← TRACKER
THREAT VECTOR // HIGH

REALITY COLLAPSE

High  ·  CURRENT VALUE 76/100

Synthetic media indistinguishable from reality. Information ecosystem integrity degrading. Consensus reality fracturing under generated content.

Current status as of 2026-09-04

Synthetic media generation moved from research curiosity to consumer software between 2022 and 2025. The 2023 wave of text-to-image diffusion tools (Stable Diffusion, Midjourney, DALL-E 3) produced photorealistic imagery indistinguishable to untrained observers; the 2024 wave of video generation (Sora, Runway Gen-3, Google Veo) closed most of the remaining gap for short clips. Voice cloning requires under a minute of source audio and runs on consumer GPUs. The result is that “the video shows X” and “the audio shows Y” have lost their default evidentiary weight for public claims.

The 2026 signal is not whether deepfakes exist (they do) or whether detection works (it does not, in adversarial conditions) but what the deployment mix looks like. Election-year deepfakes in 2024 (Slovakia, US, Bangladesh, India) had measurable but limited effect — the “liar's dividend” (Chesney & Citron 2019) is proving more consequential than any specific fake. Real evidence of misconduct can now be dismissed as AI-generated with prima facie plausibility, and the burden of proving authenticity has shifted to the accuser. Non-consensual intimate imagery — 96% of deepfake video traffic per Sensity AI's 2019 baseline — continues to be the dominant harm by volume, with US federal legislation (TAKE IT DOWN Act) passed in 2024 but state-level enforcement uneven.

Historical trajectory

DateLevel
2020-06
25
2021-01
32
2021-06
38
2022-01
42
2022-06
48
2022-11
55
2023-03
60
2023-06
63
2023-11
66
2024-03
70
2024-08
72
2025-02
74
2025-06
75
2026-01
76

Key papers

  1. Deep Fakes: A Looming Challenge for Privacy, Democracy, and National Security
    Chesney & Citron · 2019 · California Law Review 107:1753
    The paper that named the “liar's dividend” and established the legal framework for the next five years of policy work. Every subsequent deepfake-policy article cites this.
  2. Generative Adversarial Networks
    Goodfellow et al. · 2014 · NeurIPS 2014 (arXiv:1406.2661)
    The original GAN paper. Not about deepfakes as harm, but the technical substrate that made them possible. Read to understand why detection is fundamentally adversarial: the discriminator that catches fakes is exactly the training signal the generator learns to beat.
  3. The State of Deepfakes: Landscape, Threats, and Impact
    Ajder, Patrini, Cavalli, Cullen (Sensity AI) · 2019 · Sensity AI report
    First systematic measurement of deepfake prevalence and content mix. Established the still-cited 96%-is-NCII figure. Baseline for any harm-mix discussion.
  4. High-Resolution Image Synthesis with Latent Diffusion Models
    Rombach, Blattmann et al. · 2022 · CVPR 2022 (arXiv:2112.10752)
    The Stable Diffusion paper. The moment high-quality image generation became consumer-accessible and open-weight, which is when the deployment surface for deepfakes changed permanently.

Strongest counterargument

The steelman: media has been manipulable since the invention of photography, and every previous generation of manipulation technology (retouching, Photoshop, CGI, cheapfakes) produced the same predictions of civic collapse without delivering them. Modern media literacy has actually improved, not degraded, viewer skepticism; younger cohorts trust image evidence less than their parents did, which is arguably the correct calibration. Political deepfakes in the 2024 election cycle underperformed predictions substantially, and the technology's biggest actual harm (NCII) is a specific misuse case with specific legal remedies, not a civilizational threat. The “liar's dividend” argument is unfalsifiable — any weakening of trust in evidence gets attributed to deepfakes even when the underlying trust erosion has other causes. Treating deepfakes as a category-level threat obscures the specific harms that deserve specific policy responses.

Related events (10)

Events from the tracker's timeline whose tags, title, or description match this vector. Heuristic auto-match; some may be tangential.

Dead Hand systems that amplify this vector (3)

Revision history