The Deepfakes We Missed

Deepfake detection research has largely focused on a threat model inherited from the 2017–2019 wave of public concern: face-swap and talking-head manipulation targeting politicians, celebrities, and public figures. This paper argues that the dominant harms that ultimately emerged between 2022 and 2026 differ substantially from those assumptions. Real-world incidents are now concentrated around peer-generated non-consensual intimate imagery (NCII), voice-clone scam calls, emotional-manipulation fraud, and private messaging-based distribution. Meanwhile, benchmark design, datasets, and detection systems remain heavily concentrated on public-figure video manipulation. We present a large-scale classification of 438 papers published between 2017 and 2025 across five threat categories and compare research allocation against observed harm distributions. Our analysis suggests that the primary bottleneck in practical deepfake defense is no longer model capability alone, but a persistent mismatch between research priorities and deployed harms. We further identify structural causes behind this misalignment and outline three concrete research agendas for under-defended categories.

1. A classification of 438 deepfake detection papers published between 2017 and 2025 across five threat categories.
2. An empirical comparison between research allocation and observed real-world harms using evidence synthesized from IC3, IWF, AIID, and victim-centered reporting systems.
3. A structural analysis explaining why research misalignment persists, including benchmark inheritance, dataset asymmetry, and visibility-driven prioritization.
4. Three research agendas targeting under-defended harm categories with strong societal relevance and technical feasibility.

@inproceedings{raza2026deepfakeswemissed,
  title  = {The Deepfakes We Missed:
            We Built Detection Systems
            for a Threat Model That Never Fully Materialized},
  author = {Raza, Shaina},
  year   = {2026}
}