RealorRender: Germany's new deepfake detector that shows its work
Germany's Fraunhofer Institute has released RealorRender, a deepfake detection tool that achieves 85–91% accuracy and, unusually, shows users precisely which parts of an image gave it away. Developed with funding from the German Federal Office for Information Security (BSI), the system arrives as both the UK and EU tighten laws around AI-generated content — and as platforms scramble for credible detection technology.
How it works
RealorRender uses two methods in tandem. The first is standard deep learning: the algorithm classifies images based on known patterns left by AI generators. The second is more distinctive — it measures what developers call "reconstruction error." The system attempts to regenerate a suspect image using its own model. If the output matches too closely, that's a strong signal the original was machine-made rather than captured by a camera. This hybrid approach, detailed by Fraunhofer IOSB, is what pushes accuracy into the 85–91% range across tested datasets.
XAI explanation for an image in which a person was detected as a deepfake. Illustration: Fraunhofer IOSB
The explainability angle
Most current detectors return a verdict with no reasoning attached. RealorRender uses Explainable AI (XAI) to produce a heat map of the flagged image, highlighting specific regions, textures, and artifacts the model found suspicious — think microscopic lighting inconsistencies or geometric patterns characteristic of generative models. That transparency matters beyond curiosity: XAI outputs can meet evidence standards needed for legal prosecution, a direct fit for the UK's tightening regulatory environment.
Why this matters right now
The timing is pointed. The UK criminalised the creation of non-consensual intimate images in February 2026 under Section 138 of the Data (Use and Access) Act. Platforms now face a 48-hour removal window and potential Ofcom fines of up to 10% of global turnover. The UK DSIT Deepfake Detection Report identified just seven domestic providers — mostly pre-seed — against 23 US competitors, with an average funding level of around £25 million. The scale of the problem is stark: in 2023 alone, over 276,000 pornographic deepfakes were found on a single platform, 96% targeting women without consent.
On the EU side, Article 50 of the AI Act takes effect in August 2026, requiring platforms to detect and transparently label AI-generated content. RealorRender's XAI output is built to satisfy exactly that kind of transparency requirement.
What's still unclear
No commercial release date or licensing model has been announced. Fraunhofer has not confirmed API partnerships with major platforms, so whether Facebook, YouTube, or UK-based services will actually integrate RealorRender remains open. Its performance against next-generation generators — tools that may be specifically designed to evade detectors — has not been publicly quantified either. For now, it's a credible, publicly-funded prototype in a market that badly needs one.