Audrey Woods, MIT CSAIL Alliances | October 19, 2026
It’s never been harder to know if digital content is real. Generative AI tools have made it easy to create images, audio, and video for all kinds of applications, both positive and negative. On the one hand, there’s commercial value in being able to quickly make media that is convincing, appealing, and eye-catching. But, as with all technology, there is a downside, especially for banks, insurers, and payment providers. Deloitte's Center for Financial Services projects that generative AI could push fraud losses in the United States to $40 billion by 2027, up from $12.3 billion in 2023.
Graduate Student Ricardo Carrillo Silva says, “the volume of deepfakes being created will inevitably grow. In the last five years, you've gone from something blurry and inconsistent to images that are extremely photorealistic. People genuinely can't tell. Eventually it'll reach a point where these look completely real.”
Working under Dr. Amar Gupta, Carrillo is developing methods for deepfake detection that will—he hopes—stand the test of time. He says, “we want to solve the problem for the foreseeable future, not just for the best model at the moment.”
GENERALIZED DEEPFAKE DETECTION: A COMBINATION OF MODELS
“The idea behind the current project,” Carrillo explains, “is to detect deepfakes in the most generalizable way possible.” Deepfake detection models are getting better, but each one has their own methodology and, therefore, points of failure. “It varies from model to model. Some focus on specific parts of the body and look at forensic traces of how those differ from a real image. Some compare real and fake images in terms of specific noise traces. Others broadly look at what the overall image looks like.” Models are also often trained on one specific kind of generation engine, like generative adversarial networks (GANs), diffusion models, or rectified flow models. Silva’s project aims to combine deepfake detection models in an ensemble that balances out the individual weaknesses of each and offers a more robust method to detect generated images.
Carrillo and his team were inspired by the way LLMs can be made domain-specific by splitting the final neural network layer into specialized experts. For example, an educational LLM chatbot could have its last layer split into four experts for biology, math, chemistry, and history. This would allow the model to route specific questions through the relevant experts and combine the result, saving latency—since the question doesn’t have to be run through the whole model every time—and offering more specific answers. “Our pipeline works similarly: an input deepfake is sent to an initial routing module, which might send it to the GAN specialist or the diffusion specialist, or both if it looks like it could be either. Those experts produce scores, which are then combined into a meaningful output.”
An interesting finding in the preliminary stages of this project is how correlated the models are. “Models you'd think are quite different, when you look at their scores on the same set of images, there's a non-negligible correlation. They have a similar idea of what a real/fake image looks like, which is surprising.” This bodes well for the project’s goal of generalizability, since “if we find some relationship between how models fail or perform cross-family, we can build a combination model that potentially generalizes to newer types of generation, leveraging past detection techniques to catch new deepfakes.”
WHY GENERALIZABILITY MATTERS FOR FINANCIAL INSTITUTIONS
Banks and financial service providers already feel the brunt of missed deepfakes, as synthetic media is increasingly used to target the systems that verify a customer's identity. More than 40% of financial professionals have already encountered deepfakes used in fraud attempts, and roughly 90% report that fraudsters are actively using generative AI in their operations. Recent industry reporting puts deepfake fraud at about 6.5% of all fraud attempts globally, up from 0.1% in 2023, with injection attacks—where manipulated images or video are fed straight into a verification pipeline—rising 40% year over year.
This is precisely the environment a generalized detector is built for. A bank cannot know in advance which generation engine an attacker will use, and retraining a single-family detector after every new model release is neither fast nor affordable. An ensemble that routes an input to the right specialists—and that degrades gracefully when confronted with a generation method it has never seen—maps closely to how fraud and trust-and-safety teams operate: at scale, in real time, and against an adversary that updates faster than procurement cycles do.
THE FUTURE OF DEEPFAKES: A DEFENSIVE ADVANTAGE
Deepfake detection and generation can look like a cat-and-mouse game, but Carrillo is cautiously optimistic that detection has the structural advantage. "My guess is that generative AI companies first prioritize making images more photorealistic for the human eye, and second, try to fool the best detection models. I think it's harder to develop a new generative method than it is to build a detector for a specific type. So the detector side has an advantage there."
Despite the enormous buzz around generative AI tools, Carrillo understands the reservations people might have about this particular use of the technology. "Deepfakes are generally talked about as a net negative. I've only seen a few instances of generated images being used for something positive, mostly low-cost media." He wonders if that might shift as the field pushes into video, the next frontier for generative AI. "Hopefully we can find a way to make this a net positive. It's largely been used for net negatives, but I think a positive use case is possible. We just haven't found it yet." Carrillo also expects the technology's rapid improvement to eventually force a policy response. "I think we'll see some governmental regulation around this type of media, maybe mandatory watermarks. People just need to be careful about what this technology is used for."
For now, what keeps him motivated is the chance to work on something with direct, timely stakes. "It's nice to be the architect of your own work and have control over it. The ability to forge your own path through the research, and the potential to generate some positive impact if this works, that's a real motivator.”
