Trend report · gnews_meta_ig · 2026-05-25

Photographers are mad that Instagram is labelling edited photos as 'made with AI' while ignoring AI generated images - Crikey

Here's what nobody in the meta-discussion about Instagram's "Made with AI" label is telling you: the problem isn't that platforms are too sensitive. It's that their detection stack is miscalibrated, opaque, and operating on signals that are trivially easy to spoof. Photographers whose Photoshop composites get labeled AI while Midjourney exports slip through untagged isn't a glitch — it's the predictable output of a system built on metadata archaeology instead of actual signal analysis.

To understand why this keeps happening, you need to understand what platforms are actually checking in 2026. Then the fix becomes obvious.

What Platforms Scan For in 2026

The detection pipeline isn't a single gate — it's a layered stack, and each layer has a specific artifact it's looking for.

C2PA (Content Provenance Authentication) is the most structured layer. Adopted by Adobe, Google, Microsoft, and now baked into iOS and Android camera pipelines, C2PA embeds a cryptographically signed manifest inside the image file using JUMBF (JPEG Universal Metadata Box Format). The manifest carries fields like c2pa.assertions[].label, c2pa.signature, and stds.schema_org.JsonLD. When you upload an image to Instagram or TikTok, the platform reads the C2PA block if present. A generated image from DALL-E, Midjourney, or Stable Diffusion will typically carry a stds.schema_org.JsonLD assertion with "@type": "AIGeneratedMedia" or equivalent. The problem: these manifests are stripped by default by most export pipelines, so actual AI content often arrives clean — and real photos edited in Lightroom or Pixelmator get tagged because the editing software added a manifest noting compositing or enhancement.

AI metadata in EXIF/XMP is the next layer platforms check. Tools like Firefly, Ideogram, and Flux embed specific XMP namespaces: xmpMM:DerivedFrom pointing to an AI generation event, aux:SerialNumber with a vendor identifier, or strings like Generated by Adobe Firefly in the Software field. EXIF fields like Make, Model, and Software are also parsed for patterns associated with known AI pipelines. This is where Lightroom edits get caught — heavy compositing in Lightroom adds an History stack to the XMP metadata that some platform detectors flag as synthetic composition activity.

Encoder signatures are subtler. The DCT (Discrete Cosine Transform) quantization tables and chroma subsampling patterns in a JPEG carry a statistical fingerprint of the encoder that produced them. Platforms have built reference databases of what quantization tables look like when output by specific AI models versus specific phone cameras. A Midjourney export re-encoded through Photoshop will have a different DCT signature than an iPhone shot — but a well-reencoded AI image can approximate a phone's tables, making this signal unreliable on its own.

GPS and temporal metadata gaps are a behavioral signal. If a device with location services enabled produces an image missing GPSLatitude, GPSLongitude, and GPSAltitude fields — but carrying a full DateTimeOriginal — that's anomalous. Platforms weigh this inconsistently, and it explains why some batch-uploaded phone photos get flagged while AI images with injected GPS coordinates do not.

IPTC/XMP consistency scoring is an underreported layer. Platforms score metadata blocks for internal consistency: are the CreateDate, ModifyDate, and MetadataDate fields logically sequenced? Does the software chain in the History stack make sense for a real camera workflow? Inconsistencies flag content for manual review or automatic labeling.

What's Actually Getting Flagged on Instagram vs. TikTok

Instagram's "Made with AI" label uses a composite detector that checks Photoshop's internal compositing flags, luminance gradient anomalies, and XMP history stacks. The signal it's reading is "this image was significantly altered post-capture" — not "an AI model generated this." That's why a photographer's layered composite gets tagged while a Flux export does not: the Flux image has its AI metadata stripped at export, arrives with a plausible phone-equivalent software string, and carries no compositing history. It passes the metadata check invisibly.

TikTok's detection is partially metadata-gated — IPTC:AdjacencyNorm and XMP flags from AI tools trigger the "AI-generated content" label — but TikTok also runs probabilistic image detectors on upload. The gap is timing: detector confidence scores below a threshold don't trigger the label, so high-quality AI images that have been re-encoded (which degrades detector confidence) frequently go untagged. Low-quality AI images or those with intact metadata get caught. This is why the photographer community is furious: their work is metadata-rich and processed, so it trips flags. AI-generated content is often cleaner at the metadata level, so it doesn't.

The Durable Fix: Strip, Then Inject

Every durable counter-detection workflow has two steps, in order. Skipping the first makes the second worthless.

1. Strip all provenance metadata. Remove C2PA manifests, EXIF/XMP blocks, IPTC records, and any History stacks. Tools like exiftool with -all= will strip everything. For C2PA specifically, you need to remove the uuid box and any c2pa markers in the JPEG app segments — a standard EXIF stripper won't touch those. The goal is a clean binary payload: quantization tables and pixel data only, no metadata structure the platform can read.
2. Inject fresh, consistent device identity. Write a new EXIF block that looks like it came from an actual device. This means:
   • Realistic Make and Model: e.g., Apple / iPhone 15 Pro. Don't use a generic string.
   • Consistent Software field matching the device ecosystem: Adobe Lightroom 2024 for a post-processing workflow, or no Software field at all for a straight-from-camera feel.
   • Plausible GPSLatitude, GPSLongitude, GPSAltitude — coordinates that match a real location and are consistent with the DateTimeOriginal timestamp.
   • Realistic DateTimeOriginal, CreateDate, and ModifyDate in the correct temporal sequence.
   • A coherent software chain in XMP:History: e.g., Camera Capture → Apple iOS 17.2 → Adobe Lightroom Classic 13.2. Don't invent software versions that don't exist.
   • Consistent IPTC fields: Location, City, Country if GPS data is present.

   The quantization tables and chroma subsampling should match what that device would produce. An iPhone 15 Pro shoots 4:2:0 chroma subsampling at quality level ~92 — mimic that.

3. Optionally embed C2PA content credentials if the goal is maximum durability. Write a proper c2pa JUMBF block with a stds.schema_org.JsonLD assertion that describes a real camera capture event. This signals to platforms that the content has provenance — and since you control the credential, it says exactly what you want it to say.

The reason stripping alone doesn't work is that platforms flag metadata-absent images as suspicious. A JPEG with no EXIF, no GPS, and no software string from a known device will get a behavioral score penalty. The injection step isn't about lying — it's about giving the image the metadata profile of a legitimate capture so the platform's pipeline reads it as normal content rather than anomalous absence.

Why Manual Editing Gets Flagged When AI Doesn't

The core issue is that Instagram and TikTok are detecting editing process signals — not generation signals. A composite built in Photoshop has deep layer history, multiple DateTime entries, and software flags in the XMP block that read as synthetic composition. A Flux export has no compositing history, and if its metadata is stripped, it reads as a clean image. The platforms are catching the metadata footprint of post-processing software, not the pixel-level fact of AI involvement. Until detection pipelines move to actual pixel analysis at scale — which is computationally expensive and still unreliable on compressed uploads — this asymmetry will persist.

The fix is structural: strip everything, then build a clean device identity that a platform's metadata parser reads as a legitimate, processed-but-not-generated image. That workflow is exactly what Calabi automates — stripping AI provenance signals and encoder fingerprints, then injecting a coherent device profile so content passes detection pipelines cleanly.

→ Try Calabi free at calabilabs.com — 3 cleans, no card.

Related reading

• Remove the Firefly watermark
• What platforms actually scan for when detecting AI content
• Spoof iPhone 14 Pro metadata