Meta's generative AI for ads: what works and where the limits are

(updated ) 6 min read meta-adscreativeai
AI images, text, video, music and dubbing in Advantage+ creative. What works in practice, where the limits are and how to use it now the AI Act applies.

Over the past year and a half, Meta has packed its ad production with generative AI. Inside Ads Manager you can now generate images, replace backgrounds, turn static photos into video, have headlines rewritten and even add music and dubbing. According to Meta, more than 4 million advertisers now use these tools. The question is no longer whether you’ll come across them, but how to use them well. In this post I lay out what works, where the limits are and what to watch for now that the AI Act is in force.

What’s in Advantage+ creative now?

The generative features mostly fall under the name Advantage+ creative enhancements. These are the main ones you’ll come across in practice:

Images

Background generation places your product in a new setting based on a prompt, without needing a photo studio. You can also expand images (image expansion) so a photo fits neatly into every placement format, and generate separate image variations. Since May 2026, Meta groups those variations into themes such as “refined product look” and “popular in your niche”, which makes testing faster.

Text

Meta generates variations of your headline and primary text, based on what you enter yourself. Handy for quickly getting more angles into the auction, but watch the tone. More on that below.

Video

The biggest leap is in video. With Video Generation 2.0, announced at Cannes Lions in June 2025, Meta stitches up to 20 product photos together into a multi-scene clip with text overlays and music. A separate feature turns a single static photo into a short animated video.

Audio

Two newer additions: royalty-free music that Meta puts under your generated videos, and AI voice-overs that translate your video into other languages. Meta has been rolling out that translation feature more widely since mid-2026, and it wasn’t one of the eleven features from Cannes 2025.

What works in practice?

My honest experience: the generative tools are strong as an accelerator, not as a replacement.

What works well is volume and variety. Now that Meta’s algorithm uses your creative to decide who sees your ad, you need more ads that are genuinely different. AI helps you produce several backgrounds, headlines and formats quickly, so you have something to test without your production process grinding to a halt.

Background generation for catalog ads also makes a measurable difference. Meta itself cites Casetify, which saw 13 percent higher ROAS in tests with background generation. Those are Meta’s own numbers, so take them with the usual grain of salt, but the direction matches what I see in accounts: it’s a free extra layer of testing.

Image-to-video is valuable if you have little video material. A webshop with hundreds of product photos and zero videos can suddenly send moving images into the auction. That’s exactly the kind of gap these tools close.

Where the limits are

Now the less fun part, because this is where things go wrong in practice.

Fine creative control is limited. You steer with prompts, but the model interprets. Short prompts, a forgotten aspect ratio or contradictory adjectives produce messy output. The result is only as good as your input plus the model’s interpretation, and you don’t control the latter.

AI-generated text sands off nuance. Headlines turn flat and generic, and the framing rarely matches exactly what you would write. Use the text variations as raw material, not as the finished product. The same goes for the new AI instant forms from May 2026: treat them as a first draft you edit yourself.

And video is still partly in testing. Meta is rolling it out in phases, and quality varies a lot by industry and source material. Don’t count on one click producing a production-ready commercial.

The biggest pitfall is handing your creative direction over to a model. AI images are fine for varying and testing faster, but they don’t decide what your brand looks like. That responsibility stays with you.

Watch out: AI labels and the AI Act

The EU AI Act sets hard rules: the transparency obligation in Article 50 is enforced from August 2, 2026. You may not use AI-generated images that reproduce a real, recognizable face without explicit consent. And AI-generated images that show people must be labeled with the “AI-generated content” toggle in Ads Manager.

Meta also scans every ad itself and automatically adds a label to photorealistic AI images it detects through C2PA metadata or its own detection. You can’t get that label removed. Accounts that don’t disclose AI images correctly risk having their ads rejected. Meta hasn’t published an official suspension date, but waiting for enforcement isn’t a strategy. So this is no longer an optional checkbox. It’s part of your compliance.

From October 30, 2026, this becomes contractual. On that date, Meta adds the terms for generative AI in ad creative to the Self-Serve Ad Terms. You are responsible for whatever the AI features add to your ads, and you may not remove watermarks or metadata showing that something was made with AI.

How I use the generative tools

Concretely, this is what I would do:

  • Use AI for speed and volume, not for direction. Let it churn out backgrounds, variations and formats, and pick the winners yourself.
  • Rewrite every AI text. Use the variations as a starting point, but put your own brand voice and proposition back in.
  • Give enough context in your prompt. Describe the product, how it’s used, the mood you want and the aspect ratio. A thin prompt gets you thin output.
  • Fill your video gap with image-to-video. If you have little moving material, this is a logical first step, as long as you review the output critically.
  • Get your AI labels in order. Flag AI images with people correctly and assume Meta will label them itself if you don’t. Not disclosing is a suspension risk.
  • Keep testing on results. Let the algorithm finish learning and judge on rolling trends, not on a single day.

The common thread: generative AI lowers the barrier to producing a lot, but it doesn’t automatically raise the quality. Your value lies in selection, positioning and judgment. The machine makes, you choose.

Want to know whether your Meta account uses the generative tools well, or mostly sends mediocre variations into the auction? Request a Free Performance Quick Scan and I’ll take a look with you.

Sources

Free Quick Scan

Curious what this could bring you?

I review one ad account with read-only access and show you your 3 biggest opportunities. The report is yours to keep.

or book an intro call

Want to win back return?

Start with a Quick Scan and find your 3 biggest opportunities.