AI Ad Creative Review: A Pre-Launch QA Workflow for Singapore Teams

Why AI ad creative needs its own QA step

Generative tools have changed the economics of ad production. A team that once briefed a designer for three banner variants can now produce thirty headline, image and voiceover combinations in an afternoon. That is the upside. The downside is that every one of those thirty variants is a published advertisement, and every one of them carries the same obligations as an ad that took a week to make: the claims have to be true, the people shown have to have agreed to appear, and the platform it runs on has to accept it.

Most teams already have an approval loop, but it was designed for low volume. A brand manager asked to approve 150 AI-generated variants will skim, and skimming is where mistakes slip through: a product benefit the model invented, a price that is out of date, a face that looks uncomfortably like a real person, or a disclaimer that got cropped out of the 9:16 version.

This guide sets out a practical pre-launch review workflow for AI-assisted ad creative, written for Singapore marketing teams running paid social, search and display. It is not legal advice, and regulated sectors such as finance, health and alcohol have extra rules on top of this.

The rules that apply to every AI-generated ad in Singapore

There is no separate rulebook for AI-made ads in Singapore. The same standards apply whether an image was shot, illustrated or generated. Teams get caught out when they treat generative output as a draft; the rules treat it as a finished advertisement the moment it is published.

The Singapore Code of Advertising Practice

Advertising in Singapore is self-regulated by the Advertising Standards Authority of Singapore (ASAS) through the Singapore Code of Advertising Practice (SCAP). Its preamble sets the baseline: all advertisements "should be legal, decent, honest, and truthful". Several of its guidelines map almost directly onto the risks that generative tools introduce:

• Matters of fact (Guideline 1.1). Claims about objectively ascertainable facts should be capable of substantiation, and advertisers and agencies are expected to hold that substantiation ready to produce to ASAS on request. A language model that writes "fastest delivery in Singapore" has created a factual claim you now need evidence for.
• Testimonials (Guideline 3). Testimonials must be genuine and relate to the personal experience of the person giving them, and Guideline 3.3 warns that testimonials based on fictitious characters must not be framed to give the impression that real people are involved. A generated "customer" photo next to a generated quote is precisely that risk.
• Protection of privacy (Guideline 13). Ads should not portray or refer to any person or their property without prior permission (13.1), and should not manipulate a person or their property, "such as through electronic morphing", to create a misleading or untruthful presentation (13.3). Face-swaps, AI-edited product demonstrations and synthetic spokespeople all sit close to this line.
• Fair competition (Preamble 1.3). Ads should avoid denigration, exploitation of goodwill and imitation. Prompting an image model "in the style of" a competitor's campaign is a fast route to an imitation complaint.

The Code is self-regulatory rather than statute, but ASAS can ask for ads to be amended or withdrawn, and media owners generally will not carry ads that ASAS has ruled against. Read the current text on the ASAS site rather than relying on summaries; the version referenced here is the Code as published on asas.org.sg.

Platform policies on synthetic media

The major ad platforms have added specific rules for AI-generated or digitally altered content, mostly concentrated on political and social-issue advertising:

• Meta. Meta's policy on ads about social issues, elections or politics requires advertisers to disclose when such an ad contains a photorealistic image or video, or realistic-sounding audio, that was digitally created or altered to show a real person saying or doing something they did not, a realistic-looking person or event that does not exist, or altered footage of a real event. Meta states it will reject ads that should have been disclosed and were not, and that repeated failure can lead to penalties.
• Google. Google's political content policy update requires verified election advertisers to prominently disclose when ads contain synthetic content that inauthentically depicts real or realistic-looking people or events, across image, video and audio. Minor edits such as cropping or colour correction are excluded.

Most commercial brands will never run a political or social-issue ad, so these duties may not apply to you. They are still a useful benchmark for where platforms draw the line between routine editing and consequential synthetic content. Check each platform's current policy centre before a campaign, because these pages change.

A five-gate review workflow

The goal is not to have a lawyer look at every variant. It is to put the right check at the right point, so that most problems are caught automatically or by the person who created the asset, and only genuine judgement calls reach a senior reviewer.

Gate 1: Lock the inputs before generating anything

Most AI ad errors start in the brief, not the output. Before anyone opens a generation tool, the campaign should have a single approved source document containing:

• The product facts the ad is allowed to state, each with its substantiation (a test report, a price list, a terms page).
• Claims that are explicitly off-limits, such as "best", "No. 1" or "guaranteed" unless you can prove them.
• Mandatory disclaimers and where they must appear.
• Approved brand assets: logos, product shots, colour values and any licensed talent imagery, with the scope of each licence.

Feed this document to the model as context, and instruct it to use only these facts. It will not make hallucination impossible, but it gives reviewers a fixed reference to check against rather than their memory.

Gate 2: Automated checks on every variant

Some checks are mechanical and should never consume human attention:

• Claim diffing. Compare every factual statement in the copy against the approved facts list. Anything not on the list gets flagged, not blocked, for the human reviewer.
• Banned-term scan. Superlatives, absolute guarantees and regulated terms for your sector.
• Price and offer consistency. Every price, discount and end date in the copy must match the live offer. The SCAP asks that quoted prices include GST where it applies, which is an easy thing for generated copy to drop.
• Format integrity. Confirm the disclaimer, logo and call to action survive every crop and aspect ratio. This is where resized variants most often fail.
• Language coverage. If you run English, Chinese, Malay and Tamil versions, check that each translation carries the same claims and disclaimers. See our guide to multilingual marketing with AI for the wider workflow.

Gate 3: Human review of people, likeness and imagery

This is the gate that should never be automated away. A reviewer looks at every variant that contains a human face, voice or identifiable property and answers three questions:

1. Is this a real person? If yes, do we have written permission covering this use, this channel and this period? SCAP Guideline 13.1 expects prior permission.
2. Could a viewer mistake a synthetic person for a real customer, employee or expert? If the image sits beside a testimonial or an endorsement, the answer is probably yes, and Guideline 3.3 applies. Either use a genuine, consented testimonial or remove the implied endorsement.
3. Does the image show the product doing something it cannot? An AI-enhanced "before and after", a food shot with generated steam and gloss, or a device screen showing features that do not exist can all amount to a misleading presentation.

Log the answer for each asset. A dated record of who checked what is far more useful than a recollection if an ad is ever questioned.

Gate 4: Platform policy pre-check

Before upload, the campaign owner confirms which platform policies apply. For most brands this is a short list: the platform's general advertising standards, any restricted category your product sits in, and, only if the ad touches social or political issues, the synthetic content disclosure rules above. When a disclosure is required, add it at upload rather than hoping the ad is not reviewed. A rejected ad costs a day; an account-level penalty can cost a quarter.

Gate 5: Post-launch monitoring and takedown

Assign an owner to watch comments and platform notifications for the first 48 hours of any campaign built on generated creative. Keep a takedown procedure that can pause every variant of an asset family at once, not one ad at a time. If a claim turns out to be wrong, the fix should propagate back to Gate 1's approved facts document so the model does not regenerate the same mistake next week.

Who owns what

The workflow fails when everyone assumes someone else checked. A lean split for a Singapore marketing team might look like this:

• Campaign owner: maintains the approved facts document, signs off Gate 4, owns takedowns.
• Creator (human or AI operator): runs Gate 2 checks and fixes flagged items before anything goes to review.
• Brand or compliance reviewer: handles Gate 3 and any claims Gate 2 could not resolve.
• Legal or regulatory adviser: consulted only on escalations, new claim types or regulated-sector campaigns.

If personal data is involved, for example ads built from customer lists or personalised with customer attributes, bring your data protection officer into Gate 1. Our PDPA guide for AI marketing covers the consent and purpose questions in more detail.

Keeping review fast as volume grows

A review process only works if people follow it under deadline pressure. A few habits keep it light:

• Review families, not individual ads. Group variants that share the same claims and imagery and review the family once, then spot-check the rest.
• Pre-approve reusable components. An approved headline bank, disclaimer set and product imagery library means most variants are recombinations of already-checked parts.
• Treat new claims as the expensive part. A new headline built from approved facts is low risk. A new fact is high risk. Route them differently.

This is also where an AI marketing platform earns its keep. The same system that generates variants can run the Gate 2 checks, hold the approved facts document, and route only flagged items to a human. The human-oversight principles in IMDA's agentic AI guidance are a useful frame for deciding which steps an automated agent can take on its own and which need a person. For the broader view of what to automate first, see the marketing tasks worth automating.

The closing observation

Generative AI makes the creative part of advertising cheap. It does not make the accountability part cheap, and the teams that do well with it are the ones that redesign review for volume instead of hoping the old sign-off habit will stretch. Lock the facts, automate the mechanical checks, keep human eyes on people and likeness, and write down what you checked. That is enough to run AI-assisted creative at scale without the ad account, or the brand, paying for a shortcut.

Frequently Asked Questions

Do I have to label AI-generated ads in Singapore?

There is no general Singapore rule requiring every AI-generated ad to be labelled. The Singapore Code of Advertising Practice applies to all ads regardless of how they were made, so claims must be substantiated and people must not be manipulated into a misleading presentation. Platforms such as Meta and Google require disclosure of realistic synthetic content in political, election and social-issue ads. Check each platform's current policy for your ad category.

Can I use an AI-generated person in an ad?

Generally yes, as long as the image does not imitate a real individual and is not presented as a real customer, employee or expert. The SCAP's testimonial guidelines say testimonials based on fictitious characters must not be framed to give the impression that real people are involved, so avoid pairing synthetic faces with endorsement-style quotes.

Who is responsible if an AI tool writes a false claim?

The advertiser. Under the SCAP, advertisers and advertising agencies are expected to hold substantiation for factual claims and produce it to ASAS on request. The tool that generated the wording does not change who is accountable for publishing it.

Related Reading

• PDPA Compliance for AI Marketing in Singapore — consent and data use when ads are personalised
• AI Marketing Compliance and IMDA's Agentic AI Framework — where human oversight belongs
• Multilingual Marketing in Singapore — running EN/ZH/MS/TA creative with AI
• Marketing Automation Singapore — how Helixx automates campaign operations