AI for Ad Creatives: What Works Today and What Doesn’t

AI tools can now write ad copy, generate images, clone voices, make talking avatars and cut video. Some of that is useful every day. Some of it produces ads that look fine in a demo and fall flat in a feed. And some of it can get you in trouble.

We use AI in our own creative work at adthief.ai, so this isn’t a post against it. It’s an honest list of where it earns its place, where it doesn’t yet, and the rules we follow.

Where AI works well today

Research and summarising

This is the most underrated use. AI is good at reading a lot of text and pulling out patterns:

• Summarising hundreds of your reviews (and competitors’ public reviews) into the problems, outcomes and doubts customers mention most.
• Grouping a swipe file of competitor ads by angle, hook type and offer.
• Turning a long customer interview transcript into a list of phrases worth using.

The output is only as good as the input. Feed it real customer words, not your own description of your product.

First drafts and variations

Once you have a winning ad, AI is fast at making variations of it: ten new hooks on the same angle, a short and a long version of the body, the same message rewritten for a different customer type. This is where it saves the most time.

The key is that a person chooses and edits. AI drafts tend to sound alike: smooth, generic, a bit too enthusiastic. Read every line aloud and cut what a real person wouldn’t say. Our hook formulas are a good structure to give the AI, instead of asking for “catchy hooks”.

Adapting for formats and placements

Resizing, extending a background to fit a vertical format, removing a background, making a still from a video, writing captions and subtitles. These are routine edits where AI tools are reliable and save hours. Meta’s own Ads Manager also offers some generative features, such as text variations and image expansion. Check what they change before you let them run.

Translation drafts for European markets

AI translation is good enough for a first draft of ad copy in German, French, Italian or Spanish. It isn’t good enough to publish unchecked. Ads depend on tone, idioms and how people in that country actually talk about the product. Have a native speaker review every line, and expect them to rewrite the hooks.

Where AI still struggles

Product accuracy

Generated images often get your product slightly wrong: the label, the cap, the colour, the size. A customer who receives something that looks different from the ad has a reason to complain. For product shots, start from real photos of the real product and use AI for the background and setting, not the product itself.

Real-feeling people

AI avatars and synthetic presenters have improved, but in a feed, many people still notice something is off, and an ad that feels fake loses the trust that makes UGC-style ads work in the first place. Real people telling real experiences remain hard to beat. See our guide to UGC ads.

Knowing what will sell

AI can tell you what ads usually look like. It can’t tell you which of your ten new ads will win in your market next month. That still comes from testing with real budget. Treat AI output as hypotheses, and let results decide. Our creative testing guide shows how.

Originality

Ask an AI for a “scroll-stopping hook for a protein powder” and you’ll get the same few lines everyone else gets. The more specific your input (real customer quotes, a real mechanism, a real offer), the more original the output.

Rules to keep AI ads honest

These aren’t optional:

• Never generate fake reviews or testimonials. Not as text, not as a synthetic person “saying” them. Presenting invented customer experiences as real is misleading advertising under EU consumer law, and fake reviews are explicitly banned.
• Never invent claims. AI will happily write “clinically proven” or “doctors recommend” for a product with no such evidence. Every claim must be one you can prove, and in health, beauty and food, one you’re allowed to make.
• Don’t imitate real people. No synthetic voices or faces of real people without their clear permission.
• Label synthetic people and voices. If a realistic person or voice in your ad is generated, make that clear. Rules on labelling AI-generated content are tightening in the EU, and platforms have their own policies on it.
• Check rights. Make sure you have the right to use what you feed in (photos, music, footage) and what the tool gives back for commercial use.

A simple workflow for a small team

Here’s how AI fits into one creative cycle without taking over:

1. Research (AI + person): AI summarises reviews and a swipe file of competitor ads. A person picks the one angle to test.
2. Brief (person): One page: the angle, the customer, the proof, what not to say.
3. Drafts (AI): Ten hooks and three body versions from the brief.
4. Edit (person): Pick the best three hooks, rewrite them to sound human, check every claim.
5. Production (both): Real product photos or real people on video; AI for backgrounds, resizing, captions and cut-downs.
6. Test (budget): Launch as a batch, judge with written rules.
7. Variations (AI + person): For the winner, AI drafts new hooks and formats; a person approves them.

AI makes steps 1, 3, 5 and 7 faster. It doesn’t replace the choices in steps 2, 4 and 6. That’s the part that decides whether the ads work.

This is how we run AI ad creatives for clients. If you’re doing it yourself, the free Paid Ads Workbook includes a step for writing the rules you give the AI about your brand, so its drafts start closer to how you actually talk.