Why it matters

A picture's record of where it came from is written into the file, and the file is the first thing the internet changes. This page sets out what people are publicly running into, who it hurts first, the honest case against provenance, and what Moolam does about each part. Every source was read in September 2026 and is linked so you can check it.

The label falls off#

When The Washington Post uploaded an AI video to eight major platforms in October 2025, only YouTube disclosed it, and only in the description. No platform it tested kept the Content Credentials data. In September 2026 The Verge posted Gemini images carrying both C2PA and Google's SynthID watermark to Instagram, and neither was labelled AI after almost two weeks, while real photos touched by a background remover were. OpenAI added a watermark to its images in May 2026 because, as PetaPixel put it, metadata "can be tampered with and/or removed". A screenshot removes it too.

Not every platform strips it. LinkedIn keeps the credential and shows it, according to Lumethic's platform survey. But a picture rarely stays on one platform.

Credit goes with it#

Facebook pages with over a million followers take wildlife photographers' pictures, regenerate them with AI so copyright scanners miss them, and repost them with no credit, PetaPixel reported in August 2025. On Pinterest, artists say years-old hand-drawn work gets tagged "AI modified" and they have to prove otherwise, again and again. In April 2026 Instagram stopped recommending accounts that mostly repost other people's work, which tells you how common it had become.

Nobody can say where a picture came from#

When Columbia's Tow Center asked seven AI chatbots for the place, date and photographer of ten news photos, 14 of 280 answers got all three right. In August 2025 Grok placed an AFP photo from Gaza in Yemen, and a French lawmaker who shared the real photo was accused of spreading disinformation. It cuts both ways: in March 2026, fake forensic analyses were used to call real war photos AI. In a Malwarebytes survey of 1,500 adults, 85 percent said they can no longer tell real from AI, up from 66 percent a year earlier.

Opt-outs do not travel#

A creator's "not for AI training" usually lives in one place: a website file, a platform setting or a metadata field. The picture leaves without it. Cloudflare says Perplexity used undeclared crawlers that ignored robots.txt even on test sites that blocked everything; Perplexity denies it. In December 2025 a Hamburg appeal court held that a photographer's opt-out did not count because it was written in plain words rather than a machine-readable form. In August 2026 Cara, a portfolio site built for artists who refuse AI training, was scraped three times, starting with 12 million images.

Marking is now a legal duty#

Article 50 of the EU AI Act has applied since 2 August 2026. A grace period to 2 December 2026 for systems already on the market depends on the Digital Omnibus being formally adopted. Providers of AI image generators have to mark what they produce in a machine-readable way. The Commission's Code of Practice was drafted with over 187 participants and asks for signed metadata plus an invisible watermark, and for a way others can check content. Rules and standards reads those texts in full.

Who needs this first#

Makers of AI images and the agents that generate them, because the law now asks them to mark output and let others check it. Photographers and artists, because a first-registered record with a date is something a reposter cannot produce. Newsrooms and fact-checkers, because "which record does this copy belong to" is the question they are failing to answer with chatbots.

The case against, answered honestly#

The critics are serious. RAND argues C2PA only works if every tool from camera to post complies, which an open system will never manage. A 2026 security analysis says C2PA misses its own security goals. Researchers have removed invisible watermarks without knowing the scheme, and C2PA signing on Android phones has been forged. Free tools to strip every mark are a click away, and anyone who means to deceive will not register at all.

Moolam agrees with most of that. It does not rely on the mark staying in the file, which answers the stripping. It cannot catch a picture nobody registered, and it proves who registered first, not who made the picture, as the trust model says. Cropping is its real hole: a 10 percent crop is not matched, and Fingerprints prints that number next to the good ones. And by the Code's own words, fingerprinting alone is not enough for Article 50, so Moolam is one layer beside a watermark, never the whole answer.

What Moolam does about each#

How it works follows one passport through every step, and Say how AI may use a picture is the creator's side of the AI-use record.