Brand safety is not a longer prompt. It is a production system that controls what may enter a model, which claims may leave it, who owns the work and how every published asset can be traced. AI can shorten ideation, adaptation and versioning, but it also makes polished errors cheap. The operating model must therefore optimise approved, effective assets—not the number of generations.
This guide is current to 31 July 2026 and addresses UK creative and marketing operations. Copyright, trade marks, passing off, data protection, advertising rules and sector-specific restrictions may all apply. Campaigns reaching other countries need a separate review. Obtain legal advice for disputed rights, likenesses or high-risk claims.
Define the production boundary
Divide the workflow into activities with different consequences:
| Stage | Appropriate AI role | Required evidence | Human owner |
|---|---|---|---|
| Brief | Structure supplied objectives and constraints | Approved proposition and audience | Brand or product lead |
| Ideation | Generate varied concepts | Rights-safe reference set | Creative lead |
| Production | Draft copy or controlled visual variants | Asset licence and claim sources | Designer or copywriter |
| Adaptation | Resize, translate or localise approved master | Master version and local rules | Channel owner |
| Approval | Highlight differences and missing checks | Review checklist and provenance | Named approver |
| Distribution | Populate approved fields and placements | Final asset hash and media plan | Campaign owner |
Prohibit autonomous publication during the first deployment. Generating a thumbnail is materially different from releasing an ad, changing spend or replying in the brand’s voice.
Define asset classes. A background texture may receive lighter review than a child-facing video, medical claim, financial promotion, political message, celebrity likeness or testimonial. The class—not deadline pressure—sets the approval route.
Build an approved source library
A model should not browse the organisation’s entire drive. Create a versioned library containing current brand rules, product facts, prices, legal lines, substantiated claims, tone examples, logos, fonts, licensed assets and prohibited treatments.
Each source needs:
- owner and approval date;
- territory, media and duration of permitted use;
- licence or employment basis;
- talent, property and music releases;
- claim evidence and expiry;
- accessibility requirements;
- audience or placement restriction;
- model-training permission, if relevant; and
- withdrawal and replacement route.
The IPO’s copyright overview explains that protection arises automatically for original literary, artistic, music, film and other works. Its notice on digital images and photographs warns that most web images are likely protected and generally require permission or an applicable exception. “Found online” is not a licence.
Require contractors and agencies to identify inputs, tools and reuse rights. A commissioning fee does not necessarily transfer copyright; the contract should cover ownership, licence scope, model training, source disclosure, warranties and remediation.
Maintain a rights and likeness ledger
For every asset, record the prompt or instruction, model and version, generated components, reference assets, edits, contributor, licence, approvals and publication destinations. Keep enough intermediate evidence to reconstruct a dispute without retaining sensitive prompts indefinitely.
Create explicit blocks for:
- living-person likeness without documented permission;
- a voice clone without a specific release;
- protected characters, logos or packaging used as a style shortcut;
- “in the style of” a living creator in commercial production;
- stock assets whose licence excludes model training or modification;
- music with unclear composition, recording or performer rights;
- customer content reused beyond the original purpose; and
- confidential launch material sent to a shared training service.
The government’s March 2026 report and impact assessment on copyright and AI reflects an active policy area; it does not replace the current law or settle a particular licence. Teams should record the position at production time and recheck policy before major procurement.
Treat claims as controlled data
Copy generation must retrieve from an approved claim register. Each claim should link to substantiation, product version, qualifying wording, audience, territory and expiry. The system may shorten or combine only within approved rules.
UK marketing and advertising law requires advertising to be legal, decent, truthful, honest and socially responsible, with accurate descriptions. The CAP Code remains media-neutral. ASA’s June 2026 guidance on AI and deepfakes makes clear that automated creation or distribution does not shift responsibility from the advertiser.
Route these claims to specialist review:
- health, beauty, nutrition or therapeutic effects;
- financial, credit or investment outcomes;
- environmental or “sustainable” comparisons;
- superiority, market leadership or quantified savings;
- endorsements and customer results;
- scarcity, deadlines and price reductions;
- regulated products or age-restricted categories; and
- guarantees created from modelled rather than observed data.
For green claims, use the CMA’s Green Claims Code and retain lifecycle scope and comparison evidence. For testimonials, ASA guidance requires evidence that an endorsement is genuine and accurately represented. A synthetic spokesperson must not be presented as a real customer.
Design brand rules that can be tested
Turn a subjective brand deck into checks with examples and tolerances. Define approved logo files and clear space, type hierarchy, colour values, image treatments, reading level, prohibited phrases, inclusive-language principles, caption style and channel-specific dimensions.
Automated checks can catch wrong hex values, missing legal lines, unapproved claims and unsafe crops. They cannot decide whether a concept is culturally sensitive or strategically distinctive. Require a creative review that compares the output with the brief, not only the rules.
Measure sameness. Track recurring composition, stock gestures, plastic texture, excessive glow, meaningless interface graphics, pseudo-scientific motifs and demographic stereotypes. A high prompt-success rate can still produce an archive that looks synthetic and interchangeable.
Use a reference board for each campaign, not one universal “premium AI” aesthetic. Document what should vary—camera position, medium, setting, casting, rhythm—and what must remain recognisable. Reject output because it is generic, even when technically compliant.
Establish provenance without overclaiming it
Record a cryptographic hash of the approved master and each exported rendition. Where tools support it, attach Content Credentials using the current C2PA 2.4 specification. C2PA can carry source and edit assertions; it does not prove that every claim is true or every input was licensed.
Store:
- asset identifier and campaign;
- source ingredients and licence references;
- generation and editing applications;
- material human edits;
- reviewer and approval timestamp;
- disclosure decision;
- exported channel version; and
- revocation or correction status.
Decide when audiences need disclosure based on context, platform requirements and risk. A labelled AI background and an undisclosed synthetic expert have different deception potential. Do not use a provenance badge as a substitute for a clear explanation where a reasonable viewer could otherwise be misled.
Protect personal and confidential material
Prompts can contain customer profiles, unreleased product details, employee names, photographs and campaign strategy. Map what each service stores, uses for training, shares with sub-processors and transfers internationally. Prefer enterprise configurations that disable unrelated reuse and provide deletion, access control and export.
Apply the ICO’s AI and data-protection risk toolkit. Get a lawful basis and additional condition where special-category data is used; do not infer ethnicity, health or sexuality simply to diversify generated casting.
Separate talent-consent evidence from broad creative access. A designer may need an approved headshot, not the performer’s address, contract value or ID. For customer-derived content, honour deletion and objection across the asset library, embeddings and future campaigns.
Secure the creative supply chain
Generation platforms, plugins, asset managers and ad accounts form one attack path. Follow the NCSC’s secure AI development guidance and the government’s AI Cyber Security Code of Practice.
Use single sign-on, multi-factor authentication, role separation and expiring agency access. Restrict plugins, scan uploads and isolate confidential campaigns. No generation tool should receive direct permission to publish or raise media spend during the pilot.
Exercise failure cases:
- a prompt injection hidden in a client document;
- an expired licence still marked approved;
- a model silently changes after sign-off;
- a compromised freelancer account exports source files;
- a generated QR code points to the wrong domain;
- an adaptation drops a mandatory qualifier;
- an ad platform pairs safe copy with an unsafe image; and
- provenance metadata is stripped during export.
Maintain a kill switch by campaign and asset hash. Correct every channel, not only the content-management system.
Measure the whole production outcome
Baseline brief-to-approval time, revision rounds, approval defects, external spend, accessibility failures, claim corrections, rights queries and channel rejections. Count human review and rejected generations; generation time alone is meaningless.
Quality sampling should score:
- brief fidelity;
- factual and claim accuracy;
- brand distinctiveness;
- craft and visual coherence;
- rights completeness;
- inclusion and stereotype risk;
- accessibility;
- channel compliance; and
- post-publication performance.
Compare approved assets from similar briefs, with reviewers blind to production method where practical. Do not reward AI versions merely because there are more of them.
Include downstream correction cost. A cheap asset that requires retailer replacement, retranslation or complaint handling is not efficient production.
Run a 90-day production pilot
| Period | Activity | Gate |
|---|---|---|
| Days 1–15 | Select low-risk asset class, baseline work, classify claims and rights | Brief and approval boundary accepted |
| Days 16–35 | Build source library, rights ledger and brand checks | Every input has owner and permitted use |
| Days 36–55 | Shadow-produce variants against existing campaigns | Quality and review load beat baseline |
| Days 56–75 | Publish a limited, reversible campaign | No unapproved claim, asset or audience |
| Days 76–90 | Compare cost, quality, defects and performance | Scale, narrow, revise or stop |
Pause for an unlicensed or disputed source, unauthorised likeness or voice, fabricated testimonial, material claim without substantiation, disclosure that could mislead, personal-data leak, cross-client retrieval or inability to remove a published asset quickly. Also pause if approval defects rise even while output volume increases.
Scale only when approved-asset cycle time improves, first-pass approval rises, rights records remain complete and campaign outcomes do not deteriorate. Expand one asset class at a time.
Related archive guides cover AI in media production, generative marketing personalisation and social-media AI.
Make provenance part of the craft
A mature creative operation does not hide the machine or fetishise it. It gives people a clear brief, lawful materials, room for judgement and an auditable route from idea to publication. The result should feel intentional because it is intentional.
If an asset cannot be traced to approved facts, rights and a named decision, it is not brand-safe production. It is an attractive unknown with the company’s logo attached.
