28 February 2024 in Journal of Online Trust and Safety
The increasing reach of deepfakes raises practical questions about people’s ability to detect false videos online. How vulnerable are people to deepfake videos? What technologies can help improve detection? Previous experiments that measure human deepfake detection historically omit a number of conditions that can exist in typical browsing conditions. Here, we operationalized four such conditions (low prevalence, brief presentation, low video quality, and divided attention), and found in a series of online experiments that all conditions lowered detection relative to baseline, suggesting that the current literature underestimates people’s susceptibility to deepfakes. Next, we examined how AI assistance could be integrated into the human decision process. We found that a model that exposes deepfakes by amplifying artifacts increases detection rates, and also leads to higher rates of incorporating AI feedback and higher final confidence than text-based prompts. Overall, this suggests that visual indicators that cause distortions on fake videos may be effective at mitigating the impact of falsified video.
3 December 2025 in bioRxiv
Recent large-scale vision fMRI datasets have been invaluable resources to the vision neuroscience community for their deep sampling of individual subjects and diverse stimulus sets. However, practical limitations to the number of subjects, stimuli, and trials that can be collected prevent individual fMRI datasets from reaching the scale necessary for modern modeling approaches and robust conclusions. Here, we introduce MOSAIC (Meta-Organized Stimuli And fMRI Imaging data for Computational modeling), a fMRI dataset aggregation framework designed to leverage the richness of individual datasets for computationally intensive modeling and robust tests of generalization. MOSAIC is composed of eight large-scale vision fMRI datasets totaling 93 subjects, 430,007 fMRI-stimulus pairs, and 162,839 naturalistic and artificial stimuli. A shared fMRI preprocessing pipeline and a filtered test-train split minimizes dataset-specific confounds and test-set leakage when aggregating the datasets. Crucially, additional datasets can be integrated into MOSAIC post hoc, allowing MOSAIC to evolve according to the community’s interests. We use MOSAIC to show that perceptually diverse stimulus sets consistently improve decoding accuracy and stability, carrying implications for future fMRI stimulus set design. We then jointly train brain-optimized encoding models across subjects and datasets to predict fMRI activity of all visual cortex and even the whole brain. In silico functional localizer experiments performed on these digital twin models can recover subject-specific category-selective cortical regions, thereby validating our approach. Together, MOSAIC provides a scalable and community-driven solution to build robust, larger-scale models of human vision.