Calabi · Labs Try free →

Trend report · gnews_meta_ig · 2026-06-03

TikTok will let you choose how much AI-generated content you want to see - TechCrunch

TikTok will let you choose how much AI-generated content you want to see - TechCrunch

Last month, TikTok quietly rolled out a user-facing control: the ability to limit how much AI-generated content appears in your feed. The setting is buried in Settings > Content Preferences > AI-Generated Content, but its existence is significant. It signals that platforms are no longer treating AI content as a neutral fact—they're treating it as a preference. And that shift changes everything for creators, regulators, and anyone trying to publish synthetic media without it getting shadowbanned, suppressed, or labeled.

What Platforms Actually Scan For in 2026

The detection stack has matured significantly. It's no longer just "does this image look AI?" Modern pipelines are looking for specific, documentable signals—metadata fields, structural artifacts, and provenance gaps that are far harder to fake than a blurry pixel.

C2PA Content Credentials

The Coalition for Content Provenance and Authenticity standard is now embedded in metadata headers for content produced by major AI tools. When you export a video from Sora, Runway, or Adobe Firefly, it writes a c2pa box into the file with fields like actions (what processing occurred), generator (the tool used), and assertions (structured claims about the content's origin). Platforms read this. Meta and TikTok check for c2pa.claim_generator strings matching known AI tool identifiers—com.openai.sora, com.stability.stablediffusion, com.midjourney.journey. If that field is present and claims an AI origin, the content enters a secondary review queue.

The critical field is stitch_assertion, which indicates whether the content was assembled from multiple sources. A stitched assertion flags content that was edited, composited, or otherwise processed after generation—and that's exactly what creators do when they touch up an AI image for commercial use.

AI Metadata Fields That Get Flagged

Beyond C2PA, specific EXIF and XMP tags are checked:

The problem for creators is that this metadata survives most "metadata removed" workflows. Stripping EXIF doesn't remove C2PA boxes in MP4/MOV containers. A clean export from DaVinci Resolve or FFmpeg still carries the container-level provenance data unless you explicitly strip the c2pa atom using tools like ffmpeg -C2PA_remove or a dedicated stripping pipeline.

Encoder Signatures

Each AI video generator has a structural fingerprint. Runway Gen-3 produces specific temporal artifacts in the I-frame sequence. Sora generates characteristic motion-blur patterns in high-motion sequences that differ from physically captured footage. Pika and Kling have their own compression signatures.

Detection models trained on these specific generators look for statistical anomalies in motion vectors, DCT coefficients, and chroma subsampling patterns. The signature isn't visible to the human eye but is readable by classifier models that platforms run on uploaded content before it reaches any human moderator.

Missing GPS and Provenance Gaps

Authentic camera captures have GPS coordinates, device identifiers, and capture timestamps. AI-generated content has none of these unless explicitly injected. A file that has otherwise clean metadata but lacks any GPS data or camera-specific fields like Make, Model, or LensModel triggers a provenance mismatch flag. The platform expects phones to have these fields. A video missing all of them looks like it came from nowhere—and "nowhere" means AI.

This is why stripping alone doesn't work. You can remove every AI metadata field and still fail the provenance check if the file doesn't look like it came from a real device.

What Actually Gets Flagged on Instagram and TikTok

Based on creator reports and platform disclosures:

The Durable Fix: Strip + Inject

Stripping AI metadata is necessary but not sufficient. The durable solution requires two steps in sequence.

Step 1: Strip thoroughly.

Remove all of the following:

Use a pipeline that parses the container format specifically—MP4box, FFmpeg with metadata filters, or specialized tools that handle HEIC and ProRes containers where C2PA is often embedded at the atom level.

Step 2: Inject clean phone identity.

After stripping, the file looks like it came from nowhere. You need to give it a legitimate device identity:

This combination—stripping every AI signal and injecting a coherent, plausible phone identity—survives platform scrutiny because it looks like content that came from a real device, captured at a real place, at a real time.

Why Stripping Alone Fails

Creators who only strip metadata often see their content still flagged. The reason is the provenance gap: the file has no device identity, no GPS data, no capture timestamp—and real photos and videos always have these. A file that's completely clean of AI metadata but has no metadata at all is itself a signal. Platforms treat empty metadata as a red flag because it suggests the content was sanitized.

Injection solves this. A file with a coherent device identity and plausible GPS coordinates passes the provenance check because it looks like what it claims to be: content from a phone.

The platforms are getting smarter, but they're optimizing for human preference—not absolute detection. When you give them a file that behaves like a real camera capture, they treat it like one. The only durable fix is the combination: strip every AI signal, then inject a clean, consistent phone identity.

→ Try Calabi free at calabilabs.com — 10 cleans, no card.

10 free cleans. See the forensic proof before you download.
Try free →

Related reading

Related pages
Blog
All articles
Tool
Open the Sanitizer
Pricing
$12/wk · $34.99/mo