Meta is bringing in a landscaper to their walled garden.
Muse Image, Meta’s AI image generation tool, is still on its way to Advantage+ creative, giving advertisers another way to generate and optimize creative without ever leaving the platform. On the surface, there’s plenty to like. More creative variation, faster production, less resource needed to keep campaigns fed with new assets. For brands that struggle to produce enough creative, I think it’s absolutely worth a test. But there’s a trade-off here that I don’t think we’re talking about enough: the more creative decisions we hand over to platforms, the less we may understand about what’s actually working.
Creative testing has always been valuable because you leave knowing more about what works. We change a variable, see what happens, and use that information to make the next piece of creative better. Maybe it’s the hook, the product shot, the format, the message or the CTA. Whatever it is, there’s something we can take away from that campaign and apply to the next one.
That gets much harder when the creative is constantly changing. If Meta is generating variations and tweaking creative throughout a campaign, without an approval process or detailed reporting on exactly what changed and why, we might know that performance improved without knowing what actually improved it. If something is always changing, you reduce your ability to isolate the thing that worked in the first place. Eventually, creative learnings start becoming less like actual learnings and more like “Meta found something that worked.”
And that matters beyond the performance of one Meta campaign. Creative learnings should help brands get smarter about their audiences. They should inform what you do on TikTok or YouTube, what goes into the next brief, what your creative team produces, and what you decide to test next. If all of that learning stays inside Meta’s walled garden, we might be making the garden nicer, but I’m less convinced we’re becoming better gardeners.
AI-generated creative isn’t the problem
To be clear, I think AI-generated creative has a lot of value for production and optimization. Where I draw the line is assuming that means the process needs to happen inside the media platform.
I’d much rather brands have their own relationships with AI and creative partners, decide their own comfort levels, and manage that process outside the platform. That gives them more control over what gets generated and how it gets used, but it also gives them a better chance of actually understanding what they’re learning along the way. Every brand is going to have a different appetite for AI-generated creative, and they should be able to make that decision themselves rather than being pushed toward whatever a platform decides is the best way to optimize.
The bigger question for me is how much of the process we’re comfortable putting in one place. Meta already has an enormous amount of performance data. It delivers the media, controls the optimization and reports back on the results. As it moves further into generating and adapting the creative itself, another part of that process moves behind the same walls. Creative generation, media delivery, optimization and measurement can increasingly happen within one ecosystem.
There are obvious advantages to that. Meta has access to a scale of performance data that an individual advertiser doesn’t, and using that data to make faster creative decisions could create efficiencies that would be difficult to replicate manually. But advertisers should understand the trade they’re making. The more Meta learns on your behalf, the more important it becomes to ask how much of that learning actually makes its way back to you.
Efficiency is easy to measure. Sustainable growth is harder.
The obvious answer to all of this is performance. If handing over more control makes campaigns more efficient, does it really matter that we understand every decision the platform made along the way?
Maybe. But I think we have a challenge proving that these tools actually improve performance over time. I’m sure there will be some vanity metric or source to say it’s working (as platforms tend to do), but I’m more interested in the harder question: where is the brand that has genuinely driven incremental year-on-year growth because it handed more of its creative decision-making over to the platform?
That’s a much higher bar than making an individual campaign more efficient. My concern is that we could end up gaining efficiency at the expense of sustainable performance and growth. A system can get very good at optimizing toward the signals it can see, but if the advertiser comes out of that process with less understanding of its audience and fewer learnings it can apply elsewhere, I’m not sure we should automatically call that progress.
This is also why I don’t think the answer is to avoid tools like Muse Image altogether. If a client asked me whether they should use it, I’d say yes, test it, particularly if creative volume and variation are already a challenge. Everything is worth a test when there’s a clear reason for testing it. I just wouldn’t hand over the whole campaign on day one.
I’d start with a specific subset of activity, establish what we’re trying to learn before the test begins, and make sure there’s something meaningful to compare it against. And I’d go into it accepting that it might not work. Testing new technology means being comfortable with some of that investment struggling or failing; otherwise, it isn’t much of a test.
Muse Image is one product, but the choice it represents is going to keep coming up. Platforms are going to get better at generating creative, optimizing it and making decisions on advertisers’ behalf, and there will be plenty of situations where letting them do that makes sense. The question for advertisers is how much control they’re willing to relinquish in exchange for that efficiency, and whether they’re still learning enough to make better decisions themselves.
AI can absolutely help us make creative faster and at greater scale. I’m here for that. I’m just not ready to outsource the learning with it.
