The clause that expires your corpus in year ten

The contributor agreement is the document the entire supply side turns on, and the clause that damages it most is not the one that gets written about.

#§40a, not §29

The well-known part first, because it is manageable. §29 UrhG says copyright is not assignable inter vivos — it passes only by testamentary disposition or in the partition of an estate. What a German contributor can grant is a Nutzungsrecht under §31, simple (§31(2)) or exclusive (§31(3)), limited as to place, time and content (Global Law Experts; touring-artists). A US-style "hereby assigns all right, title and interest" clause is construed at best as a broad exclusive licence. Draft the exclusive licence and the problem is solved.

That is a diligence finding at any acquisition, and it is invisible until someone looks for it. The fix is small and it has to go in before the first German contributor signs, not at the point where re-papering means re-contacting several hundred people who have no reason to help you.

The fix is an ongoing royalty or corpus-participation component, however modest, that takes the grant outside the lump-sum trigger. It does three jobs at once:

1. It defeats the §40a ten-year cliff.
2. It is the cheapest available defence to a §32a "bestseller" claim — the unwaivable adjustment where the agreed consideration is strikingly disproportionate to the returns. If the corpus is later licensed to a frontier lab for a very large sum, German contributors have a statutory claim to a share, and a contractual revenue share makes the disproportion argument much harder to run. §32 and §32a cannot be contracted away, so the only available move is to make them hard to invoke.
3. It satisfies the French L131-4 proportional-remuneration default, which permits lump sums only in enumerated cases.

The published precedent for how to size it is the Shutterstock Contributor Fund — a pool from dataset-licensing revenue distributed in proportion to the volume of a contributor's content in purchased datasets, at "a 20% average corporate royalty rate of revenue received by Shutterstock for data licenses," with opt-out since January 2023 (Shutterstock). The lesson from how those funds landed is that the existence of a share buys goodwill only if the number is defensible — Adobe's Firefly bonus produced the figure of roughly $1 for 25,000 assets, and that number did more reputational damage than paying nothing would have (see what creatives actually objected to).

Two more German provisions to build for rather than argue about. §32d lets authors require annual reports on exploitation and proceeds, even under lump-sum contracts (Dorsey) — build per-contributor exploitation reporting into the platform at the start, because retrofitting it onto a corpus already sold to four labs is genuinely hard, and being unable to report is itself a breach. §31a requires writing plus a revocation right and a separate §32c remuneration claim for types of use unknown at contracting; "AI training" was arguably a known use by 2026, so this is receding, but it is live for any archival material.

#The finding that reframes the competitive position

Contra Labs' published terms are the market's creator-friendly frontier: "By participating, you license your work to Contra Labs" — a licence, not an assignment — plus an author byline, no platform fee taken from the expert's rate, an explicit right to decline projects, and payment within 7 business days (contralabs.com/jobs/photographers; Contra Help Center).

Read those terms against European law and something useful falls out.

And the corollary is a live question to raise about the competition. Mercor's CIIA takes an irrevocable assignment and adds a "perpetual, irrevocable, royalty-free worldwide license, with the right to sublicense" over pre-existing Worker Background IP — meaning a designer who brings their own type library, brush set or component kit to a task has arguably licensed it forever. Outlier takes an irrevocable worldwide assignment plus an unconditional, irrevocable moral-rights waiver, with no attribution anywhere in the terms (Mercor legal docs; clause analysis; Outlier Terms of Use).

Neither is enforceable as drafted against a German or French contributor. A US competitor recruiting in Europe on US paper therefore has an unquantified defect in its chain of title over the European portion of its corpus. [WEAK] on severity — the CIIA text sits behind a login and this rests on a secondary analysis. But it is a legitimate diligence question about a competitor and a legitimate reassurance to a buyer about yourself, and it is one of the few places where the generalist platforms are structurally worse off than a specialist.

#The rest of the document, clause by clause

Grant. Exclusive licence, worldwide, for the full term of copyright, sublicensable, with a present assignment as a fallback where permitted and a severability clause so the licence survives where the assignment fails. Enumerate the uses in L131-3 style — reproduction, storage, adaptation, translation, public communication, and expressly "use as training, fine-tuning, evaluation, benchmarking and reward-modelling data for machine learning systems, including by third-party licensees." "All rights, all media, in perpetuity, throughout the universe" is not valid French drafting. Do not rely on general words anywhere.

Byline. Grant it, name the contributor in the dataset card and any publication, and let them opt out at their election. It is near-compulsory in Germany and France, it costs nothing, and credit was among the top objections that turned artists against unpaid lab programmes. Note one genuine tension: France's L113-2 collective-work route — where a legal person takes the initiative, edits and discloses the work under its own name and the contributions merge such that individual rights cannot be separately attributed — makes the company the original owner, which would substantially simplify the French position (Dreyfus). A byline policy cuts directly against the merger requirement. These two goals are in conflict and someone has to choose; the byline is probably worth more than the French simplification, but that is a decision to take deliberately.

Decline rights. Keep them. They are a recruiting asset and they support the independent-contractor characterisation that the worker-classification question turns on.

Exclusivity. Be exclusive on the artefact — this critique, this trajectory, this rating set — and never on the person. Personal exclusivity is a classification risk, a recruiting killer, and under the EU Platform Work Directive close to an admission of employment. It also runs into the December 2026 presumption head-on. And remember §40a caps German artefact exclusivity at ten years for lump-sum grants regardless of drafting.

The employer and client warranty — the most important clause in the document. The contributor warrants that the work is original to them; contains no employer-owned material and breaches no employment, moonlighting or IP-assignment obligation; contains no client material, confidential information or third-party licensed assets (fonts, stock, brushes, plugins, component kits, LUTs); and contains no third-party personal data. Employment IP-assignment clauses are the single largest practical barrier to recruiting employed designers, and this is where that risk lands legally.

A warranty alone is worth little, because an indemnity from an individual freelancer is close to worthless. Pair it with an onboarding attestation naming the contributor's employer and confirming they have checked their own contract, and a positive-consent step at every submission.

The confidential brief, which is the sharpest risk in the product line. A screen recording of real client work captures the brief, unreleased brand assets and possibly customer data. The contributor cannot license what they do not own, and the injured party is the client — who has no contract with you, no notice, and a straightforward claim in confidence and in copyright. The controls, in order of effectiveness:

• Use de-novo briefs for trajectory capture. This is the only structurally safe answer. It costs realism. Pay the cost.
• Where real work is used, require documented written client consent, obtained by the contributor on your form, retained by you, referenced by ID in the record.
• A pre-submission review gate with a documented redaction protocol covering brand marks, client names, file paths, unreleased creative and visible third-party data — Contra's published protocol masks the recorder window, live camera, avatars, names and profile photos (HF dataset card); yours has to go further.
• An automated pre-screen for logos, faces, email addresses and document text before human review.
• A contractual takedown and re-delivery obligation to your buyer, so contamination discovered post-delivery has a defined remedy short of a claim.

A contributor who quotes a confidential brief inside a written critique is the same problem in miniature: the submission gate asks explicitly, the reviewer redacts, the record notes that redaction occurred. Do not rely on contributors to self-police a rule they heard once at onboarding. They are describing work they are proud of, and the client's name will appear.

Data protection and governing law. The contributor is a data subject — identity, pay, per-rater reliability scores — and in trajectory capture potentially handles third-party personal data. Article 13 notice, a stated lawful basis for reliability scoring, a DPIA where the platform allocates work automatically. Choose an EU seat: a US choice of law will not displace mandatory German and French provisions anyway.