What AI creative tools do
Creative tools fall into a few groups, and each group has a different place in production and a different review point. Generation produces a new image, video clip, voice track or piece of music from a text prompt and, often, a reference image. Editing changes an existing asset: removing or replacing a background, extending an image beyond its original frame, swapping an object, relighting, retouching, upscaling. Variation takes a concept and produces versions that differ in one or more elements. Adaptation reshapes a finished asset for other sizes, placements and durations. Assembly, which sits mostly inside the ad platforms, combines supplied assets into ads per placement and may add enhancements the advertiser did not make.
| Use | What the tool produces | Where it sits in production | Review point |
|---|---|---|---|
| Concepting | Moodboards, rough visuals, storyboard frames | Before a shoot or a design brief | Direction only; nothing ships |
| Generation | New images, clips, audio from prompts | Supporting and background imagery; placeholders; some hero work | Brand fit, rights, factual accuracy of what is depicted |
| Editing | Backgrounds, extensions, object swaps, retouching | Finishing existing photography and video | Extended areas and edges; anything that changed the product |
| Variation | Versions differing in headline, hook, colour, layout, subject | Building a test set from a proven concept | One variable per test; brand rules held constant |
| Adaptation | Resized, cropped, extended, re-timed versions | Delivering one asset to many placements | Composition, crop safety, legibility, logo placement |
| Assembly | Ads built per placement from supplied assets, with optional enhancements | Inside the ad platforms | What the platform added or changed; whether to turn enhancements off |
The practical centre of gravity is editing, variation and adaptation rather than generation from nothing. Generated hero imagery is possible, but it is where brand risk, rights questions and uncanny detail concentrate, and most teams use generation for supporting material while photography and design still carry the brand.
Variations for testing
The strongest case for AI in creative is the one creative testing has always wanted and rarely afforded: several versions of a proven concept, each differing in a single element, delivered at the same quality. Without tools, producing eight hooks for a video or six subject treatments for a static ad meant eight or six rounds of design time, so tests were small and infrequent. With tools, the production cost of a variation set falls to the cost of reviewing it.
What does not change is the method. A test needs a hypothesis about which element matters, versions that differ in that element and nothing else, enough delivery for each version to be read, and a decision rule set before the test starts. The creative testing guide covers the method on Meta's platform, including how to isolate the creative variable from the delivery system, and it applies to any platform. Generated variations do not shorten the test; they only make the set cheaper to build.
Two habits keep variation sets honest. The first is to vary what the hypothesis names. If the question is whether a benefit-led hook beats a problem-led hook, the variations change the hook and hold the subject, colour, layout and offer constant; a tool asked loosely for "variations" will change everything at once and leave the test unreadable. The second is to read results on the metric the campaign is bought for. A variation that wins on CTRShare of impressions that result in a click. may lose on conversions, and the KPIA metric tied directly to an objective. that decides the winner is the one that connects to the business, not the one that moves first. The discipline of the A/B TestControlled comparison of two variants. is the same whether the variants were drawn or generated. For the words rather than the images, the AI-assisted ad copy testing guide applies the same loop to headlines and body copy across platforms.
Adaptation and resizing
Adaptation is the most reliably useful application and the least glamorous. A campaign needs the same creative in square, vertical and landscape, at several durations, in sizes for display, social, video and email, often in several languages. Tools that extend backgrounds, re-crop around the subject, reflow text and re-time video turn a job that once consumed a designer's week into a review task.
The review is about what the tool changed to make the fit. Extended backgrounds can introduce objects, textures and lighting that were never in the shot. Re-crops can push a logo or a product to the edge of the safe area or out of it. Reflowed text can wrap badly or shrink below legibility on a phone. Re-timed video can cut the moment the product appears. The check is per placement, on a real device, at the size the ad will be seen, and the easiest way to run it is a contact sheet of every adapted version with the original beside it.
Adaptation also meets the platforms' own assembly. Responsive display ads assemble headlines, descriptions, images and logos into layouts per slot, and Google documents that the system may apply enhancements, such as auto-generated video or small rendering changes, under settings the advertiser can switch off. Meta's placement-level customisation lets the advertiser supply a different crop or asset per placement. The question for adaptation is where to do it: supplying placement-ready assets keeps control, while letting the platform adapt saves effort and hands over the crop. Both are fine; the mistake is not knowing which one is in use.
Brand consistency
A generated asset has no memory of the brand. Each image is produced from the prompt and the references supplied with it, and left to itself a tool produces a generic polish that looks like every other generated image. Consistency therefore has to be built in from the brand's side.
The foundation is a written identity: the visual system, imagery principles, colour and type rules and tone that the brand identity and voice guide describes. If that document exists, its imagery section becomes the reference set for every creative prompt: the kinds of subjects and settings the brand uses, the light and colour it favours, what it never shows. Reference images from past campaigns, supplied with the prompt, carry more than adjectives do. Where a tool supports a persistent style or a trained reference, the brand team rather than individual users should own it, and it should be reviewed when the identity changes.
Consistency also has to be checked, not assumed. A generated set that looks coherent in a tool's gallery can drift across weeks of production as prompts are edited and references are swapped. A monthly review of everything generated, laid out together against the identity guidelines, catches drift that nobody notices one asset at a time. The same review is the place to catch the subtler failures: hands and text rendered wrongly, product details that do not match the real product, settings that look like a different country, people who resemble real people. Those are brand failures as much as quality failures, because they are what the brand is seen to have published.
Rights, licensing and disclosure
Three questions sit beneath every generated asset, and none of them is a quality question. They are described here in general terms; they are not legal advice, and a team should take advice on anything with commercial weight.
Can the asset be used commercially? The tool's terms decide what may be done with its output, whether the team's inputs may be used to train the tool, and what happens to uploaded references and brand material. Terms differ between consumer and business tiers of the same product and change over time, so the AI policy should name which tools are approved for commercial creative and someone should re-read the terms periodically. The disclosure and data privacy guide covers vendor terms and the handling of customer data in prompts.
Does the asset infringe anyone's rights? The exposure comes mostly from the input side: prompts that name a living person, a competitor's mark, a protected character or a named artist's style, and reference images the team does not hold rights to. A rule that prompts may not name people, brands or artists, and that references must be the organisation's own assets or licensed stock, removes most of the risk. Likeness deserves particular care: a generated person who resembles a real one, deliberately or by accident, is a problem whether or not the resemblance was intended. The wider questions about ownership of output and training data are unsettled, and the open questions on AI copyright and training data sets them out for counsel and vendors, with the provenance record to keep while they settle.
Does the use have to be disclosed? Platform rules, advertising regulation and consumer protection law are all moving towards transparency for generated or altered media, at different speeds. Both major ad platforms document disclosure or labelling requirements for digitally created or altered content in at least some advertising categories, political and social-issue advertising most clearly, and some platforms apply their own labels to content they detect as generated. For ordinary commercial creative the requirements vary, and the sober course is to decide a disclosure rule, write it into the policy and check the platforms' current pages before each campaign rather than relying on a general answer.
Platform creative tools (documented)
Both major ad platforms now include generative and assembly features in their own interfaces, and the help centres are the record of what each does. This section describes them only at the level those pages do; feature names, availability by campaign type, language and region, and the controls over them change, and the current page wins over anything written here.
Google Ads documents responsive ads that assemble supplied headlines, descriptions, images and logos per placement, and, for Display, enhancements such as auto-generated video that the advertiser can turn off. For Search, it documents an optional campaign-level feature that creates additional headlines and descriptions from the landing page and existing ads, which the advertiser can review, remove and switch off. For Performance Max it documents that the campaign may generate a video from other assets when none is supplied, and it documents generative tools within the campaign builder that produce image and text assets from a prompt and the landing page, with the generated assets shown for the advertiser to review before they are used. The responsive display ads guide covers the assembly mechanics in detail.
Meta documents a family of creative enhancements under the Advantage+ creative name. These include standard enhancements that may adjust brightness and contrast, apply templates, add music or animation, expand an image to fit a placement or improve text, and generative features that can produce background variations, extend an image, and suggest text variations. Meta documents that the advertiser can review which enhancements are applied to an ad and turn them off individually or as a group, and that enhanced versions may be shown to different people. Meta also documents labelling for AI-generated or altered content and a disclosure requirement for digitally created or altered media in ads about social issues, elections or politics.
What neither platform documents is how often an enhanced version is shown, how it is chosen, or what effect it has on results in any particular account. Where that is undocumented this guide does not guess. The practical rule is to look at what the platform has done to the ad in its preview, decide enhancement by enhancement whether it is acceptable for the brand, and turn off the ones that are not. An enhancement that changes a product's colour, crops a required disclaimer or adds music to a silent-by-design ad is a brand problem however well it performs.
A furniture retailer's generated variation test
A furniture retailer has one proven static ad: a sofa in a bright living room, headline "Delivered in 7 days". The creative lead asks a tool for 6 variations that change only the room setting (loft, cottage, flat, studio, conservatory, garden room) and keeps the sofa image, headline, logo and layout fixed. Review takes 40 minutes across the set: 2 variations are rejected because the extended background has added a second sofa leg and a window that casts light from the wrong side. The remaining 4 run against the original in a split test over 14 days with equal budgets. Click-through ranges from 0.9% to 1.4% across versions, but the version with the highest click-through produces the fewest purchases, while the cottage setting matches the original on purchases at 11% lower cost per purchase. The team adopts the cottage version, generates a second set that varies the headline only, and records in its log that background extension needs a close look at edges and shadows every time.
Review before publishing
Every generated or edited asset passes a review before it runs, and the review is the same shape whatever the tool. It has four parts. Accuracy: does the asset show the product, the price, the offer and the setting truthfully, with nothing added that does not exist? Brand: does it match the identity's imagery, colour, type and tone rules, and would the brand team recognise it as theirs? Rights and disclosure: were the prompt and references within policy, do the tool's terms permit the use, and does this placement or category require a label? Fitness: does it survive every placement at real size, with safe areas respected and text legible?
Two practices make the review cheap enough to run every time. The first is a contact sheet: all versions laid out together with the source asset, so edges, shadows, hands, text and product details can be compared at once. The second is a log: who prompted, with what references, which tool, what was changed in review and who approved. The log is what answers a later question about where an asset came from, and it is what a disclosure rule or a rights query will need. Assets that fail review are not fixed in the ad platform; they go back to the tool or the designer, because a fix made in the platform is invisible to the log. When a platform has already rejected or labelled a generated asset, the diagnosis for AI-generated ad creative that was rejected separates a labelling rule, an endorsement, a claim, a rights exposure and an altered product. The AI ad image production tutorial sets out this review as a step-by-step sequence for one generated image, from brief to export.
Common mistakes
- Generating the hero and reviewing it like a stock image. Generated hero assets carry the most brand, rights and detail risk and deserve the most review, not the least.
- Varying everything at once. A variation set that changes hook, subject and layout together cannot be read as a test.
- Judging variations on click-through alone. The version people click is not always the version people buy from.
- Trusting extended backgrounds. Edges, shadows, reflections and extra objects are where edits go wrong.
- Prompting with names. People, brands and artists in a prompt are the main source of rights exposure.
- Not looking at what the platform added. Enhancements applied by default change the ad; see them in the preview and decide.
- No log. Without a record of prompt, references, tool and approval, a rights or disclosure question has no answer.
- Letting consistency drift. Assets that pass one at a time can diverge as a set; review them together.
The applications section places creative beside the other uses of AI in marketing and gives a way to judge each by value, review cost and risk.
