Remove AI Metadata from Video with FFmpeg

Every AI tool leaves a fingerprint in your video

You generate a video with Runway. The output MP4 contains metadata fields that identify it as AI-generated. Knowing how to remove AI metadata from video is what separates content that gets distributed from content that gets flagged or suppressed. TikTok, Instagram, and YouTube can all read these fields — and they do.

This isn't speculation. The C2PA standard (Content Credentials) is specifically designed to tag AI-generated content, and major AI video tools are adopting it. Even without C2PA, tools embed encoder names, creation timestamps, and custom metadata that identify their output.

Stripping this metadata is one FFmpeg command. But knowing what to strip and verifying it's gone requires understanding what's there.

What metadata AI tools embed

Common metadata fields

Every video file has standard metadata containers. These are the fields AI tools most commonly populate:

Any of these that reference an AI tool should be removed before distribution.

Tool-specific metadata

Here's what ffprobe typically reveals for each major AI video tool. Run ffprobe -v quiet -print_format json -show_format video.mp4 on a fresh export before stripping.

Sora is the most aggressive: OpenAI is formally committed to C2PA, so every Sora output includes a cryptographically signed manifest declaring AI origin. Standard -map_metadata -1 won't remove it. Re-encoding does.

C2PA Content Credentials

C2PA is a metadata standard backed by Adobe, Microsoft, Google, and others. It embeds:

• A manifest declaring the content's origin

• Cryptographic signatures proving the declaration

• Action history (what tool generated it, what edits were made)

C2PA data lives in the file's metadata containers (XMP or JUMBF). FFmpeg's -map_metadata -1 removes the standard containers, but C2PA JUMBF data may require additional handling.

Viewing metadata before removal

Use ffprobe to see what's in your file:

ffprobe -v quiet -print_format json -show_format -show_streams video.mp4

Look for these sections in the output:

{
  "format": {
    "tags": {
      "encoder": "Lavf60.3.100",
      "creation_time": "2026-02-01T12:00:00.000000Z",
      "comment": "Generated by AI Tool",
      "title": "AI Generated Content"
    }
  },
  "streams": [
    {
      "tags": {
        "handler_name": "VideoHandler",
        "encoder": "AI Video Encoder v2.1"
      }
    }
  ]
}

Any field that references an AI tool needs to go.

Stripping metadata with FFmpeg

Basic metadata removal

ffmpeg -i input.mp4 -map_metadata -1 -c:v copy -c:a copy output.mp4

-map_metadata -1 removes all metadata from the output. -c:v copy -c:a copy copies streams without re-encoding. Fast, but doesn't change the video content itself.

Deep metadata removal

Some metadata survives -map_metadata -1 because it lives in stream-level containers. For complete removal:

ffmpeg -i input.mp4 \
  -map_metadata -1 \
  -map_metadata:s:v -1 \
  -map_metadata:s:a -1 \
  -fflags +bitexact \
  -flags:v +bitexact \
  -flags:a +bitexact \
  -c:v copy -c:a copy \
  output.mp4

• -map_metadata:s:v -1: Removes video stream metadata

• -map_metadata:s:a -1: Removes audio stream metadata

• -fflags +bitexact: Prevents FFmpeg from adding its own metadata

• -flags:v +bitexact and -flags:a +bitexact: Prevents codec-level metadata

Nuclear option: re-encode everything

For maximum metadata removal, re-encode the entire file. This destroys any metadata embedded in the encoding structure:

ffmpeg -i input.mp4 \
  -map_metadata -1 \
  -fflags +bitexact \
  -c:v libx264 -crf 22 -preset fast \
  -c:a aac -b:a 128k \
  -movflags +faststart \
  output.mp4

Re-encoding creates a completely new file. No metadata from the original survives. The only downside: slight quality loss and longer processing time.

API call for metadata stripping

curl -X POST https://renderio.dev/api/v1/run-ffmpeg-command \
  -H "Content-Type: application/json" \
  -H "X-API-KEY: your_api_key" \
  -d '{
    "ffmpeg_command": "-i {{in_video}} -map_metadata -1 -map_metadata:s:v -1 -map_metadata:s:a -1 -fflags +bitexact -c:v libx264 -crf 22 -c:a aac -b:a 128k {{out_video}}",
    "input_files": { "in_video": "https://example.com/ai-video.mp4" },
    "output_files": { "out_video": "clean.mp4" }
  }'

Handling C2PA specifically

C2PA content credentials use JUMBF (JPEG Universal Metadata Box Format) or XMP. FFmpeg's metadata stripping covers XMP in most cases. For JUMBF:

Re-encoding the file (-c:v libx264 instead of -c:v copy) is the most reliable way to remove JUMBF data, because it creates a completely new file structure.

If you want to verify C2PA removal, use the c2patool:

# Install c2patool (Rust-based CLI)
cargo install c2patool

# Check for C2PA data
c2patool input.mp4

# Should return "No claim found" after stripping
c2patool output.mp4

Verification: confirm metadata is gone

After stripping, verify nothing remains:

# Check format-level metadata
ffprobe -v quiet -print_format json -show_format output.mp4 | grep -i "encoder\|comment\|title\|creation\|tool"

# Check stream-level metadata
ffprobe -v quiet -print_format json -show_streams output.mp4 | grep -i "encoder\|handler\|vendor\|comment"

# Check for any remaining tags
ffprobe -v quiet -show_entries format_tags -show_entries stream_tags output.mp4

The output should be empty or contain only generic FFmpeg encoder strings.

Batch metadata stripping

Strip metadata from multiple AI videos:

import requests

API_KEY = "ffsk_your_key"
HEADERS = {"Content-Type": "application/json", "X-API-KEY": API_KEY}

videos = [
    "https://storage.example.com/runway-output-1.mp4",
    "https://storage.example.com/runway-output-2.mp4",
    "https://storage.example.com/kling-output-1.mp4",
    "https://storage.example.com/sora-output-1.mp4",
]

for i, url in enumerate(videos):
    res = requests.post(
        "https://renderio.dev/api/v1/run-ffmpeg-command",
        headers=HEADERS,
        json={
            "ffmpeg_command": "-i {{in_video}} -map_metadata -1 -map_metadata:s:v -1 -map_metadata:s:a -1 -fflags +bitexact -c:v libx264 -crf 22 -c:a aac -b:a 128k {{out_video}}",
            "input_files": {"in_video": url},
            "output_files": {"out_video": f"clean_{i+1}.mp4"}
        }
    )
    print(f"Video {i+1}: {res.json()['command_id']}")

How different platforms check for AI metadata

Not every platform handles this the same way. Knowing which one you're posting to matters.

TikTok checks metadata primarily through its content processing pipeline on upload. It reads encoder strings and, increasingly, C2PA credentials. TikTok doesn't publicly disclose exactly what triggers suppression, but the consistent report from creators who've tested this: videos from known AI tools get lower initial distribution than camera-recorded content. The TikTok video fingerprinting guide covers what their system actually checks.

YouTube reads C2PA credentials and uses them to add an "AI-generated" label on videos, not to suppress them. As of 2026, YouTube requires creators to disclose AI-generated content in sensitive categories (news, elections, health). If you're posting in those categories and the metadata confirms AI origin, expect a disclosure label to appear automatically.

Instagram follows Meta's emerging AI labeling policies. Meta is a C2PA coalition member, so Reels and Feed posts from tools that embed C2PA credentials (Sora, some Runway outputs) get an "AI info" label. Less aggressive than YouTube's approach, but increasingly automated.

LinkedIn has lighter metadata enforcement at the moment. Standard metadata stripping is enough for most professional use cases.

The practical takeaway: for TikTok, strip everything and re-encode. For YouTube and Instagram, your main concern is C2PA credentials specifically.

Metadata stripping vs. re-encoding: which approach fits

If your video comes from Runway or Kling and doesn't use C2PA, fast stream copy is fine. If it comes from Sora or any tool explicitly committed to C2PA, skip straight to re-encoding. The quality difference at CRF 20–22 is not visible to human eyes; the file size might actually shrink if the original was encoded inefficiently.

For context on what full metadata stripping looks like across different file types and containers, the comprehensive metadata stripping guide covers MKV, WebM, and other containers beyond MP4.

Beyond metadata: additional steps

Stripping metadata is necessary but not sufficient for avoiding AI detection. Platforms also use:

• Visual pattern analysis: AI video has characteristic smoothness and consistency

• Audio pattern analysis: AI-generated speech or music has specific artifacts

For comprehensive AI detection avoidance, combine metadata stripping with:

1. Add sensor noise: -vf "noise=alls=6:allf=t" (mimics real camera noise)

2. Crop edges: -vf "crop=iw-6:ih-6:3:3" (removes potential edge artifacts)

3. Adjust colors: -vf "eq=brightness=0.01:saturation=1.02" (shifts away from AI defaults)

4. Modify audio: -af "asetrate=44100*1.005,aresample=44100" (shifts pitch slightly)

Combined command:

ffmpeg -i ai-video.mp4 \
  -vf "crop=iw-6:ih-6:3:3,noise=alls=6:allf=t,eq=brightness=0.01:saturation=1.02" \
  -af "asetrate=44100*1.005,aresample=44100" \
  -map_metadata -1 -fflags +bitexact \
  -c:v libx264 -crf 22 \
  -c:a aac -b:a 128k \
  output.mp4

This strips metadata, adds natural-looking noise, adjusts colors, and modifies audio. The result passes most automated detection checks that rely on metadata and simple visual patterns.

FAQ

Does stripping metadata actually prevent AI detection on TikTok?

Metadata removal reduces one detection signal, but it's not a complete solution on its own. TikTok uses visual analysis in addition to metadata. Stripping metadata is step one; for TikTok specifically, see making AI video undetectable on TikTok for the full picture.

Will re-encoding noticeably degrade video quality?

At CRF 20–22 with libx264 -preset fast, the quality loss is invisible at normal viewing sizes. You're looking at 99%+ PSNR versus the original. If the source was already a lossy encode (all Runway/Kling/Sora outputs are), you're doing generational compression, but the difference is imperceptible unless you're doing pixel-level comparison.

Can I strip C2PA credentials without re-encoding?

Sometimes. C2PA data stored in XMP containers gets removed by -map_metadata -1. C2PA data stored in JUMBF (which Sora uses) typically survives stream copy. The only reliable way to remove JUMBF is to re-encode, because it creates a completely new file structure that doesn't carry the original containers.

How do I verify all metadata is gone after stripping?

Run ffprobe -v quiet -print_format json -show_format output.mp4 | grep -i "encoder\|comment\|title\|creation\|tool" on the output. The result should be empty or contain only generic FFmpeg strings. For C2PA specifically, use c2patool output.mp4 — it should return "No claim found."

Does the -map_metadata -1 flag affect audio quality?

No. It only affects metadata containers, not the audio or video streams. Combined with -c:v copy -c:a copy, the streams pass through untouched.

If you're also trying to avoid detection patterns beyond metadata, the making AI video undetectable on TikTok guide covers the complementary visual and audio steps.