Beyond Generation: Using Multimodal LLMs as Detectors of LLM-Generated Deepfake Facial Images

Authors: Olivia Salamon, Virginia N. L. Franqueira, Shujun Li

Published: 2026-10-15 12:16:09+00:00

AI Summary

This paper explores the use of Multimodal Large Language Models (MLLMs) to detect deepfake facial images, investigating whether they identify synthetic media through physical inconsistencies or digital artifacts like watermarks. The study benchmarks six MLLMs, revealing high detection accuracy for some models, with GPT-5.2 achieving 97.9% in pairwise comparison, primarily by recognizing structural inconsistencies. However, it also finds that invisible watermarks can reduce detection accuracy, suggesting they might suppress the very artifacts MLLMs rely on.

Abstract

As image generation models become increasingly capable of producing highly realistic content, traditional forensic techniques are struggling to keep up. This paper investigates whether MLLMs can instead be used as a form of forensic analysis by identifying synthetic media. Specifically, the study aims to determine whether these models detect fake images by recognising physical inconsistencies, or whether they rely on digital artefacts such as embedded watermarks.Six MLLMs: GPT-4o, GPT-5, GPT-5.2, Gemini 2.5, Gemini 3, and Seed 1.6, are benchmarked using a dataset of 400 images. Fifty images are authentic photographs, while the remaining 350 are generated using seven leading image models: Gemini 3, Gemini 2.5, Flux.2 Pro, GPT-Image-1, GPT-Image-1.5, Seedream 4.0, and Seedream 4.5. The evaluation is structured in three stages: single-image classification, pairwise comparison, and a focused self-detection experiment on Gemini. Several MLLMs achieved a high detection accuracy, with GPT-5.2 reaching a peak of 97.9% in the pairwise experiment, against specific generators and an overall mean pairwise accuracy of 82.5% by identifying structural or physical inconsistencies within images. However, results from the self-detection experiment indicate that the presence of invisible SynthID watermarks leads to a 10% reduction in Gemini's detection accuracy compared to the clean control group. One possible explanation is that SynthID embedding reduces some of the visible artefacts the models rely on for detection.These results imply that MLLMs can detect structural artefacts without relying on watermarks, but performance varies substantially across model generations. The findings from comparison C5 also suggest that SynthID embedding may suppress some of the structural artefacts used during detection.

Key findings

MLLMs can effectively detect structural artifacts in deepfake images, with GPT-5.2 showing superior performance (97.9% in pairwise). The presence of invisible SynthID watermarks can paradoxically reduce detection accuracy by up to 10% for Gemini models, possibly by masking structural inconsistencies. The study also found that MLLMs did not hallucinate justifications, instead providing reasonable, albeit sometimes incorrect, reasoning.

Approach

The authors benchmark six MLLMs (GPT-4o, GPT-5, GPT-5.2, Gemini 2.5, Gemini 3, and Seed 1.6) using a dataset of 400 images (50 authentic, 350 AI-generated). They evaluate detection performance across three stages: single-image classification, pairwise comparison, and a targeted self-detection experiment on Gemini models to assess the impact of watermarks.

Datasets

Flickr-Faces-HQ (FFHQ) for authentic images; generated images from Gemini 3, Gemini 2.5, Flux.2 Pro, GPT-Image-1, GPT-Image-1.5, Seedream 4.0, and Seedream 4.5.

Model(s)

GPT-4o, GPT-5, GPT-5.2, Gemini 2.5, Gemini 3, Seed 1.6

Author countries

← Previous