Estimating Visual Attribute Effects in Advertising from Observational Data: A Deepfake-Informed Double Machine Learning Approach

A team of researchers has published a novel AI framework, DICE-DML (Deepfake-Informed Control Encoder for Double Machine Learning), that tackles a fundamental challenge in digital marketing: measuring the true causal impact of specific visual elements in advertisements. Traditional causal inference methods fail because standard vision AI models entangle attributes like a person's appearance with background confounders, producing biased results. DICE-DML solves this by using generative AI to create deepfake-altered image pairs, allowing it to isolate the variation of a single attribute (e.g., changing only skin tone) while holding everything else constant. This creates a clean dataset where the 'treatment' effect can be measured without contamination.

The technical approach combines deepfake-generated image pairs with an adversarial learning process called DICE-Diff, which analyzes the difference vectors between paired images to cancel out background noise and reveal a pure 'treatment fingerprint.' A final orthogonal projection step geometrically removes treatment-axis components. In simulations, DICE-DML reduced root mean squared error by 73-97% versus standard Double Machine Learning, with a 97.5% improvement at the null effect point, demonstrating robust error control. Applied to a dataset of 232,089 Instagram influencer posts, the framework estimated the causal effect of skin tone on engagement (likes), finding a marginally significant negative effect for darker skin tones. This provides marketers and platforms with a rigorous, AI-powered tool to audit for visual bias and optimize creative content based on causal evidence, not correlation.