AI vs AI: ChatGPT can be surprisingly effective at spotting deepfakes, study finds!

While the rise of AI has delivered numerous benefits and potential advantages, it has also introduced some challenges, such as the emergence of deepfakes. These are synthetic media that use AI to manipulate or superimpose existing video, audio or images to produce hyper-realistic but fabricated content, presenting significant concerns around misinformation and deception.

But what if we turned the tide on AI-driven misinformation by deploying the same AI technologies that are revolutionising our world? What if we use ChatGPT to actually spot deepfake images?

A research team led by the University at Buffalo has applied large language models (LLMs) to the task of spotting deepfakes of human faces. Their study revealed that although LLMs currently lag behind state-of-the-art deepfake detection algorithms in performance, their natural language processing capabilities could make them a more practical detection tool in the future.
ChatGPT's unexpected talent for spotting deepfakes with semantic savvy
Trained on an expansive dataset comprising roughly 300 billion words from the internet, ChatGPT identifies statistical patterns and relationships between words to craft its responses. The latest iterations of ChatGPT and other large language models (LLMs) have also expanded their capabilities to include image analysis. These multimodal LLMs leverage extensive databases of captioned images to decipher the connections between words and visuals.

The Media Forensics Lab team embarked on an experiment to see if GPT-4 with vision (GPT-4V) and Gemini 1.0 could differentiate between real human faces and those generated by AI. They tested these models with thousands of images, both authentic and deepfake, asking them to spot any signs of manipulation or synthetic artefacts.

In their findings, ChatGPT achieved an accuracy rate of 79.5% in detecting synthetic artefacts in images produced by latent diffusion, and 77.2% accuracy on StyleGAN-generated images. This performance is on par with earlier deepfake detection methods, indicating that with the right prompts, ChatGPT can reasonably identify AI-generated images.

A key advantage of ChatGPT is its ability to articulate its reasoning in clear language. For instance, when analysing an AI-generated photo of a man wearing glasses, ChatGPT accurately noted. “the hair on the left side of the image slightly blurs” and “the transition between the person and the background is a bit abrupt and lacks depth”.

In contrast, existing deepfake detection models typically provide the likelihood of an image being real or fake without explaining the rationale behind their conclusions. Even when examining the model’s underlying mechanisms, some features remain beyond our comprehension.

This distinction arises from ChatGPT’s reliance solely on semantic knowledge. Traditional deepfake detection algorithms are trained on extensive datasets of labelled images to discern real from fake. However, LLMs utilise their natural language capabilities to develop a sort of common-sense understanding of reality—including typical facial symmetry and the authentic look of photographs.

Overall, the study concluded that ChatGPT's blend of semantic knowledge and natural language processing offers a more user-friendly tool for deepfake detection for both developers and end users.

With that said, there’s still a massive room for improvement. ChatGPT’s performance lags behind the latest deepfake detection algorithms, which boast accuracy rates in the mid- to high-90s. This shortfall is partly because LLMs miss signal-level statistical differences, which are often imperceptible to the human eye, but detectable by specialised algorithms for identifying AI-generated images.

The study was recently published on the arXiv preprint server and can be accessed here.