Risk & Compliance
Condensed from Deliverable 4 of the research pass
(ec-workspace/research/world-model-data-flywheel/04-risk-memo.md), where
every holding, EULA clause, and policy is cited to a primary source (fetched
and quoted directly). Anything here that changes product behavior should go
through counsel before shipping. Related shipped docs:
Privacy — Data Collection,
Consent Versioning.
The framing that drives everything
The business contains two distinct copyright acts:
- The buyer's act of training on gameplay — where 2025 US rulings are relatively favorable (Bartz: training on lawfully acquired works = "quintessentially transformative" fair use).
- Our act of reproducing, compiling, and selling copies of copyrighted audiovisual works — where the same rulings are actively unfavorable: Bartz held that assembling and retaining an unauthorized library is the infringing act ($1.5B settlement, ~$3,000/work). A data seller sits on the library side of Bartz and cannot borrow the trainer's fair-use posture.
Contributor consent clears the contributor's rights. It does nothing for the publisher's rights, because under essentially every EULA the player does not own the game's audiovisual output.
Ranked risk register
| # | Risk | Severity | Likelihood | Primary mitigation |
|---|---|---|---|---|
| 1 | Publisher copyright claim against the seller for compiling/selling footage | Critical — statutory damages per work; model invalidation | Medium today, rising as publishers monetize their own data (Origin Lab channel makes unlicensed sellers direct competitors of rightsholders, collapsing fair-use factor four) | Title whitelist now; publisher partnerships as end-state; sell annotations/IDM/service, not footage catalogues (Products & Publishers) |
| 2 | Express EULA AI/harvesting bans — verified verbatim: Frontier cl. 3(g) bans any ML/AI use; Valve video policy: "can't sell or license your videos to others for a payment of any kind"; Nintendo: do-not-touch | Critical for affected titles | High that clauses spread (Frontier-style is the template); medium for enforcement against us specifically | Per-title EULA audit with dated snapshots; hard blacklist; prefer ML-licensed titles (Blizzard SC2 precedent) |
| 3 | Minors in the contributor pool — COPPA (under-13) now; COPPA 2.0 Senate-passed Mar 2026 (under-17), pending House; voidable contracts; child-data-labor narrative | High — FTC action; "paying kids for data" is a kill-shot headline | High absent real age assurance | Hard 18+ with KYC-grade verification — a checkbox is not a mitigation |
| 4 | Voice data — BIPA voiceprint class actions (107+ filed 2025); GDPR biometric-adjacent explicit consent; in-game voice chat captures non-consenting third parties (CIPA all-party consent, $5K/violation) | High — class-action magnet | Medium-high | Mic-only capture; never record incoming voice channels; BIPA-grade written consent; contractual ban on buyer voiceprint derivation |
| 5 | EU AI Act transparency boomerang — GPAI buyers must publish training-content summaries (fines from Aug 2026) → publishers will discover their titles in buyer disclosures | Medium-high | High (mechanical) | Assume zero secrecy; per-title clearance docs designed to survive public disclosure; provenance ledger as product |
| 6 | Anti-cheat flags / publisher kill-switch — kernel anti-cheats treat input hooks as keylogger-shaped; a publisher can blacklist the client per-title at any time | Medium-high (supply destruction; contributor bans) | High for Vanguard-class titles | No hooks in protected processes; per-title technical vetting; pursue whitelisting relationships (Overwolf model — start early, takes years); contributor ban-indemnity policy |
| 7 | Contributor consent-scope claims (Lehrman v. Lovo: paid contributors + vague downstream-AI disclosure → surviving state-law claims) | Medium | Low if done right | Granular, revocable, plain-language consent covering sale/resale/model classes; SAG-AFTRA 2025 IMA as the benchmark; re-consent on new use classes — extends the shipped consent versioning model |
| 8 | Jurisdictional fragility of buyers' fair-use premise (Ross: not fair use; UK: no commercial TDM exception; Japan Art. 30-4 proviso vs sold databases) | Medium (demand-side) | Medium | Contractual risk allocation to buyer; monitor Ross appeal |
| 9 | Competitive fast-follow (Medal input tier; Overwolf quests; NitroGen-lineage reconstruction) | Medium-high | Medium, 12–18 mo | Speed to contracts; longitudinal histories; provenance standard (Competition & Moat) |
| 10 | EU Platform Workers Directive (transposition deadline 2026-12-02) — quest/validation/reward scoring "almost certainly triggers the presumption" of algorithmic management (per the summer GTM plan's own analysis), and directed-collection-as-a-service increases the exposure (spec'd task work under electronic supervision) (added by vault alignment review 2026-07-08) | High (reclassification risk in EU, where the contributor base lives) | High that the presumption applies; medium that enforcement reaches a pilot-scale program | Employment counsel before the founding experiment's paid campaigns; contractor-structure review; the unresolved IP-assignment model (work-for-hire vs tiered licensing, flagged open in the GTM plan) must be settled in the same pass |
| 11 | Activity-not-title breach on public surfaces — EC Data Strategy: "No specific game titles appear in capture-side marketing, investor materials, or bounty solicitation… non-negotiable." The data-engine docs name capture titles throughout — fine internally, a standing-rule breach if the docs site or derived copy is public (added 2026-07-08) | Medium (Grokster-inducement surface) | High if the Cloudflare Pages deployment is public | Confirm the docs deployment is access-gated (now a legal requirement, not just competitive hygiene); title names never move to capture-side public copy |
The two most plausible deal-breakers, in full
1 · Publisher IP, sharpened by publishers entering the data market
The chain: contributor consent never reaches the publisher's copyright → the seller's act (reproduce → compile → sell for profit) is the fact pattern 2025 courts treated worst (Bartz library-side; Thomson Reuters v. Ross on commercial substitution; US Copyright Office guidance against commercial exploitation of competing troves) → Frontier-style "no ML/AI use" clauses and Valve's "never sell your videos" policy add ready-made contract claims → and once publishers sell licensed gameplay data through Origin-Lab-style channels, an unlicensed seller competes with the rightsholder in an established licensing market, which collapses fair-use factor four in the US, defeats the Art. 30-4 proviso in Japan, and is opt-out-blocked in the EU.
Counterweight: the Medal/General Intuition equilibrium — $2.3B valuation, billions of user clips, no public suit — shows enforcement can stay latent for years. But note why their posture is more defensible than footage resale: semantic actions rather than raw footage, first-party lab rather than reseller. Latent ≠ absent, and our product order (Products & Publishers) is designed around exactly this asymmetry.
No court has ruled on third-party sale of recorded gameplay for AI training. Every conclusion above maps from adjacent holdings — which cuts both ways, and is why the whitelist + partnership path is strategy, not paranoia.
Hedge worth funding (2026-07-08, from the EA-internal read in the open-platform vision): the read holds that the reconstruction/spatial products (video → 3D reconstruction → community knowledge) are the defensible transformative use, while action/input capture is "the exposed flank" — the inverse of this memo's footage-first weighting. Both can be true (footage-resale = the contract exposure; input capture = the novel-theory exposure), but until counsel question #6 is answered, the spatial/wiki thread doubles as legal diversification, not just product vision.
2 · Minors plus voice
The "get paid to play games" demographic skews under-18. Paying minors for recorded screen/mic/input data implicates COPPA today, COPPA 2.0 imminently, voidable contracts, and a child-data-labor narrative regulators and journalists will find irresistible. Mic capture separately drags in BIPA voiceprint classes, GDPR explicit consent, and all-party-consent wiretap exposure whenever teammates' voices leak into a recording. Either alone is manageable; together, an unmitigated pipeline (checkbox age gate + full audio capture) is the most likely trigger of an action that arrives before any copyright suit does.
The non-negotiables (inherited by every roadmap phase)
These are Phase 0 gates in the Roadmap; nothing collected before they're in place should be assumed sellable.
- 18+ only, KYC-grade age verification. Not a checkbox.
- Mic-only audio. Never incoming voice channels. Not yet true in the
built pipeline: the loopback game-audio track contains game-mixed voice
chat today (EC-113 open) and can capture Discord/system audio (bug-14) —
Legal/EC Data Strategy.mdalready ordered loopback "disabled entirely pending redesign"; per-process isolation or track-drop is Phase-0 engineering, and the existing corpus's game-audio tracks are quarantined pending counsel (counsel question #5). Voice product leans transcripts-as-intent (the shipped promise uploads transcripts only); double opt-in; no voiceprint derivation, contractually imposed on buyers. - Granular, revocable, plain-language consent covering video / inputs / mic / metadata / sale / resale / model classes — versioned per the shipped consent-versioning architecture; re-consent on new use classes.
- Title whitelist with dated EULA snapshots. Hard blacklist: Nintendo, Frontier-style AI-clause EULAs; no sale-purpose capture of titles whose platform policy bans it (Valve) pending partnership. Documented per-title review.
- Per-clip provenance ledger — consent version, whitelist status at capture time, processing lineage; withdrawal/redaction honored (Ego4D-grade).
- PII pipeline — blur/strip usernames, chat panes, friend lists, on-screen personal data; quests must not elicit on-screen PII (already a constraint in Quests as a Data Product §0).
- Anti-cheat posture — no hooks in protected processes; per-title technical vetting before enabling capture.
Mitigations observed working in the wild
- Publisher-side consent at the source (Microsoft/Ninja Theory: EULA + compliance + IRB) — unavailable to third parties without partnership, which is the point: partner.
- Publisher-licensed marketplace (Origin Lab, Worldmodeldata) — the only structure that fully clears both copyright layers.
- Ego4D-grade human-subjects hygiene — the reference standard for consent, withdrawal, and PII handling in human video collection.
- Buyer-side risk structuring — sell contributor rights + annotations + provenance, with indemnity scoped to contributor claims only; the training-side fair-use bet stays with the buyer, explicitly.