What questions can I ask?

For reliable AI it is crucial to use good quality data. Without the right data, the algorithm cannot make the right analysis. In other words: ‘garbage in, garbage out’.

How do you know if the data is ‘right’? This depends on the goal of the algorithm. Data that is great for one application may be useless for another. Hence the question: where does the data come from? More information about how the data was collected and what it is representative of will help you determine whether the algorithm can be meaningful for your purpose. You can think of an algorithm as a recipe and the data as the ingredients. If you want to make cake you need eggs. Where do the eggs come from? The color of the eggs does not matter for the recipe, whether they are white or brown. But whether it concerns chicken eggs, quail eggs or ostrich eggs is relevant. If your goal was to find suitable eggs to paint, then color would be relevant and size would matter less.

An algorithm executes a series of instructions. If the instructions are incorrect, so is the outcome. Hence, we ask; what happens to the data? Put differently; what were the algorithm’s instructions?

Instructions are based on human choices and assumptions. The makers of the algorithm must be able to explain their choices. This is relatively easy with a simple algorithm and limited number of variables. The makers provide rules and instructions that the algorithm uses. With a self-learning algorithm it is already a lot more difficult. In this case, the makers do not provide the instructions, but let the algorithm figure out which variables together give the result. Yet, here too the makers make certain choices. They train the algorithm with different data sets and thus choose what the algorithm can and cannot ‘learn’. They cannot see how the algorithm learns, but they do determine the quality, robustness and how representative the data is.

AI makes it possible to analyze large amounts of data and recognize patterns. Next, you have to decide what to do with the result. What choices can you make based on the source and analysis used?

If you are about to get on your bike, a weather forecast that predicts 10mm of rain will fall today isn’t very helpful. You want a prediction of the exact time and location of a shower. But for a farmer, the amount of rain is much more important, while the time of the shower is less relevant. Different algorithms will therefore be useful for different purposes. It is important to check whether the outcome fits the purpose for which the algorithm is used. By asking how data and analysis fit the outcome, you gain more insight into how well the algorithm works in achieving your goal.