As organizations and businesses use more AI, what's apparent is that the world is relying more and more on models. We're trying to understand consumer behavior and use that information to drive outcomes for organizations. But there's a huge issue in the way those models are used and built. A great example of this is the US election polling.
If we wind back the clock to 2016, you will remember that the polls said Hillary Clinton was going to win but she didn’t. There are three possible scenarios why that happened. Trump was a complete unknown at the time, voter turnout was significantly more than expected, and there was a huge amount of undecided voters. This meant that the models the pollsters were using were fundamentally wrong because they were built on assumptions that were incorrect. The polling models that were being used made sense prior to the election, but the assumptions weren’t updated for what was actually happening on the ground. As a result, the polls were wrong.
Fast forward to 2020, the polls today tell us that Biden is ahead. There are several fundamental things that may decide this election but these aren’t built into the polling models. There are significantly less undecided voters than last time, Trump is a very well known entity and his funding is on the decline, and there is the impact of COVID. The relative recency of Trump's own bout of COVID will have a huge impact on the way people vote in this election. After all, we’ve never had an election in the United States during a global pandemic. But these factors haven’t been taken into account in the polling models and we don’t know to what degree these things are going to impact voter behaviour.
The models that are created for polling make assumptions about a wide range of things like where you live, your socio economic background and your propensity to vote. There is a whole lot of science that goes behind the modelling that tries to profile the individuals polled and then extrapolates that out to the whole community. If the assumptions that you make are actually false, the margin for error in your model is much greater.
As we’ve seen this year, things can change dramatically very quickly and if you don’t update your models for these changes they won’t be accurate. Taking all of this into account, I wouldn’t trust the polls in the US election as I don’t think there is enough data to really understand how people are going to vote on the day.
There is a real lesson that businesses can learn here when they’re modeling and using data science. While AI algorithms are self learning, the reality is that the underlying assumptions that build any model change. So if you're building analytics that rely on these types of models, then you need to be aware that your models are fallible.
You can't just implement AI, you've got to continuously look at and manage it. Because how you build your business today, and the assumptions you build it on, will change. If you're using AI, make sure that you're constantly updating and maintaining your models and your base assumptions. This will ensure that the outcomes you seek are achievable.
What is augmented analytics?
Augmented analytics is a key driver for many businesses to modernize their BI capability. Learn what it is, why it's important and the data value it delivers.