Synthetic Pasts (2024–2026) is a research programme on the historicity of generative AI. The project works with open-source pre-trained models — Stable Diffusion, Mistral, Llama, and others — fine-tuned on historical data spanning literary fiction, art and comics, and scientific images and texts from the 19th and 20th centuries. The question is what kinds of pasts these systems produce when asked to render history, and how those synthetic pasts compete or coexist with existing historical knowledge. The programme situates generative AI within a longer history of media technology rather than treating it as a rupture, and develops methods that are scalable to other humanities and social-science domains.