The Ghost in the Machine - AI's Impact on Cultural Heritage (Research) Over the past decade, deep learning methods have made remarkable advancements. This progress can be attributed to various factors such as massive parallelization through the utilization of Graphics Processing Units (GPUs) for massive parallelization. This shift in hardware has significantly accelerated the training of deep neural networks, allowing researchers to tackle increasingly complex problems. Another critical factor contributing to the success of deep learning is the acquisition of vast training datasets sourced from the World Wide Web, which has become a treasure trove of information. As a result, these models have become adept at capturing intricate patterns and representations in various domains. Furthermore, the development of efficient and reusable neural network architectures has also played a crucial role in the advancement of deep learning. Putting everything together, these evolutions have paved the way for the achievement of human-like or even superhuman performance in specific domains. Notably, the emergence of pre-trained large language models has demonstrated the capability to grasp the intricate semantics of natural languages, yielding exceptional outcomes in classification, prediction, and generation tasks. Similarly, in the realm of image generation, models such as Stable Diffusion and Dall-E have showcased their prowess. Tasks that once demanded human expertise for their execution are now on the brink of being supported or entirely taken over by machine intelligence. In the subsequent sections, we will illuminate some recent breakthroughs in AI-assisted search and retrieval systems within the domain of cultural heritage. One such example is the development of a multimodal search system for Iconclass, incorporating vision-language pre-trained machine learning models. However, it is paramount to approach the application of these cutting-edge generative AI models in scientific and research contexts with due diligence. One must remain mindful of potential inaccuracies and hallucinations that these systems can inadvertently produce. It's worth noting that deep learning and large language models constitute only a specific subset of artificial intelligence, falling under the broader category of machine learning. Symbolic knowledge representation represents another distinct subdomain of AI, distinguished by its mathematical rigor and formalism. In this realm, any inaccuracies or inconsistencies in underlying assumptions can be readily identified and rectified. Knowledge graphs built upon ontologies present a viable avenue for enhancing the explainability of black-box statistical deep learning systems. Furthermore, they possess the capacity to flag false or counterfeit information. As a result, future information systems are poised to embrace hybrid solutions that amalgamate symbolic and subsymbolic AI approaches to combine the strengths of both paradigms, offering not only reliable but also trustworthy results.