ByteDance's MPA deal buys peace at the output layer and leaves the training bill unpriced

The reason this landed where it did is that one half of the problem has an implementation and the other half has a docket. Face blocking, filters against copyrighted characters, C2PA-style credentials, watermarking and continued monitoring are what public summaries of the memorandum describe ^[5], and each one is a ticket somebody can close: a blocklist, a classifier, a signing step, a review queue. None of them require either party to agree on what a back catalogue is worth as training material ^[4].

The count is the story. Five named output-side controls, zero input-side commitments ^[1]. The guardrails are said to apply across products such as TikTok, CapCut and Dreamina ^[2], so the same five controls imply at least fifteen control-to-surface pairings before anyone argues about a licence ^[2]. That is a moderation programme with a provenance layer bolted on, which is roughly how the dev.to analysis reads it: closer to content moderation than to rebuilding the economics of model training ^[9].

Output-side leverage is legible in a way input-side leverage is not. A model that emits Tom Cruise, Spider-Man or Elsa produces a screenshot that a journalist and a judge can both read without expert testimony ^[12]. A training-data claim requires proving the relevant use, surviving fair use, defining damages, and not accidentally building a licensing structure that hands the largest labs a moat ^[8]. One of those can be resolved by a release; the other cannot be resolved by anyone shipping software.

Worth being clear about the evidence. This is one analysis, resting on a 36Kr report that the agreement is output-side governance rather than a training licence ^[4], on the MPA's statement that ByteDance took feedback after its cease-and-desist letter and implemented safeguards ^[3], and on the association's own responsible-innovation page, which the same account dates to a February 2026 demand to stop infringing outputs and put safeguards in place ^[6]. The memorandum's operative text is not in front of us. Treat the five controls as the published shape of the deal, not as an audited spec.

For anyone building on a video or image model, the sequencing matters more than the parties. The concessions available today are the ones a rightsholder can verify from the outside, which means character and likeness blocking, provenance credentials and distribution throttling arrive as launch gates while the input question stays open. You can budget a filter. You cannot reserve against a price that has not been set, and the dev.to piece is direct that both sides prefer it unset for now: studios do not know whether litigation yields a strong entitlement or a messy middle, and AI firms do not know whether paying early lowers risk or advertises that the asset has a clearing price ^[13].

There is a second-order reason not to treat the filter work as temporary compliance overhead. If training ends up mostly treated as fair use, rightsholders get pushed toward exactly this territory: output control, brand enforcement and downstream revenue shares ^[11]. In that world the blocklist and the credential pipeline you build to satisfy a cease-and-desist become the surface on which the commercial negotiation happens later. The alternative path is uglier arithmetic. Licensing every film, performance, script and visual asset that plausibly appears in training makes transaction costs explode, with portfolio deals favouring large studios and smaller artists still struggling to collect ^[10].