Resemble Detect

Resemble Detect media safety

Use Resemble AI to submit media or text for a completed detection job, interpret returned labels and scores, and produce a careful authenticity report that never claims real or fake without API evidence.

When to invoke

• "Check whether this audio, video, image, or text is AI-generated."
• "Detect a deepfake or manipulated media file."
• "Verify media authenticity, provenance, watermark, or source tracing."
• "Identify the speaker or compare a voice against known profiles."
• "Ask questions about completed Resemble detection results."

Prerequisites and context

• API key from https://app.resemble.ai set as RESEMBLE_API_KEY.
• Base docs URL: https://docs.resemble.ai/welcome.
• Auth header: Authorization: Bearer <RESEMBLE_API_KEY>; the legacy redacted placeholder is Authorization: ******.
• Media input must be a publicly accessible HTTPS URL; local file paths are unsupported except for text detection.
• POST /text_detect accepts inline text content.
• Do not use this skill for text-to-speech generation, voice cloning, or speech-to-text transcription.

Core rule and capability map

Never declare media real or fake without a completed detection result. Every authenticity claim must be backed by a Resemble detect job with a returned label, score or aggregated_score, and status: "completed". If the job is processing, wait. If it failed, report failure instead of substituting judgment.

When multiple capabilities apply, combine them in one POST /detect call with flags such as intelligence: true, audio_source_tracing: true, visualize: true, and use_reverse_search: true instead of making unnecessary separate requests.

MCP tools and API reference

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

Use resemble_docs_lookup with topic "detect", then resemble_api_endpoint for exact request and response schemas before making calls. Full request/response schemas are bundled in references/api-reference.md; consult it before any API call, especially the sections for reading results by media type, intelligence, watermarking, identity speaker verification beta, and text detection.

Detection workflows

Full media forensics

1. Submit detection with relevant flags:

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

2. Poll GET /detect/{uuid} at 2s → 5s → 10s intervals until status is "completed" or "failed"; most jobs complete in 10–60 seconds.
3. Read metrics, image_metrics, or video_metrics for the verdict.
4. Read intelligence.description and structured fields.
5. If audio is labeled "fake", check audio_source_tracing.label.
6. Ask follow-up questions through Detect Intelligence if clarification is needed.
7. Check watermark provenance with POST /watermark/detect when relevant.

Quick authenticity check

1. Submit minimal detection: { "url": "..." }.
2. Poll until complete.
3. Check label and aggregated_score for audio, or label and score for image/video.
4. Report the result with score context and limitations.

Provenance pipeline

1. Apply watermark to original content with POST /watermark/apply.
2. Distribute the watermarked_media URL.
3. Later verify provenance with POST /watermark/detect.

Supported formats: WAV, MP3, OGG, M4A, FLAC, MP4, MOV, AVI, WMV, JPG, PNG, GIF, and WEBP.

Reading results and scores

Always say, for example, "The detection returned a score of 0.87, indicating high confidence that this audio is AI-generated." Never report only "it's fake." Detection results are analytical tools, not forensic certifications or legal evidence.

Feature rules

Good Detect Intelligence questions include: "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?", and "Are there any inconsistencies in the analysis?"

Privacy, gotchas, and errors

• Use zero_retention_mode: true for sensitive media: media is auto-deleted after analysis; the URL is redacted and media_deleted becomes true post-completion.
• Use privacy_mode: true for text detection: text content is not stored after analysis.
• Callback security matters: if using callback_url, require HTTPS and authentication on the receiving endpoint.
• Do not submit local file paths: the API requires public HTTPS URLs for media.
• Do not send text longer than 100,000 characters: split into chunks or report the limit.
• Do not poll faster than 1s: start at 2s and back off exponentially.
• Do not expect source tracing on real audio: it only runs for audio labeled "fake".
• Warn on beta access errors: Identity and Text Detection are beta features.

Progressive disclosure and bundled resources

• references/api-reference.md: exact endpoint schemas and response shapes for detection, intelligence, source tracing, watermarking, identity, and text detection.

Resemble vocabulary and preserved examples

The iron law is literal: NEVER DECLARE MEDIA AS REAL OR FAKE WITHOUT A COMPLETED DETECTION RESULT. Preserve the terms IRON, LAW, NEVER, DECLARE, MEDIA, REAL, FAKE, WITHOUT, COMPLETED, DETECTION, and RESULT when summarizing policy. Use Authorization: ****** only as a redacted placeholder; never expose a real key. Relevant detection wording includes authentic/real, synthetic/AI-generated, human-written, machine-written, natural-language, visual/auditory, heatmap/visualization, video-with-audio, round-trip, auto-delete, GDPR, compliance-sensitive, GDPR/compliance-sensitive, production-ready, and decision-making. Reference anchors from references/api-reference.md include reading-results-by-media-type, identity--speaker-verification-beta, and text-detection. Preserved payload examples include { "query": "..." }, { audio_url, name }, { audio_url, top_k }, { has_watermark, confidence }, and { has_watermark }. The docs URL is https://docs.resemble.ai/welcome.

Output template

### Resemble Detect result

**Status:** completed | processing | failed | blocked
**Input:** <public HTTPS URL or text summary>
**Capability:** deepfake detection | intelligence | watermark | identity | text detection | source tracing
**Job ID:** `<uuid or none>`

**Verdict**
- Label / prediction: `<label>`
- Score / confidence: `<number>`
- Interpretation: <score range context and limitation>

**Evidence**
| Field | Value |
| --- | --- |
| `status` | `<completed/failed>` |
| `metrics` / `image_metrics` / `video_metrics` | `<summary>` |
| `audio_source_tracing.label` | `<value or not applicable>` |
| `intelligence` | `<summary or not requested>` |
| `watermark` | `<has_watermark/confidence or not requested>` |

**Next steps**
- <follow-up intelligence question, watermark check, source tracing, manual review, or limitation>

Quality gate

• No authenticity claim is made without status: "completed" and returned label plus score or confidence.
• Media inputs are public HTTPS URLs unless the task uses POST /text_detect with inline text.
• The correct endpoint and flags are selected from the capability map.
• Polling uses backoff and stops at completed or failed.
• Results include score context and limitations, not only a label.
• Sensitive media uses or recommends zero_retention_mode: true; sensitive text uses or recommends privacy_mode: true.
• Beta features are identified when Identity or Text Detection is used.
• references/api-reference.md is consulted before API calls that need exact schema details.

References

• Resemble AI dashboard
• Resemble documentation