Invisible watermarking involves inserting a hidden, imperceptible signal into digital content—such as images, audio, or text. Long used as a tool for copyright protection, it has become an essential component of transparency and security in AI systems: today, it is one of the few ways to make AI-generated content detectable and traceable. This presentation introduces Content Seal OSS, the suite of open-source watermarking models developed at FAIR (Meta), which covers all three modalities: PixelSeal for images and video, AudioSeal for audio, and TextSeal for LLMs. We will explore where the watermark can be inserted into a generation pipeline (during post-processing of the output, or natively within the model) and what each choice entails in terms of latency, computational cost, and integration effort. Beyond research, we will discuss what is actually deployed in production and the constraints that drive these decisions, followed by the limitations of the approach: robustness to transformations, security against an adaptive adversary, and large-scale detection. We will conclude by discussing the relationship with other provenance standards and emerging regulations such as the AI Act. Description (short version, depending on available space in the program): Invisible watermarking inserts an imperceptible signal into an image, sound, or text, and is currently one of the few ways to make AI-generated content traceable. This session presents Content Seal, the open-source suite of watermarking models developed at FAIR (Meta) for all three modalities, the options for integrating it into a generation pipeline (latency, cost, robustness), and the limitations of the approach when faced with an adversary.