Agencies industrialized AI creative. The paperwork hasn’t caught up.

The race, in one timeline

Lay the announcements end to end and the pattern is unmistakable: each holding company now operates an AI layer that generates, adapts and versions creative at industrial scale — dentsu wired Adobe’s GenStudio into its identity stack in February 2025 to the same end. Volume per human reviewer has exploded by design. That is precisely the environment in which the public AI-creative failures we catalogued happen — a lower defect rate multiplied by a much larger volume still ships defects.

Problem one: the client may not own the work

The US Copyright Office settled the core question in its January 2025 report on AI and copyrightability: purely AI-generated output is not protectable, prompts alone do not make a human the author, and no new legislation is needed — existing law suffices. For an agency, that means a deliverable produced end-to-end by a generation platform may be something the client cannot own or enforce as IP: a competitor could run substantially similar imagery with no copyright recourse.

Human authorship in the workflow — selection, arrangement, substantial creative modification — changes the analysis, which is why how much AI was used, and where, is no longer a production detail. It is a fact that determines what the client is buying. Advertising counsel now recommends handling this contractually: written disclosure of AI use in deliverables, explicit liability allocation, and attention to the generation platforms’ own terms of use.

Problem two: disclosure is becoming a duty, not a courtesy

The industry’s own codes moved first: the ANA’s ethics code calls for clear notice of generative AI use in ads where non-disclosure would materially mislead, and the UK’s IPA and ISBA published joint principles for generative AI in creative advertising. Regulation is close behind: from 2 August 2026 the EU AI Act’s deployer duty makes disclosure of realistic AI imagery a legal obligation for whoever publishes it — and the supply chain has started passing rules downstream, the way Adobe Stock already requires contributors to label generative-AI content and bans real-person references in prompts.

For agencies the awkward part is that the duty usually lands on the client — the deployer — while the knowledge of what was AI-generated lives with the agency. That asymmetry is exactly what contracts exist to fix.

What the missing policy looks like

Three clauses and one habit close most of the gap. Clause one: AI-use disclosure per deliverable — what was generated, with which tool, and what was human-modified, recorded at handoff. Clause two: ownership warranty scoped honestly — which parts of the deliverable carry copyright, which are unprotectable AI output. Clause three: liability allocation for defects, IP overlap and disclosure failures. The habit: a verification gate before anything ships — because none of the paper matters if a six-fingered hand or a garbled headline reaches the placement.

The gate is also what makes the disclosure clause true: you cannot honestly certify what was AI-generated, or that it is defect-free, without inspecting the final asset. That inspection is what Chekr automates — per-creative scans with pinned findings, provenance read from the file, and an exportable record. One free scan shows what the per-deliverable record looks like.