Over the past three decades, Topology Optimization (TO) has yielded new and surprising solutions across length scales and engineering disciplines, often outperforming conventional low-weight designs. However, although the number of convincing design examples is vast and growing, industry adoption remains relatively limited except in the aerospace and automotive sectors. Surveys of design practitioners indicate that the will to adopt is there, but inclusion in design practice is hindered by TO's large time commitment and black-box nature. To address both simultaneously, human-in-the-loop TO has recently been developed. In human-in-the-loop TO, the otherwise fully automated design method is interrupted to enable interactive human input. This way, the human user can infuse their preferences and expertise to enrich the design framework and guide the automated algorithm towards a solution they are likely to accept as ready for fabrication. The design problem is kept simple to enable fast computation, typically using a minimum-compliance formulation. Even so, successful human-in-the-loop demonstration examples include designs in which human input improves performance of complex criteria such as buckling, stress concentrations, manufacturability, and aesthetics. An inherent weakness of all human-in-the-loop approaches is that they rely on their users to make sound design decisions. This work aims to support good decision-making by developing an AI co-pilot for human-in-the-loop TO. The AI co-pilot used a Machine Learning (ML) model to predict where to make changes. It is formulated using a U-Net architecture. The model is trained on several topological characteristics, including the longest topological member and the most complex node. It is implemented within an existing human-in-the-loop TO framework as an AI co-pilot and demonstrated to improve metrics such as buckling performance and manufacturability.