Authenticating reality: why marking the real matters more than marking the synthetic

For enterprise brands, publishers, and media networks, the primary threat is no longer just identifying what a machine generated. It is proving the integrity of genuine visual assets. Meeting that challenge requires moving past metadata containers and embedding tracking signals directly into the pixel layer itself.

C2PA and JPEG Trust: two standards, one shared fragility

The market often treats C2PA as the exclusive answer to content verification, but it is not the only framework available. The European AI Office's Code of Practice does not mandate C2PA exclusively, recognising that alternative standards exist. A notable example is the JPEG Trust project, which establishes a framework for trust in imaging systems, designed to balance security, privacy, and provenance data in a way that complements rather than duplicates C2PA.

Adoption of C2PA is also accelerating at the hardware level. Major camera manufacturers including Sony, Nikon, Canon, and Leica are integrating C2PA compliance directly into their device firmware, ensuring that an asset's cryptographic manifest is created at the exact moment of capture.

Yet whether utilising C2PA manifests or JPEG Trust architectures, metadata-based solutions share a foundational vulnerability: the provenance information resides outside the visual payload itself. Social media platforms routinely strip metadata to reduce bandwidth, format conversions break cryptographic hashes, and a simple screenshot completely severs the image from its historical record. Once that container is removed, the asset becomes anonymous and unverifiable.

SynthID and forensic watermarking: a necessary distinction

To address the limitations of metadata, the market's attention has gravitated toward Google's SynthID, which has become the most widely discussed watermarking technology in the AI content space. However, its function is frequently mischaracterised. Even if SynthID is technically a pixel-level watermarking tool, it is not a generalist forensic recovery tool. It is a model-dependent marker, purpose-built to signal that an asset was generated by Google's Gemini models. It operates within Google's own infrastructure and does not have the capacity to recover stripped metadata across third-party platforms, nor to track real-world imagery across the open web. It only checks for the presence of its own signature, meaning it can exclusively identify content produced by Google's own models like Gemini, Imagen and their derivatives.

This distinction matters enormously. The real gap in the content integrity market is not the identification of AI-generated images, it is the protection of authentic ones. Real photography, corporate visual assets, and unaltered documentary imagery represent the highest-value targets for manipulation and misappropriation, yet they remain entirely underserved by AI-centric tools. Solutions like Imatag are built precisely for this purpose, specialising in the watermarking of genuine content rather than the labelling of synthetic output.

The mechanics of pixel-level proof

Genuine asset protection requires imperceptible, signal-based digital watermarking. Rather than appending external data packets to a file, this method works by manipulating pixel values directly, using spatial domain embedding or frequency domain transformations such as the Discrete Cosine Transform. The signal is woven into the luminance and chrominance values of the image, making it inseparable from the visual data itself.

The practical consequence is significant: the forensic signal survives the analog hole. Aggressive lossy compression, geometric cropping, format conversion, and even screen capture cannot remove it. The watermark persists regardless of how the asset is distributed or degraded.

This creates a powerful recovery workflow when paired with metadata standards. When a platform strips a C2PA manifest or a JPEG Trust container, the embedded watermark acts as an immutable key. Automated systems can scan the pixel layer, detect the signal, and programmatically re-link the original provenance records to the asset. The chain of trust is restored without any human intervention.

The legislative horizon: a framework still taking shape

The regulatory picture around content authentication is evolving, but significant gaps remain. The EU AI Act establishes machine-readable marking requirements strictly for AI-generated outputs, with no equivalent obligation for authentic, real-world content.

The closest any legislature has come to addressing this gap is in California. Through the California AI Transparency Act (SB 942, signed 2024) and its amendment AB 853 (signed October 2025), California has built the most advanced provenance framework enacted by any jurisdiction to date. AB 853 notably requires camera and recording device manufacturers sold in the state from 2028 to offer users the option to embed provenance signals into captured photos and videos at the moment of creation. It is the first time any legislature has formally touched on the authentication of camera-captured content.

However, even this law frames the question around distinguishing authentic content from AI-generated output, rather than mandating the protection of real assets in their own right. Rather than placing the burden of enforcement on photographers or brands, future legislation should aim to integrate provenance signaling as a default infrastructure requirement. By moving from a model of individual obligation to a framework where authenticity is woven into the standard capture and distribution process, regulators can ensure that the protection of “the real” becomes a structural guarantee rather than an optional strategic choice.

‍