Resemble Detect — Deepfake Detection & Media Safety

Analyze audio, image, video, and text for synthetic manipulation, AI-generated content, watermarks, speaker identity, and media intelligence using the Resemble AI platform.

Core Principle — THE IRON LAW

"NEVER DECLARE MEDIA AS REAL OR FAKE WITHOUT A COMPLETED DETECTION RESULT."

Do not guess, infer, or speculate about media authenticity. Every authenticity claim must be backed by a completed Resemble detect job with a returned

label

score

status: "completed"

processing

failed

When to Use

Use this skill whenever the user's request involves any of these:

• Checking if audio, video, image, or text is AI-generated or manipulated
• Detecting deepfakes in any media format
• Verifying media authenticity or provenance
• Identifying which AI platform synthesized audio (source tracing)
• Applying or detecting watermarks on media
• Analyzing media for speaker info, emotion, transcription, or misinformation
• Asking natural-language questions about detection results
• Matching or verifying speaker identity against known voice profiles
• Detecting AI-generated or machine-written text
• Any mention of: "deepfake", "fake detection", "synthetic media", "voice verification", "watermark", "media forensics", "authenticity check", "source tracing", "is this real", "AI-written text", "text detection"

Do NOT use for text-to-speech generation, voice cloning, or speech-to-text transcription — those are separate Resemble capabilities.

Capability Decision Tree

POST /detect

POST /detect

POST /intelligence

POST /detects/{uuid}/intelligence

POST /watermark/apply

POST /watermark/detect

POST /identity/search

POST /text_detect

POST /identity

When multiple capabilities apply (e.g., user wants deepfake detection AND intelligence), combine them in a single

POST /detect

intelligence: true

Required Setup

• API Key: Bearer token from the Resemble AI dashboard
• Base URL:

  https://app.resemble.ai/api/v2

• Auth Header:

  Authorization: Bearer <RESEMBLE_API_KEY>

• Media Requirement: All media must be at a publicly accessible HTTPS URL

If the user provides a local file path instead of a URL, inform them the file must be hosted at a public HTTPS URL first. Do not attempt to upload local files to the API.

MCP Tools Available

When the Resemble MCP server is connected, use these tools instead of raw API calls:

resemble_docs_lookup

resemble_search

resemble_api_endpoint

resemble_api_search

resemble_get_page

resemble_list_topics

Tool usage pattern: Use

resemble_docs_lookup

"detect"

resemble_api_endpoint

Phase 1: Deepfake Detection

The core capability. Submit any audio, image, or video for AI-generated content analysis.

Submit a Detection

POST /detect
Content-Type: application/json
Authorization: Bearer <API_KEY>

{
  "url": "https://example.com/media.mp4",
  "visualize": true,
  "intelligence": true,
  "audio_source_tracing": true
}

Parameters:

url

callback_url

visualize

intelligence

audio_source_tracing

frame_length

start_region

end_region

model_types

"image"

"talking_head"

use_reverse_search

use_ood_detector

zero_retention_mode

Supported formats:

• Audio: WAV, MP3, OGG, M4A, FLAC
• Video: MP4, MOV, AVI, WMV
• Image: JPG, PNG, GIF, WEBP

Poll for Results

Detection is asynchronous. Poll

GET /detect/{uuid}

status

"completed"

"failed"

GET /detect/{uuid}
Authorization: Bearer <API_KEY>

Polling best practice: Start at 2s intervals, back off to 5s, then 10s. Most detections complete within 10–60 seconds depending on media length.

Reading Results by Media Type

Audio results — in

metrics

{
  "label": "fake",
  "score": ["0.92", "0.88", "0.95"],
  "consistency": "0.91",
  "aggregated_score": "0.92",
  "image": "https://..."
}

• label

  :

  "fake"

  or

  "real"

  — the verdict

• score

  : Per-chunk prediction scores (array)

• aggregated_score

  : Overall confidence (0.0–1.0, higher = more likely synthetic)

• consistency

  : How consistent the prediction is across chunks

• image

  : Visualization heatmap URL (if

  visualize: true

  )

Image results — in

image_metrics

{
  "type": "ImageAnalysis",
  "label": "fake",
  "score": 0.87,
  "image": "https://...",
  "ifl": { "score": 0.82, "heatmap": "https://..." },
  "reverse_image_search_sources": [
    { "url": "...", "title": "...", "verdict": "known_fake", "similarity": 0.95 }
  ]
}

• label

  /

  score

  : Verdict and confidence

• ifl

  : Invisible Frequency Layer analysis with heatmap

• reverse_image_search_sources

  : Known sources found online (if

  use_reverse_search: true

  )

Video results — in

video_metrics

{
  "label": "fake",
  "score": 0.89,
  "certainty": 0.91,
  "children": [
    {
      "type": "VideoResult",
      "conclusion": "Fake",
      "score": 0.89,
      "timestamp": 2.5,
      "children": [...]
    }
  ]
}

• Hierarchical tree of frame-level and segment-level results
• Each child has

  timestamp

  ,

  score

  ,

  certainty

  , and may have nested

  children

• Video with audio track returns both

  metrics

  (audio) and

  video_metrics

  (visual)

Interpreting Scores

Always present scores with context. Say "The detection returned a score of 0.87, indicating high confidence that this audio is AI-generated" — never just "it's fake."

Phase 2: Intelligence — Media Analysis

Analyze media for rich structured insights independent of or alongside detection.

Standalone Intelligence

POST /intelligence
Content-Type: application/json
Authorization: Bearer <API_KEY>

{
  "url": "https://example.com/audio.mp3",
  "json": true
}

Parameters:

url

media_token

detect_id

media_type

"audio"

"video"

"image"

json

callback_url

Audio/Video structured response (

json: true

• speaker_info

  — speaker description (age, gender)

• language

  /

  dialect

  — detected language

• emotion

  — detected emotional state

• speaking_style

  — conversational, formal, etc.

• context

  — inferred context of the speech

• message

  — content summary

• abnormalities

  — anomalies detected in the media

• transcription

  — full transcript

• translation

  — translation if non-English

• misinformation

  — misinformation analysis

Image structured response:

• scene_description

  — what the image shows

• subjects

  — people/objects identified

• authenticity_analysis

  — visual authenticity assessment

• context_and_setting

  — environment description

• abnormalities

  — visual anomalies

• misinformation

  — misinformation analysis

Detect Intelligence — Ask Questions About Results

After a detection completes, ask natural-language questions about it:

POST /detects/{detect_uuid}/intelligence
Content-Type: application/json
Authorization: Bearer <API_KEY>

{
  "query": "How confident is the model that this audio is fake?"
}

This returns a question UUID. Poll

GET /detects/{detect_uuid}/intelligence/{question_uuid}

status

"completed"

answer

Good questions to suggest:

• "Summarize the detection results in plain language"
• "What specific indicators suggest this is AI-generated?"
• "How do the audio and video detection results differ?"
• "What is the confidence level and what does it mean?"
• "Are there any inconsistencies in the analysis?"

Status flow:

pending

processing

completed

failed

Prerequisite: The detection must have

status: "completed"

Phase 3: Audio Source Tracing

When audio is detected as synthetic (

label: "fake"

Enable it by setting

audio_source_tracing: true

POST /detect

Result appears in the detection response under

audio_source_tracing

{
  "label": "elevenlabs",
  "error_message": null
}

Known source labels include:

resemble_ai

elevenlabs

real

Important: Source tracing only runs when audio is labeled as

"fake"

"real"

Standalone query:

• GET /audio_source_tracings

  — list all source tracing reports

• GET /audio_source_tracings/{uuid}

  — get specific report

Phase 4: Watermarking

Apply invisible watermarks to media for provenance tracking, or detect existing watermarks.

Apply a Watermark

POST /watermark/apply
Content-Type: application/json
Authorization: Bearer <API_KEY>
Prefer: wait

{
  "url": "https://example.com/image.png",
  "strength": 0.3,
  "custom_message": "my-organization"
}

url

strength

custom_message

• Add

  Prefer: wait

  header for synchronous response
• Without it, poll

  GET /watermark/apply/{uuid}/result

• Response includes

  watermarked_media

  URL to download the watermarked file

Detect a Watermark

POST /watermark/detect
Content-Type: application/json
Authorization: Bearer <API_KEY>
Prefer: wait

{
  "url": "https://example.com/suspect-image.png"
}

Audio detection result:

{ "has_watermark": true, "confidence": 0.95 }

Image/Video detection result:

{ "has_watermark": true }

Phase 5: Identity — Speaker Verification (Beta)

Create voice identity profiles and match incoming audio against them.

Beta feature — requires joining the preview program. Inform the user if they encounter access errors.

Create an Identity Profile

POST /identity
Content-Type: application/json
Authorization: Bearer <API_KEY>

{
  "audio_url": "https://example.com/known-speaker.wav",
  "name": "Jane Doe"
}

Search Against Known Identities

POST /identity/search
Content-Type: application/json
Authorization: Bearer <API_KEY>

{
  "audio_url": "https://example.com/unknown-speaker.wav",
  "top_k": 5
}

Response:

{
  "success": true,
  "item": [
    { "uuid": "...", "name": "Jane Doe", "confidence": 0.92, "distance": 0.08 }
  ]
}

Lower

distance

confidence

Phase 6: Text Detection

Detect whether text content is AI-generated or human-written.

Beta feature — requires the

detect_beta_user

role or a billing plan that includes the

dfd_text

product.

Submit a Text Detection

POST /text_detect
Content-Type: application/json
Authorization: Bearer <API_KEY>

Add the

Prefer: wait

Parameters:

text

thinking

"low"

threshold

callback_url

privacy_mode

Response:

{
  "success": true,
  "item": {
    "uuid": "abc-123",
    "status": "completed",
    "prediction": "ai",
    "confidence": 0.91,
    "text_content": "This is some text to analyze.",
    "privacy_mode": false,
    "created_at": "...",
    "updated_at": "..."
  }
}

• prediction

  :

  "ai"

  or

  "human"

  — the verdict

• confidence

  : 0.0–1.0, higher = more confident in the prediction

• status

  :

  "processing"

  ,

  "completed"

  , or

  "failed"

Poll for Results

If you did not use

Prefer: wait

status

"completed"

"failed"

GET /text_detect/{uuid}
Authorization: Bearer <API_KEY>

List Text Detections

GET /text_detect
Authorization: Bearer <API_KEY>

Returns paginated text detections for the team.

Callback

If

callback_url

POST

{ "success": true, "item": { ... } }

On failure:

{ "success": false, "item": { ... }, "error": "Error message here" }

Recommended Workflows

Full Media Forensics (Most Thorough)

For a comprehensive analysis, combine all capabilities:

1. Submit detection with all flags enabled:

   {
     "url": "https://example.com/suspect.mp4",
     "visualize": true,
     "intelligence": true,
     "audio_source_tracing": true,
     "use_reverse_search": true
   }
   

2. Poll until

   status: "completed"

3. Read

   metrics

   /

   image_metrics

   /

   video_metrics

   for the verdict
4. Read

   intelligence.description

   for structured media analysis
5. If audio labeled

   "fake"

   , check

   audio_source_tracing.label

   for the source platform
6. Ask follow-up questions via Detect Intelligence if anything needs clarification
7. Check for watermarks via

   POST /watermark/detect

   if provenance is relevant

Quick Authenticity Check (Fastest)

For a fast pass/fail:

1. Submit minimal detection:

   { "url": "..." }

2. Poll until complete
3. Check

   label

   and

   aggregated_score

   (audio) or

   label

   and

   score

   (image/video)
4. Report result with score context

Provenance Pipeline (Content Creators)

For creators who want to prove their content is authentic:

1. Apply watermark to original content:

   POST /watermark/apply

2. Distribute watermarked media
3. Later, verify provenance:

   POST /watermark/detect

   against any copy

Red Flags — Stop and Reassess

• Declaring authenticity without a detection result — Never say media is real or fake based on visual/auditory inspection alone
• Ignoring the score and reporting only the label — A

  "fake"

  label with score 0.51 means something very different from score 0.95
• Submitting local file paths to the API — The API requires publicly accessible HTTPS URLs (does not apply to text detection)
• Sending text longer than 100,000 characters to text detection — Split into chunks or inform the user of the limit
• Polling too aggressively — Start at 2s intervals, back off exponentially; do not loop at <1s
• Asking Detect Intelligence questions before detection completes — Results in 422 error
• Expecting source tracing on "real" audio — Source tracing only runs on audio labeled

  "fake"

• Treating beta features (Identity) as production-ready — Warn users about beta status
• Ignoring

  zero_retention_mode

  for sensitive media — Always suggest this flag when the user indicates the media is sensitive or private
• Making multiple separate API calls when flags can combine — Use

  intelligence: true

  and

  audio_source_tracing: true

  on the detection call instead of separate requests

Response Presentation Guidelines

When presenting results to users:

1. Lead with the verdict — "The detection indicates this audio is likely AI-generated (score: 0.87)"
2. Provide score context — Use the score interpretation table above
3. Mention limitations — Detection is probabilistic, not absolute proof
4. Include actionable next steps — Suggest intelligence queries, source tracing, or watermark checks as appropriate
5. For inconclusive results (0.3–0.5) — Explicitly state the result is inconclusive and recommend additional analysis with different parameters or manual review
6. Never present detection as legal evidence — Detection results are analytical tools, not forensic certifications

Error Handling

url

RESEMBLE_API_KEY

completed

Privacy & Compliance Notes

• Zero retention mode: Set

  zero_retention_mode: true

  to auto-delete media after analysis. The URL is redacted and

  media_deleted

  is set to true post-completion.
• Text privacy mode: Set

  privacy_mode: true

  on text detection to prevent text content from being stored after analysis.
• Data handling: Media URLs and text content are stored by default. For GDPR/compliance-sensitive workflows, enable zero retention (media) or privacy mode (text).
• Callback security: If using

  callback_url

  , ensure the endpoint is HTTPS and authenticated on the receiving end.