麻豆社

In a world where seeing is no longer believing, verification demands more than intuition. Deepfakes are becoming part of the everyday challenge of deciding what is real, whether you are a journalist working under deadline pressure, a policymaker thinking about public trust, or a member of the public trying to make sense of what you see online.

and has identified the proliferation of deepfakes as a particular problem area. That is why 麻豆社 Research & Development has been exploring authenticity and verification technologies, and why deepfake detection has become a key area of work for us. Our aim has been not only to build a strong detector, but to develop it in a way that reflects the needs of real-world verification.

We are not approaching deepfake detection in isolation. 麻豆社 R&D is already actively involved in provenance and authenticity work, including C2PA, and detection sits alongside that as part of a broader verification toolkit. Having a range of tools in our arsenal is necessary, especially in a world where no single method can answer every verification question on its own.

A broad toolkit matters because different technologies provide various kinds of evidence. Provenance can offer very strong confidence when it is present and intact, while . Detection is different. It is probabilistic rather than definitive, but it is often the most broadly usable signal because it can still be applied when provenance metadata or watermarking are absent. In practice, the goal is not to find a single silver bullet, but to combine complementary signals with human judgement.

Journalism has shaped our deepfake detection project from the beginning, rather than benchmark performance alone. We worked early with journalists to understand what would help in verification workflows. This included a user study with 麻豆社 Verify, trials in other 麻豆社 settings, and, more recently, collaboration with the 麻豆社’s internal training teams so that development and education could inform one another. Our aim was not simply to build a detector, but to understand what kind of evidence and interface would genuinely support editorial decision-making. That journalism-first approach has shaped not only the wider tool, but also how we think about what a high-performing detector needs to do in practice.

Our work clearly highlights that users do not just want a verdict. They want help to interpret it. That insight shaped the tool itself. Rather than treating detection as a standalone answer, we designed the system to show multiple forms of evidence. Alongside model output, the deployed version includes checks and external verification signals such as , allowing users to build a fuller picture when authenticity is uncertain. The goal is not to replace human judgement, but to strengthen it with better evidence.

Seen in that light, explainability has been a major theme throughout the project, but it has also proved to be one of the hardest problems. We recognised early on that users needed help interpreting the verdict, yet explainability means different things depending on who you are talking to and how technically confident they are. For some users, it means understanding what the model may be responding to. For others, it means knowing how much weight to place on a result and what to do next. There is no single explanation format that works equally well for everyone.

We have experimented with a range of technical approaches to support interpretation, including concept bottleneck models, LLM-based explainers, gradient-based visualisation, and other methods. Some of these are promising research directions, but in practice we have not yet found them reliable enough. Hallucination, inconsistency, and low reliability are serious problems when the aim is to support real decisions rather than produce plausible-sounding text. That is why we have tried to stay disciplined about what our models do and do not claim. Explainability is a genuine requirement, but it is not yet a solved feature.

One direction we have found more useful is partial manipulation detection. Rather than simply outputting whether an image is AI-generated or AI-manipulated, this approach can highlight the regions most likely to have been altered. That gives users something more objective and inspectable than a fluent narrative explanation. It does not remove uncertainty, but it can help people understand why a result may have been triggered and where further scrutiny might be needed. This is an area we have pursued in collaboration with the and presented at one of the largest AI conferences, NeurIPS.

Yet surfacing evidence is only part of the challenge. Users also need to understand what the tool is showing them and what it is not showing them. That is why our work has also included research with academics from on how users perceive AI-assisted media detection and how the interface can better support interpretation. Users don鈥檛 always know what a detection output means, and a tool can only be used responsibly if its limitations are as clear as its strengths. At the same time, it is important to emphasise that this work sits on top of a strong core detection capability. Our efforts around interface, interpretation, and explainability are not there to compensate for a weak detector. They are about making a strong detector more usable, more responsible, and more valuable in practice.

Alongside the model and the user-facing work, we have also been contributing to the wider research ecosystem. As part of developing the detector, we created the 麻豆社-PAIR dataset, short for Paired Authentic and Inpainted References. Designed to support the training and evaluation of partial manipulation detection, 麻豆社-PAIR contains images that are partially manipulated by multiple state-of-the-art generative models, allowing detectors to learn not only whether an image may have been altered, but also where that alteration is likely to be. It is currently one of the largest and most comprehensive datasets of its kind, and . That matters because access to strong, representative data remains one of the major bottlenecks in deepfake detection. Building a strong internal capability matters, but helping to strengthen the broader ecosystem matters too.

One challenge in this space is that there is still no single standard way to evaluate and compare deepfake detection models. Performance can vary depending on the dataset, the manipulation type, and the evaluation setup, which makes competing claims difficult to judge consistently. , while government-backed challenge activity could be an important step towards more credible and independent evaluation. 麻豆社 R&D has contributed to that effort over time, first in an advisory capacity to the earlier challenge process in 2024, and later by participating directly with our own model in the latest iteration of the challenge in 2026.

One important milestone in the project was our participation in the , where our detector was among the strongest performers in the image detection category in terms of accuracy. That result is strong, independent validation of the core image detection capability itself. But the wider significance of the project lies in what we are building around that model: a journalism-first verification capability shaped by user needs, complementary evidence, and ongoing work on interpretation and explainability. We had also begun developing a video detector, but it was not ready to share at the time of the challenge. The challenge validated the core image detector performance. The wider project is about building the capability around it.

The deepfake detection project is about more than a single benchmark result. It combines a strong core detector with a journalism-first approach that makes our work more useful, more interpretable, and more valuable in practice. It helps to build a verification capability that combines complementary signals, reflects journalistic needs, treats explainability as a serious open problem, and contributes data and research to others researching in this field. The future of verification will not be built on one signal alone, nor on model scores alone. It will depend on how well we combine evidence, context, and human judgement. An important next step for us is to strengthen the video detection capability, building on the stronger progress we have already made in image detection.