Why Instagram and TikTok flag Photoshop exports as AI, and how to remove the C2PA metadata

Photographers and video editors keep running into an odd problem: a file that was almost entirely hand-edited shows up on Instagram or TikTok with an "AI info" label. A guide published on dev.to by Sebastian Balog traces the label to metadata that Adobe's own software writes into exports, and walks through a free, scriptable way to catch and remove it before posting.

The trigger, according to the post, is any generative feature: one Generative Fill stroke in Photoshop or a two-second Generative Extend in Premiere Pro. After that, the export carries Content Credentials based on the C2PA standard, which Instagram, Facebook, TikTok, YouTube and LinkedIn can read and translate into a label. Balog cites Adobe's FAQ as saying that when its generative tools create or modify content, this metadata is attached automatically and cannot be switched off for that content. Photoshop's "Export As" metadata option set to none only governs camera and copyright information; it does not touch the C2PA record.

The two markers to watch

The post narrows the labelling down to two data points. The first is the C2PA manifest itself, stored as JUMBF boxes labelled c2pa: inside APP11 segments for JPEGs, and in a top-level uuid box for MP4 and MOV files. The second is an IPTC field, DigitalSourceType, which gets set to values such as trainedAlgorithmicMedia or compositeWithTrainedAlgorithmicMedia.

Both can be checked from the command line. c2patool reports whether a manifest is present, and exiftool can test the IPTC tag with a short conditional expression, which the guide provides as a one-liner for checking a single file.

Automating the check with a folder watcher

Rather than inspecting files by hand, Balog's script watches an export folder on macOS using fswatch and checks each new JPEG, PNG, MP4 or MOV as it lands. Two implementation details matter. The script waits until a file's size stops changing before testing it, because Premiere Pro and Media Encoder write large files over minutes while Photoshop saves a temporary file and renames it — and even then, a render that pauses mid-write can fool the heuristic. It also filters by extension and skips its own cleaned output, so the sidecar and cache files that editing software generates do not set it off. On Windows, he notes, the equivalent would have to be rebuilt around PowerShell's FileSystemWatcher.

Removing the metadata without breaking the file

The cleaning step is where a naive approach causes damage, the post warns. Wiping all metadata with exiftool also deletes the colour profile and orientation data. For photos, Balog recommends removing everything but copying the ICC profile and orientation back from the original. For video, he suggests a stream-copy remux with ffmpeg that drops metadata without re-encoding the footage.

Verification closes the loop: c2patool should report that no claim is found in the cleaned file, and ffmpeg's streamhash comparison between input and output confirms the video streams themselves were left untouched.

A paid alternative, with a disclosure

The same post promotes MetaClean, a Mac and Windows metadata app that packages this workflow as a product. Its watch-folder mode checks finished files in the background, prompts only when Content Credentials or an AI-origin tag are found, saves a verified clean copy beside the original with colour profile, orientation and video streams intact, and offers a batch mode for existing archives. It processes everything locally and costs a one-time $5 after a three-day trial. Balog discloses that he developed the app, so the piece is part tutorial and part promotion; the scripted approach is the free path.

Why it matters

C2PA provenance has moved from a standards document into shipping software, and it behaves differently from ordinary EXIF data. Adobe attaches it automatically and offers no export setting that disables it for generative content, and major platforms are already reading it. In one sense the label is the standard working as intended: surfacing how a file was made. In practice, a single generative edit reclassifies an otherwise manual piece of work as AI-assisted in the eyes of Instagram or TikTok. Creators now face a choice between accepting that label and actively managing metadata they cannot turn off at the source, and this guide lays out the mechanics of that decision either way.