Generated reference: edit sources under docs/helix/, then npm run site:build.

Generated from HELIX docs. Source: docs/helix/00-discover/prior-art-research.md. Edit the source, not this file or site/ HTML.

ddx:
  id: aibadge.research.prior-art
  type: research-note
status: draft
activity: 00-discover
created: 2026-07-22

Prior Art Research: Content Origin Labeling & Attributable Provenance

Research disposition: done (time-boxed survey for genesis / vision) Decisions this research enables: product gap, non-goals, and differentiation before PRD Product name (locked): Innsigle (etymon innsigli; repo slug still aibadge)

Research questions

1. What already exists for labeling AI vs human content (brand and UI)?
2. What already exists for verifiable provenance (crypto, standards, pipelines)?
3. Where does a brand-first product with DKIM-like individual attribution fit without colliding with C2PA or pure badge movements?
4. What market/regulatory pressure makes “why now” concrete?

Findings summary

Gap this project can own: a recognizable brand system for origin states (at least human / AI / mixed) plus optional portable assertions that bind a claim to a person or organization the way DKIM binds a message to a domain—without becoming an AI detector or a full C2PA camera/edit stack.

Competitive detail

Not By AI (notbyai.fyi)

• Offer: Downloadable badges (“Written/Painted/Produced by human, not by AI”); membership project pages; free non-commercial use.
• Trust model: Voluntary self-assertion under a “90% human” rule. Site states it is not an AI detection tool; accountability stays with the badge user.
• Strength: Clear brand, media recognition, low friction for humans who want to signal craft.
• Implication for us: Owning “human-only decorative badge” alone is a crowded, already-branded niche. Differentiation requires multi-state labeling and/or verifiable attribution, not a second human-only sticker.

C2PA and Content Credentials (CAI)

• Offer: Open technical standard (C2PA) + CAI tooling for cryptographically signed manifests (“nutrition labels”) covering creation, edits, generative AI, and publishing. Hardware (e.g. cameras, chips) and major vendors participate.
• Trust model: Hash + sign; tamper-evident; optional identity; inspection via verify UIs. Explicitly does not declare whether content is “real”—it records process claims that tools and creators choose to attach.
• Strength: Correct industrial answer for media supply chains and interoperability.
• Implication for us: Do not compete as “C2PA but smaller.” Align, interoperate later if needed, and own a simpler brand + personal assertion layer for web/text/publishing UX where C2PA is overkill or absent.

SynthID and multi-layer vendor provenance

• OpenAI and Google public materials (2024–2026) push stacked approaches: Content Credentials for detailed context + SynthID when metadata does not survive.
• Implication for us: Model vendors will keep closing “this file came from our model.” They will not own creator-facing brand language or “I personally attest this piece is human / AI-assisted / AI-generated” for arbitrary publishers.

Platform and policy layer

• Platforms apply their own labels; policy and media literacy remain separate from creator-owned marks.
• EU AI Act Article 50: transparency duties for certain AI interactions and synthetic content; provider machine-readable marking; deployer disclosure for deepfakes and some AI text informing the public on matters of public interest (with human-review/editorial-responsibility exceptions). Code of Practice on transparency of AI-generated content supports compliance; obligations are legal even when the code is voluntary.
• Implication for us: “Why now” is regulatory + social demand for clear origin language, not only enterprise media tooling. Opportunity: help publishers and individuals speak origin clearly and stand behind claims, without claiming to be a compliance product of record (assumption to validate in frame).

DKIM analogy (design constraint for later tooling)

Claim vs fact (critical product honesty): Signed badges assert who said what about origin, not what an oracle detected. That keeps us out of the AI-detector arms race and aligned with C2PA’s “nutrition label, not reality oracle” posture.

Positioning options supported by evidence

Research recommendation for vision: lead with A+B (brand first, tooling that makes assertions attributable). Explicitly reject C and D as primary identity for v1.

Assumptions to validate in frame (not treated as facts)

1. Creators will care about signed claims, not only graphics—worth testing with 5–10 target users.
2. “Mixed / AI-assisted” is a useful public state, not just a legal hedge.
3. Individual key discovery can be simple enough (profile URL or DNS) without a heavy identity provider in v0.
4. Interop with C2PA is a later bridge, not a launch dependency.
5. Regulatory disclosure demand increases badge adoption even if we are not a compliance suite.

Open research (parked; not blocking vision draft)

• Trademark / naming collisions for short marks (“Badge,” “Mark,” “Seal,” “Credentials”).
• Whether publishers want AI badges as pride/compliance signal or only human badges as craft signal.
• Threat model: badge stripping, claim replay, key compromise (belongs in threat model + security requirements later).
• Text vs image vs audio first surface (assumption in vision UX: start where brand is visible—sites, posts, bylines).

Sources (survey anchors)

• Not By AI product site and 90% rule framing — https://notbyai.fyi/
• Content Authenticity Initiative / Content Credentials how-it-works — https://contentauthenticity.org/how-it-works
• C2PA coalition context (via CAI materials)
• OpenAI content provenance / SynthID + Content Credentials public posts (2026)
• Google DeepMind SynthID announcements
• EU AI Act Article 50 summaries and Code of Practice on transparency of AI-generated content (obligations from 2 Aug 2026)

Decision log