Human versus machine: How AI sees our world

We now use many AI tools in our everyday lives, from image recognition to speech processing. But does artificial intelligence see the world through the same "eyes" as we do? Researchers at the Max Planck Institute and Justus Liebig University Giessen investigated this question. They wanted to know whether AI perceives objects, buildings, or other living beings in the same way as humans, or whether the strategies differ. They used natural images to determine this.

The four study authors developed a method for identifying and comparing various key dimensions that humans and AI pay attention to when recognizing objects. These dimensions represent various properties of objects, from purely visual aspects like "round" or "white" to semantic properties like "animal-related" or "fire-related." Some combined both visual and semantic elements.

The analysis was based on 5 million human judgments and a deep neural network (DNN) model of nearly 2,000 object images – and revealed clear differences, which the team of scientists has now published in the journal "Nature Machine Intelligence." While humans primarily pay attention to the meaning of an object, asking themselves "what do I know about it?", AI is more focused on external features such as shape and color.

This visual bias leads to the AI ​​reaching similar results, but often being wrong upon closer inspection. For example, in an "animal-related dimension," the AI ​​excluded many images of animals and included some that didn't show any animals at all. These images then only showed images of nature, cages, or bars. This demonstrates that similar is not the same. And similar results are sometimes based on completely different paths that led to them.

Despite partially similar decision-making, humans and AI appear to pursue different thinking strategies. According to the researchers, this raises questions about the reliability and trustworthiness of AI systems.

The study thus provides not only insights into the "mindset" of artificial systems, but also comparisons with human approaches. This is valuable for research in areas such as computational cognitive neuroscience. According to the Max Planck Institute, this not only allows for a better understanding of how AI processes information, but also provides insights into human cognition and the design of safe and reliable AI systems.