Quality-assurance checklist for AI-generated creative

Put a named human at the release boundary

This researched draft provides a sourced checklist and a human approval boundary only. It has no real generated or tested defect record, makes no empirical defect claim, and does not complete the page's required evidence gate. Within that limited scope, the release boundary is explicit: a named human with release authority must inspect the final exported asset and every intended channel preview, review the evidence record, and record APPROVE, REVISE, or REJECT. Any unresolved blocking defect produces a no-go decision; an agent, model, automated score, or approval of an earlier draft cannot release the asset.

This boundary is grounded in NIST AI 600-1 rather than presented as model behavior. NIST suggests establishing minimum performance or assurance criteria within deployment-approval go/no-go processes, including human-oversight roles and responsibilities in the system inventory, and sharing pre-deployment testing with actors who have release-approval authority. For a creative team, the named approver and recorded final-export decision are an editorial implementation of those controls, not a claim that NIST mandates this exact form.

A useful review separates four questions: does the concept answer the brief, are depicted facts and words supportable, is the file technically fit for its placement, and are rights, provenance, and disclosure handled? Record the asset ID, brief version, source files, model or tool, prompt or edit history, destination, reviewer, evidence checked, unresolved defects, corrections, and final decision so the human decision can be audited.

Gate 1

Concept and claimsCheck the brief, audience, depicted facts, offer, and every express or implied claim.

Gate 2

Visual and technical integrityInspect the full file and placement previews for artifacts, drift, text errors, crop, ratio, and contrast.

Gate 3

Rights and provenanceTrace source inputs, permissions, releases, human edits, and retained provenance records.

Gate 4

Disclosure and human releaseApply the rules for the actual channel, then record an explicit approve, revise, or reject decision.

Approve the message before polishing the pixels

Read the creative without the brief beside it and write down what a reasonable viewer could conclude. Check names, prices, dates, quantities, locations, product features, comparisons, quoted people, interface states, and before-and-after implications. NIST defines generative-AI confabulation broadly enough to include confidently false material, output that diverges from the prompt, and contradictions within the same context. For visual QA, that means a plausible label, chart, package, map, receipt, testimonial, or product screen is not evidence that its content is true.

For advertising, the FTC says advertisers and agencies need a reasonable basis for objective express and implied claims before dissemination. The approval boundary is therefore simple: link each objective claim to evidence already held by the team, soften or remove a claim that exceeds that evidence, and reject invented proof. A disclosure cannot rescue an unsupported performance number, a fabricated testimonial, or a depiction that creates a false product capability.

Use a three-state claims pass. Mark copy and depictions PASS only when the exact claim and context match the evidence; REVISE when the idea is supportable but the wording or visual implication overreaches; REJECT when the asset relies on a false fact, nonexistent endorsement, fabricated interface, or unavailable proof. Keep subjective brand judgments in the concept review so they do not blur this evidence decision.

Use documented defect categories without pretending they are project examples

Review the exported file at 100% zoom, then inspect faces, hands, product edges, repeating structures, reflections, transparent objects, labels, logos, jewelry, cables, wheels, shadows, and contact points. Compare each region with the approved source or reference. Describe only what is actually observed in the project asset, with a location, expected state, observed state, and severity; do not pre-fill a defect log with invented anatomy, lettering, or geometry failures.

Adobe Stock's current quality guidance supplies documented inspection categories, not real AI-generated defects for this page. It tells contributors to check for softness and unintended blur at 100%, minimize noise and artifacts, use clean masks without visible selection errors, remove colored halos, keep the main subject sharp, balance lighting and color, and avoid excessive cropping. These categories can structure a future review, but they do not satisfy this page's requirement for captured, task-matched defect examples.

For video, repeat the review through time. Watch once at normal speed for concept and continuity, once muted for visual drift, and once frame by frame around cuts, occlusions, fast motion, and the last frame. Reject unexplained identity swaps, geometry changes, appearing text, temporal flicker, broken contact, or a final frame that cannot support the intended edit. This page does not claim a failure rate for any model.

Proof text, ratio, crop, and smallest placement

Do not trust generated lettering. Transcribe every visible word into the review record and compare it character by character with approved copy, including punctuation, legal qualifiers, currency, units, URLs, and brand capitalization. Check accidental text in backgrounds and product packaging too. If the channel can render live text, prefer that for essential information; if words must remain in pixels, review the final export rather than the editable source.

Build the asset at the placement's required ratio, then preview every delivered crop. Google Ads' current Performance Max guidance is one channel-specific example: it lists horizontal 1.91:1, square 1:1, and vertical 4:5 image assets, requires important content within the center 80% safe area, and notes that its AI-label setting does not itself guarantee regulatory compliance. Those specifications should not be generalized to another publisher. Save a separate checklist row for each actual destination and recheck whenever that destination changes its requirements.

For web creative whose text is intended to be read as text, WCAG 2.2's Level AA contrast criterion is at least 4.5:1 for normal text and 3:1 for large text, with listed exceptions. Test the smallest real placement, not only a large artboard. A headline that technically survives a crop can still fail because it becomes unreadable, competes with the subject, or loses the qualifying words that make the claim accurate.

Check rights, provenance, and disclosure as different questions

Create an input ledger before release: source filename or URL, creator or owner, license or permission record, model or property release when applicable, restrictions, and the human changes made after generation. A tool producing an image does not by itself establish that every depicted logo, person, character, artwork, location, or source input is cleared for the intended use. Escalate ambiguous rights instead of treating a model name, prompt, or “AI-generated” label as permission.

In the United States, the Copyright Office says generative-AI output can receive copyright protection only where a human author determines sufficient expressive elements; human-authored material perceptible in an output and creative human arrangement or modification may qualify, while mere prompting does not. That is a copyrightability statement, not a blanket ownership or non-infringement determination. Keep the human contribution record, but do not turn it into a legal conclusion the source does not make.

Preserve Content Credentials when the workflow supports them and verify the final exported asset rather than assuming metadata survived. The C2PA 2.4 technical specification describes a tamper-evident, cryptographically verifiable provenance architecture and avoids judging whether assertions are good or bad. The official C2PA 2.4 Explainer separately says provenance alone cannot establish that content is true, accurate, or factual. Provenance can strengthen the audit trail; it does not replace claim review, rights review, or human approval.

Finally, determine disclosure from the actual jurisdiction, channel, audience, and use. The FTC's current Advertising FAQs says qualifying information needed to prevent deception should be clear and conspicuous and that fine print cannot contradict another statement or clear up the misimpression it leaves. Put the approved disclosure where viewers encounter the creative, verify that it survives export and crop, and capture the final channel preview.

Consolidate this unsupported standalone page into the workflow hub

The content-map contract requires real defect examples as well as a human approval boundary. This standalone page cannot supply the real defect records: the research pass produced no generated or tested asset and found no relevance-qualified community defect example. The allowed overlap-guard outcome therefore applies. Route the sourced checklist, the human go/no-go boundary, and the explicit unmet-evidence notice to the broader MCP creative workflow hub; do not advance this route as an independent guide.

Consolidation does not convert the missing examples into evidence or mark the gate complete. A later standalone pass would still need captured source assets, unedited outputs, precise defect locations or timestamps, and recorded human decisions. Use the generation-manifest guide to preserve the model, prompt, source input, job ID, timestamps, and selected output needed for that future evidence pass.

Do not publish this route until the missing evidence exists

This researched draft cannot satisfy its canonical evidence requirement because it has no real, task-matched defect record. Under its overlap guard, route only the sourced claim, technical, rights, provenance, disclosure, and human-approval checklist to /guides/mcp-creative-workflows. The parent may carry that limited material and the evidence warning; neither this draft nor consolidation may represent that real defect evidence exists. Keep this route draft, unreviewed, unpublished, and outside the publication manifest.

The 2026-08-09 last30days pass returned 83 items and was degraded: Reddit ended partial after HTTP 429, X was unconfigured, and no evidence cluster cleared the relevance floor for an AI creative output QA checklist with real defects. Zero items were classified relevant; seven were only marginal. No asset was generated or independently tested. Adobe Stock quality guidance was re-fetched successfully and still supplies inspection categories rather than a captured AI-output defect. The required evidence gate remains open. This page therefore does not report practitioner consensus, defect frequency, preferred tools, model rankings, test outcomes, or performance improvements.

The surviving checklist is a release-control framework, not a guarantee of accuracy, accessibility, ownership, regulatory compliance, or platform acceptance. Requirements vary by asset type, jurisdiction, campaign, and destination. Recheck current rules for the actual publication channel and involve the appropriate specialist when a rights, safety, regulated-claim, or disclosure question exceeds the creative reviewer's authority.