It was reported just this month that more than 30% of the data used by Google for one of their shared machine learning models were incorrectly labeled. Besides the model itself is full of errors, the training data that was used by that model was also full of errors. It's impossible to trust Google's model if it's full of human-induced errors that computers can't correct.

The MIT study from 2021 found that almost 6% of images in the ImageNet database were mislabeled, and Google isn't the only one with major data mislabeling. Further, the study found that 10 of the most commonly used computer vision, natural language, and audio datasets contained label errors. When the data used to train these models is so poor, how can we trust or use them? That data and those models can't be trusted.

There is no question that garbage in is garbage out when it comes to AI, and there is a significant amount of bad data in AI projects. You are making the same mistake as Google, ImageNet, and others. Data collection, aggregation, cleaning, and labeling constitute over 80% of the time spent on AI projects, according to Cognilytica research.No matter how much time is spent, mistakes are bound to happen, even if the data is high quality. It is impossible to get good results with bad data.

This has been the case for decades for data-oriented projects, and now it's a problem for AI projects, which are essentially data projects. AI relies heavily on data. The data from which learning must be derived is what drives AI and machine learning projects. The poor quality of data causes AI systems to fail when organizations move too quickly with their AI projects.  Having bad data will negatively affect your AI projects, so don't be surprised if you don't have good data.