The AI Creative Rights Contradiction Index (2026)

Methodology

This study measures the density of structural rights contradictions across the creative-AI stack: the cells where a creator who complied with one layer (the tool's grant, or the platform's rule) is still exposed by another layer. It is a disagreement index for rights, not a ranking of tools. The substrate is the sourced, dated tools.json and platforms.json data layer (16 tools across image, video, music, and voice; 9 publishing platforms), every field carrying a primary source and a lastVerified date of 2026-06-08.

Four contradiction classes are scored as exact predicates on committed fields (full definitions in the pre-registration):

• Class 1, own-vs-copyright: the tool grants ownership but the output is not copyrightable (you own a thing you cannot stop a competitor copying verbatim).
• Class 2, sell-vs-defend (the indemnity trap): the tool licenses paid commercial use but offers no IP indemnification (you may sell it, but you defend any claim alone).
• Class 3, own-vs-monetize: the tool permits commercial use but the destination platform does not cleanly allow monetizing the content.
• Class 4, free-tier trap: the tool is marketed for commercial use but its free tier specifically cannot be sold.

The scorer is deterministic (pure field comparisons, no judgement in the counts), so re-running it on the deposited data reproduces the index exactly. Any field value that is not a clean yes or no (for example a "conditional" or "limited") is recorded as ambiguous and excluded from that class rather than coerced. Output copyrightability is coded against the US human-authorship rule, reaffirmed when the Supreme Court declined Thaler v. Perlmutter in March 2026.

Two density figures, and a scope note on voice. Classes 1, 2, and 4 are tool-intrinsic, so a tool-level density is reported over all 16 tools (14 of 16 = 88% carry at least one contradiction). Class 3 needs a platform, so the path-level density (the 83% headline) is reported over the 69 modality-matched tool x platform paths. Because the 9 platforms in the layer span image, music, video, and text but include no dedicated voice-distribution platform, the three voice tools (ElevenLabs, Murf, Fish Audio) are scored at the tool level and appear in the tool matrix below, but they contribute zero platform paths, so the 69-path headline covers image, music, and video only. Voice is not omitted; it is tool-level only, and adding a voice platform to the layer is the way to bring it into the path metric.

One transparency note carried from the pre-registration: the original Class 3 predicate named platform values that turned out not to exist in the data (no platform in the set hard-bans monetization). On discovering that at the data-freeze boundary, before computing any counts, the predicate was voided and re-specified on principle, in a dated addendum: a contradiction counts only where a platform genuinely endangers monetization, never where it merely requires disclosure. The strict result (5%) is the headline; a broader "platform friction" reading (91%, counting disclosure and medium strike risk) is reported only as a labelled sensitivity, because counting a disclosure step as a contradiction would inflate the index.

This is a research dataset, not legal advice. Terms change frequently and turn on your specific facts and jurisdiction; verify against each vendor's live terms before any commercial decision.

Finding 1: The contradictions are in the tools, not the platforms

Headline: indemnity gap fires 89% of paths, platform-monetize fires 5%

The loud version of the AI-rights story is about platforms banning or demonetizing AI content. The data says the opposite. Of the 69 modality-matched creator paths, the tool-versus-platform monetization contradiction (Class 3) fires on just 5%, because the nine platforms in the set have converged on permissive rules: all allow AI content, and only one (YouTube Music, coded "conditional") does not cleanly permit monetization. The real exposure is baked into the tools: the indemnity gap (Class 2) fires on 89% of paths and the copyright void (Class 1) on 39%. A creator worried about publishing platforms is looking in the wrong place; the unresolved risk sits in the tool's own terms.

Finding 2: The indemnity gap is near-universal

Headline: 13 of 14 tools license commercial use with no indemnification

The single most prevalent contradiction is the sell-vs-defend gap. Of the 14 tools with an unambiguous answer, 13 license paid commercial use while offering no IP indemnification (two tools, Google Veo and OpenAI's image models, are recorded ambiguous and excluded). If an image, track, or clip you sell is later claimed to infringe, these vendors disclaim any duty to defend or reimburse you. The lone exception is Adobe Firefly, whose commercial guarantee is backed by training only on licensed and public-domain content, and it is exactly this attribute that lets Firefly escape every contradiction class in the matrix below. For brand-critical work, indemnification is the one attribute that most separates the field, and it points to a single vendor.

Finding 3: Six tools sell you "ownership" of an uncopyrightable output

The copyright void is not a vendor choice; it is the human-authorship rule locked in by Thaler. What makes it a contradiction rather than a mere fact is the tools that pair it with an ownership grant. Six of the sixteen tools (Sora 2, Runway, ElevenLabs, Flux, OpenAI's image models, and Stable Audio) tell you that you own your output, while that same output is not copyrightable, so you own an asset you have no legal power to stop a competitor copying verbatim. The other tools grant a license rather than ownership, which is more honest about what you are getting but no more protective. Either way, the enforceable protection comes only from human authorship you add on top, meaningful selection, arrangement, and editing, never from the tool's grant. The full rule is documented in is AI image output copyrightable?.

Finding 4: Density is highest in music, and the class matters more than the composite

Per modality, path-level contradiction density is 100% for music, 80% for image, and 80% for video (voice is tool-level only, see the methodology scope note: the platform layer has no voice-distribution platform, so the voice tools carry contradictions in the matrix but generate no paths). Music tops the list because it is the only modality where the platform class also fires (Suno, Udio, and Stable Audio all meet YouTube Music's conditional monetization). But the honest reading of this whole index is that the composite is high mainly because of one near-universal class, the indemnity gap. The interesting structure is the variance around it: which class fires, which tool escapes, and, over time, whether the density drifts. A single headline percentage over-reads the market; the per-class and per-tool detail is where the decision-useful signal lives, which is why the matrix below reports every cell rather than only the composite.

The tool contradiction matrix

All 16 tools, showing which tool-intrinsic contradiction classes fire. contradiction = the class fires; clear = it does not; ambiguous = the underlying value is not a clean yes or no and is excluded from that class. The full per-path matrix (including the platform class) is in the deposited path-matrix.json.

Adobe Firefly is the only tool clear on every firing class (it indemnifies, and its free-tier value is ambiguous rather than a hard no). Google Veo is excluded on two classes for ambiguous (preview/restricted) values, which is why it also does not fire; it is a data-availability clear, not a genuine one.

Limitations

The sample is the 16 tools and 9 platforms in the maintained data layer, selected by the rule "everything in the layer," not by contradiction count; it is not a census, and expanding the layer and re-running is the correct way to test representativeness. The platform layer has no dedicated voice-distribution platform, so the three voice tools are scored at the tool level only and do not appear in the 69-path headline; the path density therefore describes image, music, and video, and a voice platform should be added to extend it. The predicates are binary over what are sometimes bounded, conditional terms (Firefly's guarantee has a cap and conditions), and ambiguous values are excluded rather than coerced, which is disclosed per class. The copyrightability coding is US-specific. The Class 3 predicate was corrected on principle after the field vocabulary was seen but before counts were computed, which gives it weaker pre-registration strength than the other classes, so it is reported strictly (5%) with the broader friction reading (91%) shown only as a labelled sensitivity. Nothing here is legal advice; verify each vendor's and platform's live terms before any commercial decision.

/research/ai-creative-rights-contradiction-2026/PRE-REGISTRATION.txt   (hypothesis + predicates, pre-committed)
/research/ai-creative-rights-contradiction-2026/ADVERSARIAL-AUDIT.txt  (B1-B6, against the real cells)
/research/ai-creative-rights-contradiction-2026/score.py              (deterministic scorer)
/research/ai-creative-rights-contradiction-2026/RESULT.json           (the computed indices)
/research/ai-creative-rights-contradiction-2026/path-matrix.json      (every tool x platform cell)
/research/ai-creative-rights-contradiction-2026/data-snapshot/        (the frozen, hashed data)