At the AI Festa 2026, ThinkForBL, a company that specializes in evaluating the reliability of AI systems, highlighted a growing concern in the field: not all data are equal when training artificial intelligence models. While many companies aim to use the largest possible volume of data to train their AI, this approach can be misleading. ThinkForBL demonstrated how models trained on different datasets—such as one with 50,000 images and another with just 2,000—can yield varying results. The goal of the demonstration was not to suggest that smaller datasets are always better, but to emphasize that the quantity of data is not the only factor that determines the quality of an AI model. The company pointed out that issues such as duplicate data, overrepresented categories, or examples that don't reflect real-world scenarios can affect how well a model performs. This means the problem isn't necessarily with the model itself, but rather with the dataset it was trained on. ThinkForBL specifically introduced the concept of "contextual bias," which occurs when individual data points are accurate but collectively do not represent the environment in which the AI will be used. As a result, a model might perform well in controlled tests but behave unpredictably when faced with new or less common situations. To address these challenges, ThinkForBL developed a tool called Reeinn, which analyzes the data used to train AI systems. Reeinn focuses on identifying redundancies and imbalances that could lead to unreliable performance in real-world applications. This approach reflects a broader shift in the AI industry, where the focus is no longer solely on a model's success rate but also on its robustness—its ability to perform consistently even when faced with unfamiliar data. This shift is especially critical in high-stakes environments such as public services, security, and defense, where AI errors can have serious consequences. Ensuring that AI systems remain reliable when deployed outside of controlled development settings is a key challenge. As a result, the evaluation of AI projects may need to evolve beyond comparing models based on performance benchmarks. Instead, the data used for training, the testing environments, and the models’ ability to maintain consistent performance in changing conditions will become more important. The adoption of international reliability standards could also play a crucial role in establishing common criteria for assessing AI systems across different providers.