Assumptions, model disagreement, and rethink triggers, updated weekly before your PM, risk, or IC discussion. Currently accepting one macro thesis and one AI-economy thesis for July.
4-model average
building category history
loading
13pt spread
Consensus sits at 66% across the four models and is — (loading). Models are mixed at a 13pt spread, so the average conceals live disagreement worth inspecting below. Category confidence is LOW (building category history).
Confidence reflects category-level track record. Stability tracks estimate movement. Models shows whether the four agree.
Loading belief trajectory...
Deepfake detection technology has been improving rapidly, with significant advancements seen in frameworks like Google's MediaPipe and existing benchmarks indicating that accuracy rates have approached the mid-90s range. However, as generative models evolve, the synthetic videos created may also become more sophisticated, potentially outpacing detection efforts. Current evaluations by the DeepFake Detection Challenge show that while tools achieve high accuracy, they often falter against more complex, nuanced deepfakes.
While deepfake generation technology is advancing rapidly, so too is the development of detection methods. Current state-of-the-art detection methods can achieve accuracies exceeding 95% on controlled datasets. By 2028, continued research and investment in adversarial training and novel artifact detection techniques are likely to push these accuracies to over 99% for most current generation synthetic video, though this assumes a relatively stable pace of advancement in generative models.
Deepfake detection has shown rapid improvement, with recent benchmarks (2023-2024) demonstrating 85-95% accuracy on specific datasets, but generalization remains challenging as adversarial techniques evolve in parallel. The core difficulty is an adversarial arms race: detection models trained on generation method X often fail on new generation methods, and synthetic video generation (particularly with diffusion models) is advancing faster than detection. However, 99% accuracy is achievable in controlled settings, and 4 years provides sufficient time for ensemble methods, multimodal detection approaches, and potential regulatory standards to mature. The main limiting factor is whether "current generation synthetic video" remains static or continues evolving—if deepfakes continue advancing at historical rates, maintaining 99% accuracy against moving targets becomes substantially harder.
Current deepfake detectors achieve 85-92% accuracy on 2023-2024 synthetic videos per NIST FRVT and Deepfake Detection Challenge benchmarks, with a 3-4% annual improvement rate since 2021; however, generative video models like Sora and Kling have closed the detection gap by 8-12% in blind tests within 6 months of release, and adversarial training loops now require 2-4x more labeled data to maintain prior accuracy levels. Structural factors include rapid diffusion of open-source generators versus slower adoption of standardized detection APIs in platforms.