Skip to content
Faisal Mahmud TusherDigital marketing, SEO & analytics
with Faisal Tusher

AI in Marketing › Troubleshooting

AI-Generated Ad Creative Rejected

Why an ad platform rejects or labels generated creative: unchecked rules on synthetic media, a synthetic person, an unsupported claim and an altered product.
Troubleshootingintermediate19 min readUpdated

First, read the rejection as the platform gave it and decide which of five problems it names

Start with the platform's own words. A rejection, a label or a removal arrives with a reason, and that reason is the diagnosis's first evidence, so copy it as given rather than paraphrasing it into what the team expected. Put it beside the asset, the brief, the script if there is one, and the production log entry if there is one. Then decide which of five problems the reason names. A labelling problem: the reason refers to digitally created, synthetic or altered media, and the question is whether the platform's documented rule for this ad category was checked and followed at upload. An endorsement problem: a generated face, voice or presenter is introduced as a customer, an employee or an expert, or says it has used the product. A claim problem: the reason names a superlative, a result, a saving, a comparison or a category claim in the copy or implied by the image. A rights problem: the asset resembles a known person, carries another brand's mark or reads as a named artist's work. A product problem: the item in the ad has a seat, a size, a colour or a capability the real product does not have. Many rejections name two at once, and the illustrative example below is one of them. Where the reason names an ordinary policy rule rather than generated media, the guide to Meta ad review and rejected ads and the guide to Google Ads policies and disapprovals cover the review cycle, resubmission and appeal.

Then fix who answers. A RACI matrix built around the four decisions that recur in generated creative takes minutes and stops the rejection bouncing between people: who is accountable for the labelling decision on each platform, who is accountable for every claim and holds the substantiation file, who answers for rights and likeness and receives the production log, and who signs off brand fit. The AI claims and consumer protection guide already divides the work between the prompter or writer, the editor and the specialist, with the specialist approving any use of a real likeness or any new or stronger claim, and the matrix records that division for the team's own roles. Two rules from the site's creative guides apply before anything is touched. A policy rejection is a signal to revise the brief, not to ask the model for a workaround, because a variant written to slip past a review still says what it says. And a fix made inside the ad platform is invisible to the production log, so the asset goes back to the tool, the designer or the script, and comes forward again through the review gate. Nothing on this page is legal advice; the platforms' help centres and the team's own adviser decide what applies.

Diagnosis table

Diagnosis

SymptomCheckLikely causeFix
The rejection or label refers to digitally created or altered media, and nobody checked the platform's rules before uploadThe platform's help centre page on generated or altered media for this ad category, read today, against what was declared at uploadThe platform's documented rules for digitally created or altered media not checked at upload timeCheck the help centre at upload time for every generated asset, use the platform's labelling tool where it requires one, and record the decision
A generated presenter, face or voice is introduced as a customer, an employee or an expert, or says it has used the productThe script and the brief: is any synthetic person presented as real, or as someone with experience of the productA synthetic person, face or voice presented as a real customer or employeeRemove the endorsement framing, label the synthetic person visibly, and never imply experience nobody has
The rejection names a claim: a superlative, a result, a saving or a comparison the copy or the image makesEach claim against the substantiation file and the approved claims list in the briefA claim the generated copy or image makes that the product cannot substantiateRewrite the claim in approved wording or drop it, and pass every variant through the claims check before launch
The asset resembles a known person, carries another brand's mark or reads as a recognisable artist's styleThe production log: did the prompt or a reference name a person, a brand, a character or an artistA prompt that named a living person, a competitor's mark or a named artist's styleRegenerate from a prompt with no names and references the business owns or has licensed, and escalate anything uncertain to whoever advises on rights
The product in the ad has a feature, a size, a colour or a capability the real product does notThe generated image beside the approved product photograph, edge by edge, especially in extended or edited areasAn extended background or edit that changed what the product is or doesComposite the real product photograph into the scene and keep edits to background and lighting
The asset went from the generator to the platform with no record of who checked accuracy, brand, rights or fitnessThe log for the asset: prompt, references, tool, edits, reviewer and approverA review that skipped the accuracy, brand, rights and fitness gate before uploadPut every generated asset through the four-part review and the approval procedure, with a log entry, before upload

Fix, verify and prevent each cause

Rules on digitally created or altered media not checked at upload

The disclosure and data privacy guide gives the rule for realistic generated or altered imagery on ad platforms in one line: follow the platform's documented rules, and use the platform's labelling tool where its help centre requires it, because the platforms document labelling requirements for realistic synthetic media and unlabelled uploads may be rejected or removed. The same guide adds the part that explains most rejections of this kind: if a platform documents a rule for generated media, it is the platform's rule that applies on that platform, whatever the team's own policy says, and the help centres are the authority and change often, so they are checked at upload time rather than from memory. The AI for creative guide describes the documentation at the level this site can: both major platforms document labelling or disclosure requirements for digitally created or altered media 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. The Meta Business Help Centre and Google Ads Help are where those rules live, and this page does not quote them, because the current page wins over anything written here.

Fix: open the help centre page for the platform and the ad category today, read what it asks for digitally created or altered media, declare or label the asset as it requires, and record the page checked, the date and the decision in the production log before resubmitting.

Verify: the resubmitted asset carries the label or declaration the page asked for, and the log shows who checked which page on which date.

Prevent: write a disclosure rule into the team's AI use policy that names the check at upload as a step rather than a memory, and make the disclosure decision part of the recorded sign-off, as the disclosure guide's checklist asks: platform labelling rules for generated media checked in the help centre at upload time, and the decision recorded with the approval.

A synthetic person presented as a real customer or employee

The AI claims and consumer protection guide draws the line that matters here. A synthetic presenter reading a product description is a production choice; a synthetic customer describing their results is a fabricated endorsement; a synthetic doctor or expert recommending a product implies expertise and experience that no one has. The standard does not change with the tool: as the guide summarises the direction of the FTC's business guidance, claims must be truthful and substantiated whether a person or a model wrote them, endorsements must be genuine, and using AI to deceive is treated as deception rather than as a technology question; in the UK the Advertising Standards Authority applies the advertising codes across media. The site states both only as direction, and none of it is legal advice. A generated presenter who says they have used the product for a month is making an endorsement nobody made.

Fix: strip the endorsement framing from the script and the on-screen text, so the synthetic person presents rather than testifies, and label the person visibly where they could be taken for real; where the message needs a real voice, use a real customer's words with their permission, as the claims guide requires for every testimonial. Never generate the likeness or voice of a real person without their documented agreement.

Verify: read the script for any claim of experience, use or recommendation by the synthetic person, and view the labelled asset in every placement at the size it will be seen, because a label that is cropped out of the vertical version is no label.

Prevent: put the three practices the claims guide names off limits in the AI use policy, polishing real reviews, composite testimonials and synthetic customers, and follow the guide's division of labour: the prompter records that the asset is synthetic, the editor checks the label and that no real person is implied, and the specialist approves any use of a real likeness or voice.

A claim the product cannot substantiate

The AI-assisted ad copy testing guide is blunt about where these claims come from: a model adds them readily, the fastest, guaranteed results, a specific saving, a number of customers, and advertising rules apply to what an ad says however it was drafted. The guide's pre-launch rule is that every factual, comparative or outcome claim in every variant is checked against the approved claims list and the evidence behind it, and any line whose claim cannot be supported is rewritten or dropped. It adds that both major platforms review ads against their published advertising policies, so the policy pages for the category are checked before drafting, the restrictions go into the brief, and a rejection is a signal to revise the brief rather than ask the model for a workaround. An image makes claims too: the producing tutorial treats an image that makes the product look larger, different or capable of something it is not as a misleading claim rather than a styling choice.

Fix: take the claim the rejection names back to the substantiation file; a claim that matches an entry runs in the approved wording only, a claim that is new, stronger or broader goes back for evidence or is softened, and a figure or outcome with no source is cut. Then read the rest of the variant set for the same habit, because a model that added one superlative has usually added others.

Verify: every variant passes the claims and compliance step of the creative approval procedure with its evidence attached, and the test the variants belong to still holds, since variants in an A/B test must differ only in the element being tested and a rewritten claim is a second variable.

Prevent: the creative brief the model drafts from carries the approved facts and claims the copy may use and tells the model what it may not do: add facts, figures, awards, comparisons or promises that are not on the list, change the offer, or invent what the brief does not supply. The claims check sits before launch as a gate, not after a rejection as a repair.

A prompt that named a person, a mark or an artist

The AI for creative guide locates the rights exposure in generated creative mostly on 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 generated person who resembles a real one is a problem whether or not the resemblance was intended. The open questions on copyright and training data set out why this site gives no legal conclusions on ownership or infringement, and list the habits many teams adopt while the law settles: avoid naming living artists, photographers, brands and real people in prompts, upload only material with clear rights to share, check output for resemblance before publication with a reverse image search and a reviewer's eye, and keep a provenance record for every asset.

Fix: read the production log for the rejected asset, or reconstruct it, and find the name, the reference or the style instruction that carried the resemblance in. Regenerate from a prompt written in plain descriptive terms, with references that are the organisation's own photography, previous approved campaign images or licensed stock, and replace any generated person who resembles a real, recognisable individual. Where the question is whether a particular output may be used at all, record what is known and escalate to whoever advises the organisation; the vendor can say what its terms permit, and counsel can say what that means.

Verify: the new asset's provenance record is complete, with the tool and plan, the prompt, each reference and the licence behind it, the human contribution, the resemblance check and who signed off, and the resemblance check found nothing.

Prevent: a rule in the AI use policy that prompts may not name people, brands, characters or artists and that references must be owned or licensed, applied at the brief and checked at the gate, removes most of the risk before an asset exists.

An extended background or edit that changed the product

The AI for creative guide describes adaptation and editing as the most reliably useful application of the tools and names where it goes wrong: extended backgrounds can introduce objects, textures and lighting that were never in the shot, re-crops can push a product or logo out of the safe area, and the review is about what the tool changed to make the fit. The disclosure guide's table treats AI-edited product photography as needing no disclosure in the usual case, provided edits are kept to background and lighting and nothing alters what the product is or does, because altered product features mislead buyers. The producing tutorial therefore composites the real product photograph, packaging and logo into the generated scene rather than relying on generated versions of them, and compares the result with the real thing for shape, colour, label, size relative to the surroundings and anything shown in use.

Fix: rebuild the asset from the approved product photograph, generate or extend only the setting, match light direction, shadows, reflections and scale where the real and generated parts meet, and check the extended areas at full size for anything the product does not have.

Verify: lay the new asset beside the approved product photograph on a contact sheet and compare edge by edge; the product should match in every detail a customer could rely on, and the extended area should hold no second seat, extra control, impossible shadow or repeated object.

Prevent: the image brief states what may be generated and what must be real, usually the product, packaging, logo and any text as real assets with the setting generated, and a monthly review lays out everything generated together against the brand's imagery rules, because sets that pass one asset at a time drift as prompts and references change.

A review that skipped the gate before upload

The producing tutorial ends with a review gate that one reviewer who did not produce the image signs off before it goes to the approval procedure: brief fit, brand, product accuracy, details such as hands, text, logos, edges and reflections, rights, likeness, labelling, placement fit and the log. The creative guide gives the same review in four parts, accuracy, brand, rights and disclosure, and fitness, and names the two practices that make it cheap enough to run every time: a contact sheet of every version beside the source asset, and a log of who prompted, with what references, in which tool, what was changed and who approved. An asset that went from the generator to the platform without that record is the sixth cause, and it is usually the reason one of the other five was not caught.

Fix: run the gate on the rejected asset now, with a reviewer other than the person who produced it, and send each failure back to the step that caused it: the brief for a claim, the prompt or the references for a rights question, the composite for a changed product, the script for an endorsement, the upload for a label. Then submit the corrected version through the creative approval procedure, with the brief and the claims evidence attached, so the brand, claims and compliance, accessibility and sign-off steps happen once and in order.

Verify: the gate record and the sign-off exist for the version that was uploaded, naming the reviewer, the approver and the version label, and the live asset matches the approved file in the asset library.

Prevent: no generated asset reaches a platform without the gate; the human review of generated creative is this gate plus the approval procedure, applied without the assumption that the words or pixels were chosen deliberately.

Prevention: a labelling check at upload, a claims check before it and a log for every generated asset

The six causes sort into three checks, and the routine that prevents the next rejection puts each check at the point in production where it belongs and gives it an owner in the RACI.

The claims check comes first, before anything is generated. The brief carries the approved facts and claims, the policy restrictions for the category from the platform's current pages, and the rule that the model may not add facts, figures, awards, comparisons or promises; the image brief says what must be real and what may be generated, and the prompt rule bars names of people, brands, characters and artists. The substantiation file is the reference for every claim a variant makes, and the person accountable for claims in the matrix holds it. Checking claims after a rejection is a repair; checking them in the brief is prevention.

The review gate comes second, after generation and before upload. One reviewer who did not produce the asset works through brief fit, brand, product accuracy, details, rights, likeness, labelling and placement fit on a contact sheet, at full size and at phone size, and a failure goes back to the step that caused it. The creative approval procedure then does what it does for any creative: brand review, claims and compliance with evidence, accessibility, sign-off with a version label, and release from the asset library rather than from a draft. Generated assets do not need a different procedure; they need the same one applied every time.

The labelling check comes third, at upload, because that is when the platform's rule applies. The help centre page for the platform and the category is read on the day, the asset is declared or labelled as the page requires, and the decision is recorded. The platforms' own enhancements are previewed in the same step and accepted or switched off one by one, since an enhancement that changes a product's colour or crops a required disclaimer is a brand problem however well it performs.

The log ties the three together. Every generated asset carries the tool and plan, the prompt, the references and their licences, the edits and the human contribution, the checks done and who signed off, and the label applied and why, kept with the asset in the library rather than in a chat history. The copyright questions page offers a simple measure for whether the habit has taken hold: the share of published generated assets with a complete record. The log is also what answers a platform's question, a rights query or a disclosure rule months later, which is why nothing is fixed in the ad platform and everything comes back through the gate. The team's AI use policy, in the sense the glossary entry gives it, is where the disclosure rule, the prompt rule, the approved tools and the off-limits practices are written down so the routine survives the next campaign and the next hire.

Rejections also damage the testing a variation set was built for. A variant that would be rejected wastes a test slot and can delay the whole ad set, and the creative testing guide needs variants that differ in one element, run long enough to read, with the decision rule set before the test starts; a claim rewritten after rejection is a second variable. Judge the surviving variants on the metric the campaign is bought for rather than on click-through alone, since the version people click is not always the version they buy from. The applications section places generated creative beside the other uses of AI in marketing and gives a way to judge each by value, review cost and risk. The AI troubleshooting hub lists the neighbouring symptoms: where the copy's problem is an invented fact rather than a claim, the diagnosis for AI content that contains invented facts traces it to the sources and the review, and where a workflow has sent generated messages to the wrong people, the diagnosis for AI automation that sent the wrong messages starts from the send log.

A mobility scooter retailer finds two problems behind one rejection (illustrative)

Illustrative example — not a real result

Underwickle Mobility

Underwickle Mobility, an invented retailer of mobility scooters with 4 showrooms and an online shop, launches an autumn campaign with 1 generated video and 6 generated static images. Within 2 days one platform rejects the video and applies a label to 1 of the statics; the other platform accepts all 7.

The marketing manager copies both reasons as given. The video's reason refers to the presenter: a generated woman in her seventies introduced on screen as "a customer of 9 years" who says the scooter "changed my life". The static's label refers to altered media. The log for the video shows the script was written by the model from a brief that asked for "a long-standing customer's story"; no real customer was involved. The log for the static is empty. Laying the static beside the approved product photograph shows what the extended background did: the generator added a second seat and a rear basket to a scooter that has neither, and nobody compared the two before upload.

The fixes take 3 days. The video is re-scripted so the presenter describes the scooter's range, folding and kerb climbing from the specification sheet and makes no claim of ownership or experience; a visible label marks her as a generated presenter, checked in all 3 placements. The static is rebuilt by compositing the real product photograph into a regenerated setting, with the extended area checked at full size. Before resubmission the manager reads the platform's current page on altered media for the category and records the page, the date and the declaration in the log. Both assets pass the review gate with the merchandising lead, who produced neither, and go through the approval procedure as version 2. The log now holds entries for all 7 assets, and the policy gains 2 lines: no generated person may be scripted as a customer or employee, and the product is never generated. The resubmitted video and static are accepted, and the following campaign's 9 assets pass without a rejection.

Common mistakes

  • Asking the model for a version the platform will accept. A rejection is a reason to revise the brief; a variant written to slip past review still says what it says.
  • Fixing the asset inside the ad platform. The fix is invisible to the log, and the next question about labels, claims or rights has no answer.
  • Checking labelling rules from memory. The help centres change; the page is read at upload time.
  • Letting a generated presenter testify. Presenting is a production choice; claiming to be a customer is a fabricated endorsement.
  • Generating the product. The real product is composited in; a generated one drifts from what customers receive, and extended backgrounds add what the product does not have.
  • Names in the prompt. People, brands, characters and artists in a prompt or a reference are where rights exposure enters.
  • Reviewing the asset like a stock image. Generated creative needs the four-part gate, a contact sheet and a reviewer who did not produce it.

Frequently asked questions

Does a label mean the platform has decided our ad is misleading?

No. A label applied to content a platform detects as generated or altered is a disclosure, and the site's guides record that some platforms apply such labels and that both major platforms document labelling or disclosure requirements for digitally created or altered media in at least some categories. A rejection is a different signal: it names a rule the ad did not meet. Read each as the platform gave it, check the current help centre page for the category rather than relying on what was true last campaign, and treat a label on an ordinary commercial ad as a prompt to confirm the asset was declared correctly, not as a verdict on the claim.

Can we fix the asset inside the ad platform and resubmit?

Resubmitting a corrected asset is normal; making the correction inside the ad platform is the mistake. The site's creative guides send a failed asset back to the tool or the designer, because a fix made in the platform is invisible to the production log, and the log is what answers the next question about labels, claims or rights. Correct the brief, the prompt, the composite or the script, run the review gate again with a reviewer who did not produce the asset, record the new version, and upload from the asset library.

Is a generated presenter allowed at all?

A synthetic presenter reading a product description is a production choice, as the site's claims guide puts it; a synthetic customer describing results, or a synthetic expert recommending the product, is a fabricated endorsement, because it implies experience and expertise nobody has. The safe practice the guides describe is a visible label where the person could be taken for real, no script line that claims to have used the product, and no likeness of a real person without their documented agreement. How the rules apply in each market is a question for a legal adviser; this page is not legal advice.

Terms used on this page

Frameworks

Where this page is used

Sources

  1. Meta Business Help Centre — Meta
  2. Google Ads Help — Google Ads Help
  3. FTC business guidance — Federal Trade Commission
  4. Advertising Standards Authority — ASA

Related