Skip to main content
TopAIThreats home TOP AI THREATS
How-To Guide

How to Detect Deepfakes: A Practitioner Checklist

Step-by-step workflow for evaluating suspected deepfake video, audio, or images. Quick-reference checklists for visual inspection, audio analysis, provenance verification, and escalation guidance.

Last updated: 2026-03-21

Who this is for: Security professionals, communications teams, journalists, trust and safety teams, and anyone who needs to evaluate whether specific content is AI-generated or AI-manipulated.

What Deepfakes Are and Why They Matter

Deepfakes are AI-generated or AI-manipulated media — video, audio, or images — designed to convincingly depict events that did not occur or words that were not spoken. They are used in three primary threat contexts:

  • Financial fraud. Impersonation of executives or trusted contacts to authorize transactions. The Hong Kong deepfake CFO fraud used real-time multi-participant video deepfakes to steal $25.6 million. The UK energy company voice cloning attack used a cloned CEO voice to extract $243,000.
  • Impersonation and harassment. Non-consensual intimate imagery, identity theft, and reputation attacks. The Taylor Swift deepfake image incident accumulated 47 million views before platform removal.
  • Disinformation. Manipulated political content designed to influence public opinion. The Slovakia election deepfake audio spread during a pre-election moratorium when candidates could not publicly respond.

No single detection technique reliably identifies all deepfakes. Detection and generation are in a continuous arms race. This guide provides a layered evaluation workflow — combining manual inspection, automated tools, provenance checks, and procedural verification — that represents the current best practice.

For the underlying science — why these methods work, where they fail, and what the incident evidence shows — see the Deepfake Detection Methods reference page.

Threat patterns this guide addresses

This guide applies to three threat patterns in the TopAIThreats taxonomy:

Step 1: Pause — Do Not Act on Unverified Content

Before analyzing the content, ensure no action is taken based on it:

  • If the suspected deepfake requests action (transfer funds, share credentials, make a statement): stop and verify first
  • If it claims to be from a known person: do not respond through the same channel
  • If it is spreading virally: do not reshare, even to debunk — amplification aids the attacker

Step 2: Preserve the Evidence

Before running any analysis, document what you have:

Step 3: Check for Content Credentials (C2PA)

Upload the file to a C2PA verification tool to check for cryptographic provenance:

Step 4: Visual Inspection Checklist (Video and Images)

Examine the content for these indicators. Each is suggestive, not conclusive — multiple indicators together increase confidence.

Face and skin

Eyes

Lighting and shadows

Video-specific (play at 0.25x speed)

Step 5: Audio Inspection Checklist (Voice Calls and Audio)

Speech patterns

Breathing and noise

Voice quality

Step 6: Run Automated Detection (Triage Only)

Use one or more automated detection systems as a supporting signal. A negative result does not confirm authenticity.

Examples of systems used for deepfake detection include:

SystemBest forAccess
Intel FakeCatcherReal-time video analysisEnterprise (Intel hardware)
Hive ModerationPlatform-scale scanningAPI (commercial)
Deepware ScannerQuick individual video checksConsumer / API
Sensity AIThreat intelligence correlationEnterprise

For how these systems work and why they fail on novel generation methods, see Deepfake Detection Methods — Automated Detection Systems.

Step 7: Verify Out-of-Band (High-Stakes Contexts)

For any content that could drive a consequential decision:

Identity claims

Source claims

Step 8: Escalate When Necessary

Not all deepfakes require the same response. Escalation depends on context:

Financial impact

If the deepfake has been or could be used to authorize financial transactions:

If the deepfake could result in litigation, regulatory action, or evidentiary proceedings:

Organizational reputation

If the deepfake targets your organization’s leadership, brand, or public communications:

Election or political content

If the deepfake relates to elections, political figures, or public policy:

Social media / low-stakes

If the deepfake is circulating on social media and does not target your organization:

Quick Decision Tree

Suspected deepfake
├── Requesting action (money, credentials, statement)?
│   └── YES → STOP. Verify out-of-band (Step 7) BEFORE anything else.

├── Has Content Credentials (C2PA)?
│   ├── YES, intact from trusted source → Likely authentic.
│   └── NO → Continue analysis (Steps 4-6).

├── Multiple visual/audio indicators present?
│   ├── YES → Treat as suspected deepfake. Escalate per Step 8.
│   └── NO / UNSURE → Run automated detection. Verify out-of-band if high-stakes.

└── Low-stakes context (social media, not targeted)?
    └── Report to platform. Do not reshare.

Where This Guide Fits in AI Threat Response

This guide covers detection — evaluating whether specific content is AI-generated. It is one part of a layered response:

  • Detection (this guide) — Is this authentic? Evaluate specific media for signs of AI generation or manipulation.
  • Detection methodsHow does detection work? Technical reference on forensic analysis, automated systems, provenance, and their limitations.
  • PreventionCan we prove content is real? Establishing authenticity at the point of creation.
  • Organizational defenseCan we prevent harm even if detection fails? Verification protocols, training, and procedural controls.
  • Incident responseWhat do we do now? Response procedures when a deepfake attack succeeds.

What This Guide Does Not Cover