Deepfake Detection: What Works, What Doesn't, and Why It Matters

Deepfake detection tools work by identifying artifacts that current generation models leave behind — subtle frequency patterns, blinking anomalies, lighting inconsistencies. These artifacts are real, but they are also moving targets: every generation of models is specifically trained to eliminate the artifacts the previous detector caught. Any detection tool has a shelf life.

What detection tools actually do well

• Catching older or lower-quality synthetic media at scale — useful for content moderation backlogs.
• Providing a risk signal, not a definitive verdict — flag for human review, not auto-removal.
• Detecting re-compressed or edited synthetic media when the artifact footprint survives compression.
• Running quickly enough to pre-screen high-volume uploads.

Provenance is the better bet

Rather than detecting fakes after the fact, the content authenticity ecosystem focuses on provenance: was this content signed by a known camera, device, or creator at the time of capture? The Coalition for Content Provenance and Authenticity (C2PA) standard attaches a cryptographic manifest to media at creation. Tools like Adobe's Content Credentials and camera firmware from Sony and Nikon already implement it.

Practical steps for deployers

1. 1For content moderation: use detection tools as a triage flag to route to human review, not as a final verdict.
2. 2For publishing: require C2PA provenance on media you source from third parties.
3. 3For internal communications: watermark any AI-generated media your organization produces so it can be identified later.
4. 4For users: media literacy is the long-game — label AI-generated content clearly and consistently.

The big idea: detection buys time but provenance wins long-term. Build workflows that require content to carry its origin story rather than hoping a detector can reconstruct it later.

Synthetic media disclosure is moving from optional to required; the design of disclosure determines whether it actually protects audiences.

What AI does well here

• Implement C2PA content credentials so provenance travels with the file
• Design visible disclosure that matches the context (overlay text on video, label on image, audio disclosure on synthetic voice)
• Document the AI involvement in production (which parts were generated, which were edited, which were unaltered)
• Build disclosure into the asset workflow so it's automatic, not afterthought

What AI cannot do

• Substitute for legal review in regulated contexts (political advertising, FDA-regulated promotion)
• Make audiences read the disclosure if it's hidden
• Replace the editorial responsibility for accuracy

AI can assist with deepfake political advertising disclosure that travels with the asset across re-shares, but ethical and legal accountability stays with the humans deploying it.

What AI does well here

• Draft policy memos covering deepfake obligations.
• Generate vendor diligence checklists referencing political advertising.

What AI cannot do

• Substitute for counsel on jurisdiction-specific obligations.
• Resolve the underlying value tradeoffs between competing stakeholders.

AI can draft AI deepfake takedown requests that cite the right platform policy section, identify the harm class, and request a clear remedy.

What AI does well here

• Match the alleged harm to the specific platform policy clause being violated
• Produce parallel notices for several platforms in one pass

What AI cannot do

• Confirm that the disputed media is in fact AI-generated
• Predict how a platform's trust and safety team will rule

Courts now require provenance metadata, expert testimony, and chain-of-custody before admitting any media that could be AI-generated.

What AI does well here

• Surface metadata anomalies for review
• Compare frames against known reference clips
• Draft authentication checklists for counsel

What AI cannot do

• Render a final admissibility ruling
• Replace a qualified forensic expert's testimony
• Guarantee that a deepfake detector is correct