input data.
Truth 3: AI mistakes will look different to human mistakes
The final truth about AI, which gets by far the least attention but which is shaping up to be among the most interesting to watch as AI adoption matures, is that AI makes very different mistakes compared to humans.
The impact of this is that human users find it difficult to empathize with AI. The concern about a lack of empathy for AI may seem like a very solid ‘nice-to-have’, but this lack of empathy is damaging because it usually manifests itself in unwarranted disillusionment with AI products in general.
This issue will become an increasingly important practical consideration for business decision-makers looking to drive AI adoption in their organizations.
It is already creating practical barriers. An example is in the growing field of AI-powered scheduling assistants, which are facing adoption challenges in part because their models make seemingly ‘stupid’ mistakes that are obvious to humans. Consider an email sent at 1AM to a group saying, “Let’s get together first thing tomorrow to finalize this, and let’s add Julie”. The AI assistant is likely to make its error from incorrectly interpreting the ambiguity in “1AM” and “tomorrow”, which most humans would be able to navigate pretty easily. On the other hand, the error of a human scheduler is more likely to be an error of oversight, forgetting to include Julie on the invite at all, for example.
This is no less of an error, but it is one that many humans have made themselves or can otherwise more easily empathize with. This empathy gap is so pronounced that researchers are starting to program ‘human-style mistakes’ intentionally into AI programs to make them more likable.
We believe that as human users get more exposure to AI performance, this will improve. Until then this reality will require understanding and expectations-setting on the part of AI sponsors within organizations mindful of the third truth: AI mistakes will look different to human mistakes.
Lessons to take from these three truths
These three truths are designed to provide a simple starting point for anyone looking to get smart, fast on the basics of AI.
Taking them one level further, they can also provide a useful guide for business owners looking to plan for AI adoption.
Understanding AI’s value as an enabler of human intelligence helps you define the right role for AI in your organization.
Appreciating the importance of data helps you focus on the right details needed for optimal AI performance.
And being aware of the differences in AI and human mistakes helps you pre-empt the empathy challenges of human users.
In all, AI can be a daunting topic, but it doesn’t have to be. We hope that, armed with a better understanding of these AI truths, you’ll now be in a better position to start planning effectively for the coming impact of AI on your industry.