Kai ChenIn 2019, Adobe, Twitter, and The New York Times launched the Content Authenticity Initiative to preserve attribution and edit history for digital media—four years before Adobe made Firefly commercially available. By August 2026, Adobe said Claude users could reach more than 50 Creative Cloud tools, allowing one assistant to move work through a sequence the final pixels cannot explain.
Adobe began with a file-level authenticity problem; agentic software has turned it into a workflow problem. AI provenance is shifting from origin markers on outputs to signed records of custody across tools, users, and approvals.
Watermarks inherit the output’s fragility
A watermark asks the output to carry its own history. Its marker must remain detectable while text, images, audio, or video get compressed, cropped, rewritten, translated, remixed, or deliberately manipulated. Creators and downstream tools routinely make those changes, so durability becomes a workflow problem.
Anthropic’s disclosure around Claude makes the limit concrete. Claude’s text watermark appears sparsely in code and factual material, and a full rewrite can erase it. Researchers have also removed watermarks from AI-generated images and inserted them into human-created images. Those attacks break a binary detector in both directions: synthetic work can lose the signal, and human work can acquire it.
Metadata-backed credentials face a transport problem too. OpenAI acknowledged in 2024 that DALL-E 3’s C2PA metadata could be removed accidentally or intentionally. Ordinary file handling and adversarial stripping can produce the same visible result: an artifact without its original context.
On August 17, 2026, Google let Gemini and Flow users remove visible watermarks from generated images, video, and music while retaining SynthID watermarks and C2PA metadata. Google thereby separated a clean export from the signals a verifier can inspect. Provider-specific marks can still help platforms identify and triage model outputs, but they cannot record who approved a later edit, whether a rights holder authorized it, or where the file traveled afterward.
C2PA turns edits into signed history
One year after launching the Content Authenticity Initiative, Adobe placed a content-authenticity tool in Photoshop beta in 2020, moving attribution into the software where edits occurred.
Adobe, Arm, Intel, Microsoft, and Truepic then started C2PA, a cryptographic standard for encoding the provenance of images, video, and audio. A recipient can inspect an attributable record of a file’s source and changes instead of classifying it from visible characteristics.
Adobe later extended Content Credentials through a web app that lets creators apply attribution and declare that they do not want their work used for AI training. A publisher, platform, or model provider can use the same inspectable record when deciding whether to accept, label, or train on the work.
C2PA authenticates only the history it preserves. Cameras still lack sufficient support for its metadata, and editing tools or distribution platforms can strip signed records. An intact credential authenticates a positive history; a missing credential does not prove that a file is false, synthetic, or unauthorized.
A visible break lets a buyer quarantine an asset before relying on its source or rights assertions.
Agents turn authorship into a sequence of actions
Adobe’s Firefly AI Assistant can orchestrate and execute multistep tasks across Creative Cloud applications. Anthropic has connected Claude to Adobe, Blender, Autodesk, Ableton, and other professional tools.
An assistant can invoke one model, call an editing application, apply several transformations, export a file, and hand it back to a person for approval. The final pixels record none of that sequence on their own.
An enterprise governing this workflow needs to know which user initiated the task, which services acted, which source material entered the process, which edits followed, and who authorized the result. The tools must produce that evidence while the work happens.
Adobe controls the creative suite; Anthropic operates across several vendors. Both architectures expose the same weak point: each handoff can drop attribution, authorization, or rights context.
Platforms and law make missing context expensive
OpenAI added SynthID support alongside C2PA Content Credentials and previewed a public verification portal. SynthID estimates model origin, while C2PA carries a history that other tools and platforms can inspect.
Meta said it would label images originating from partners including Google, OpenAI, Microsoft, Adobe, Midjourney, and Shutterstock. Downstream platforms determine whether creation-side credentials affect distribution or remain unused metadata.
Since August 2, 2026, the EU AI Act has required labels for certain AI-generated images, video, audio, and text on matters of public interest when the material is designed to appear authentic. The rule raises the cost of reaching publication without origin context.
A publisher can preserve a credentialed asset, send an incomplete one for review, or reject it before only probabilistic detection remains. That decision gives signed history commercial value.
Adobe can underwrite an intact workflow
Adobe made Firefly commercially available in 2023 and introduced generative credits, tying creation to a metered software business. It also offered businesses full indemnification against copyright-infringement lawsuits over Firefly-generated content. That promise gives a business customer a counterparty for defined copyright risk.
Firefly Foundry extends the model with content grounded in a brand or franchise’s creative universe, bringing permissions, source identity, and accountable rights holders inside the workflow.
When vendors meter AI actions rather than sell only software seats, each billed action gives them a reason to record what happened. Adobe can charge for generation, attach rights assurances to approved workflows, and preserve evidence that the work remained within those boundaries.
A buyer can require a signed record showing that approved source material passed through an indemnified Firefly workflow. A missing link sends the asset back for review instead of into a campaign.
The provenance system Adobe began building in 2019 mattered when it could name a creator and an edit. In 2026, it earns its keep when a buyer can follow every approval through more than 50 callable tools—and refuse the asset where the trail ends.