Privacy-Tiered Export

A creative team finishes a project and needs to distribute the assets. Some images go to social media — where generation metadata should be stripped for privacy, but C2PA provenance may need to be preserved for regulatory compliance. Other images go to a client — where workflow details should be removed but color profiles and IPTC attribution must stay intact. A third set goes into the legal archive — where everything is preserved, including the full generation history, as evidence of the creative process.

Each distribution context has different privacy requirements. The same asset, sent to three different destinations, needs three different metadata profiles. This is not a configuration problem — it is an architectural requirement. The system must encode distribution policies as executable rules that can be applied consistently, audited reliably, and extended as regulatory requirements evolve.

The Forces at Work

Several forces shape the need for privacy-tiered export:

• Regulatory diversity: The EU AI Act requires AI-generated content to be marked as such — which means preserving certain provenance metadata even in public distribution. California SB 942 adds disclosure requirements for AI-generated media. Meanwhile, GDPR may require removing personal data embedded in generation metadata. These regulations don't align neatly, creating a matrix of requirements that varies by jurisdiction and context.
• Client confidentiality: Workflow metadata can reveal competitive intelligence — which models a team uses, which LoRAs they've trained, what prompt engineering techniques they've developed. Sending this data to clients or publishing it on social media leaks proprietary process information.
• Provenance value: The same metadata that represents a privacy risk in one context is a provenance asset in another. An archive that preserves the complete generation history enables reproducibility, dispute resolution, and compliance demonstration. Stripping it destroys evidence.
• Manual stripping fails at scale: A team producing hundreds of images per week cannot manually decide which metadata to strip for each distribution context. The metadata persistence problem shows that manual stripping is also unreliable — partial deletion creates false confidence. The only scalable approach is policy-driven automation.

The Problem

Different distribution contexts require different metadata profiles, but most asset management systems treat export as a format conversion — change the file type, adjust the resolution, done. Metadata handling is either all-or-nothing: strip everything (losing provenance) or preserve everything (leaking process details). There is no middle ground that selectively preserves the metadata appropriate for each context.

The problem is compounded by the volume of AI-generated content. When a team produces 50 images per day and distributes to three different channels, that is 150 export decisions per day — each requiring the correct metadata profile. Manual selection is not viable. The metadata policy must be defined once and applied automatically.

The Pattern: Export Policies as Executable Rules

Privacy-tiered export treats metadata handling as a policy engine, not a binary switch. Each distribution context is defined as a named export profile that specifies exactly which metadata categories to preserve, strip, or transform.

The Four Privacy Tiers

*C2PA preservation on social media depends on platform support — some platforms strip all metadata on upload regardless.

1. Share (Maximum Privacy)

For social media and public distribution. Strip all generation-specific metadata: prompts, models, workflows, seeds, GPS coordinates. Preserve technical metadata (color profiles) and add regulatory disclosure fields (IPTC 2025.1 AI disclosure) to satisfy EU AI Act and SB 942 requirements. Optionally preserve C2PA provenance manifests where platform support exists.

2. Portfolio (Professional Display)

For portfolio websites and professional showcases. Strip generation details but preserve attribution metadata — the creator's name, copyright, and contact information should travel with the image. Preserve C2PA manifests to demonstrate provenance. Add AI disclosure fields. This tier balances privacy with professional credibility.

3. Client (Business Delivery)

For deliverables to clients and business partners. Strip workflow details and model information (protecting process IP) but preserve all attribution, color profiles, and provenance metadata. The client receives a professionally complete file with proper attribution and compliance metadata, without visibility into the creative team's proprietary techniques.

4. Archive (Full Preservation)

For legal archives, internal records, and compliance documentation. Preserve everything. Every prompt, every model name, every workflow parameter, every seed value. This is the master record — the evidence of what was created, how, and when. The archive tier exists so that the other tiers can strip metadata confidently, knowing the full record is preserved elsewhere.

Policy-as-Code, Not Manual Checklists

Each tier is defined as a machine-executable policy — a set of rules that can be applied automatically to any asset during export. The agent-first interface design pattern means these policies can be invoked by AI agents as part of automated workflows: “Export this collection with the Client profile” is a tool call, not a manual configuration.

Policies are versioned and auditable. When a regulation changes — when EU AI Act enforcement begins in August 2026, when a new jurisdiction adds disclosure requirements — the policy definition is updated once and applies to all future exports. The audit trail records which policy version was applied to each export, creating a compliance record that survives personnel changes and regulatory updates.

Consequences

Benefits

• Consistent compliance: Every export follows the defined policy for its context. No individual judgment calls, no forgotten metadata fields, no inconsistency between team members. The policy is the single source of truth for what metadata travels with each distribution context.
• Scalable distribution: Whether exporting 1 image or 1,000, the same policy applies. Bulk exports to different channels can run simultaneously with different profiles, without manual intervention.
• Regulatory adaptability: When regulations change, the policy changes in one place. All future exports automatically comply. This is particularly important as EU AI Act and SB 942 enforcement timelines approach — the system adapts to regulatory changes without requiring workflow redesign.
• Audit readiness: The combination of versioned policies and export logs means a team can demonstrate exactly what metadata was included or excluded in any past export, and which policy version governed that decision.

Costs and Limitations

• Upfront policy design: Defining the right tiers for your organization requires understanding your distribution contexts, regulatory obligations, and privacy requirements. This is intellectual work that must happen before the automation delivers value.
• Edge cases: Four tiers cover most distribution scenarios, but specialized contexts (legal discovery, regulatory submissions, insurance claims) may require custom profiles. The framework must be extensible beyond the four standard tiers.
• C2PA re-signing complexity: When metadata is stripped from a C2PA-signed asset, the cryptographic signature becomes invalid. Re-signing with a new manifest that reflects the stripped state requires C2PA infrastructure — an additional architectural component beyond simple metadata removal.

Related Patterns

• Metadata Persistence — the technical challenge that makes automated, policy-driven stripping essential rather than relying on manual metadata removal.
• Agent-First Interface Design — export policies as tool calls that AI agents can invoke as part of automated distribution workflows.
• Collection Semantics — collections as the organizational unit that export policies operate on, with privacy-aware publishing as a collection-level operation.
• Cross-Tool Provenance — the provenance data that privacy-tiered export must selectively preserve or strip depending on distribution context.