3 Reasons Why Not to Blindly Trust Predictions
Data Scientists are basically fortune tellers. They predict the future by looking at what happened in the past. However they don’t use a crystal ball; they use advanced mathematical and statistical models to find correlations and connections in large amounts of data. They project this on the present in order to predict what will happen with as much certainty as they can. That does not mean you should always blindly trust these prediction models – especially not when human lives are concerned. When using Artificial Intelligence (AI) in decision-making, always consider this trio of critical footnotes before drawing definitive conclusions.
By Erica D’Acunto, senior data scientist at ORTEC
Margin of error
Every AI prediction always has a margin of error, no matter how small. No matter how rich the historical data, and no matter how advanced the model that is applied to it: a 100% chance only exists in theory. This margin of error is often acceptable for predictions concerning capital. A bread factory wanting to predict demand in order to reduce waste? A model for investors to predict movements in exchange rates on the stock market? A predictive maintenance application that predicts when a part of machinery will have to be replaced? If the prediction model works with 98% accuracy, it naturally offers great added business value – that 2% is then negligible. But awareness of this margin of error, no matter how small, becomes much more important when the prediction affects human lives. Would you dare to go into traffic when there are self-driving vehicles that anticipate situations correctly in 98% of the cases? If tax evasion could be predicted with 98% certainty, should every person that comes out of the system be preventatively arrested? Of course, you should take data-supported advice seriously and weigh it in your final decision. However, be aware of the margin of error and do not blindly trust a prediction: do additional research and try to interpret the results.
Prediction models are as good as the data used to train them. Imagine you are teaching an AI model to distinguish cats from dogs and you then show it a picture of a fox: the model will not know what to do with it. In many cases, this ‘bias’ is however not as clear as in this example; it can be difficult to discover. Even an apparently perfect data set can produce confusing results, e.g. because a certain category is under-represented. A bias in a data set can also result from a bias in the knowledge or beliefs of the person that created the data set. As the presence of a bias is not always that clear, it is even more important to be able to recognize it. Take a web shop that wants to predict what type of shoes a certain customer likes so as to be able to make better recommendations. To train the machine learning algorithm, the customer’s purchasing history and the purchasing history of other customers is used. As many of the previous purchases were made by women, the training data represent more female than male preferences. That creates a bias in the data and thus also in the algorithm that was trained with this data. Eventually, this will produce a situation in which the algorithm’s recommendations for female customers are much better than the ones for male customers.
A few practical examples from predictors with a so-called bias are the antisemitic chatbot Tay or the LinkedIn search engine that has developed a preference for males. But in this case as well, it becomes more harmful when skewed predictors affect human lives. In the US, the police for example uses algorithms to ‘predict’ where to find criminal hotspots. Trained with historical crime data, these applications lead to an overrepresentation of police in poorer neighborhoods with a primarily black population. This in turn leads to the arrests of more black people, and this data is then entered back into the algorithm, creating a vicious cycle. Predictive policing, as this application is called, is used in the Netherlands as well.
The Black Box
On top of these two theoretical arguments, there is also a practical argument to not blindly trust the predictions of an algorithm. The fact is that it is often unclear how some of these algorithms reach their conclusions. Currently algorithms are even already used to predict whether someone is creditworthy or eligible for a job. With access to the underlying mathematical models that make these predictions, you could ascertain what kind of indicators are used by these systems. But increasingly often, the algorithms used are so complex that the choices they make can no longer be interpreted and therefore not checked. Not even by the people that built them. You might not worry about not knowing why an algorithm shows you a certain ad on the internet or how it determines whether you like a certain artist. A black box algorithm that works very well is useful, however if our goal is to learn more about a phenomenon, then we should put more effort in understanding how it draws its conclusions. Didn’t you always have to show your calculations on your math tests to prove that you understood how it worked?
Let’s take a medical example. In 2015, a deep learning algorithm was applied to a patient database with around 700,000 people to find patterns in it. Then the algorithm had to analyze the data of current patients to see what the algorithm had learned. The algorithm turned out to be capable of things including predicting, very accurately, when psychiatric patients would have a schizophrenic episode. That was a huge breakthrough for these patients, as thanks to the predictor the medication can now be administered before the episode starts instead of after it is too late. Mission successful, you might say. But how the algorithm reaches its conclusion is still a mystery, as is the actual cause of the episodes. What we know about the disorder has thus remained the same, bringing us no closer to preventing it.
Making a 100% valid and reliable prediction is unfortunately a utopia. There is always a margin of error, and models are as neutral as the data on which they are based. Therefore we use increasingly advanced models, to look for even stronger links and connections. The downside to that is that we can no longer follow the reasoning of these models at times, making their validity and reliability impossible to check. To get the most value out of AI, it is important that we acknowledge its limitations. So think about it, remain critical, and realize that sometimes a prediction should only be seen as a well-supported argument.