The 4 AI Questions
AI ad delivery does not spin out of control because AI is too powerful. It happens because you did not give it boundaries.
A case has been circulating in the industry recently: an evergreen creative that had been running for three months was swapped overnight by the platform's AI. The budget kept spending.
When the media buyer opened the ad account in the morning, the carefully tested product hero image had been replaced by a completely unrelated AI-generated image. There was no operation record in the backend, and the change was not visible in the preview.
This was not an isolated case.
We asked the same question to three media buyers on our team—one running short-form drama, one running consumer products, and one running financial services, each managing a different monthly budget size:
When ad delivery is handed over to AI, what are clients most afraid of?
Their four answers were highly consistent: creative assets become random, generation gets out of control, content triggers compliance risks, and brand tone collapses. But the solutions from the three media buyers were completely different.
This article is not about whether AI ad delivery is useful. That question is no longer worth debating. Every major advertising platform is moving toward automation. Whether you turn off Advantage+ or not, Meta will continue handing more decision-making power to algorithms.
The real question worth discussing is this:
Where are the controllable boundaries of AI ad delivery?
Will creative assets become random?
The judgment from all three media buyers: partly true.
If you dump dozens of creative assets across different styles into the system all at once, AI will indeed start aggressively mixing and matching during the cold-start phase: pairing Video A with Headline B and Description C, testing combinations almost like drawing lots. The sparser the account data, the stronger this sense of randomness becomes.
But this is not an AI problem. It is a feeding problem.
The media buyer running short-form drama shared their approach: intent-based grouping. Different types of creative assets are strictly separated by audience intent. Male-oriented and female-oriented content are run in separate accounts, different plotlines are split into different campaigns, and the copy and video within each group remain tightly bound to each other. AI is not given the opportunity to pair assets across groups.
The result: for the same product, after switching to intent-based grouping, first-day ROI volatility dropped by 35% compared with the “dump-everything-into-one-bucket” approach.
The media buyer in financial services took another route: every creative iteration and modification was logged for the system, while accepted and rejected outputs were fed back to the algorithm engineers so the system could learn what a “good” creative looked like. After several rounds of calibration, the creative adoption rate for the same team increased from 30% before the process was introduced to 80%.
The two approaches look different, but the underlying logic is the same:
You are not limiting AI’s capability. You are narrowing AI’s choice space.
When you reduce the number of possible combination paths from 10,000 to 100, AI’s “randomness” becomes “exploration within a reasonable range.”
Will generation get out of control?
The judgment from all three media buyers: absolutely true.
This is the issue that gives media buyers the biggest headache. The underlying logic of advertising-platform AI is to pursue a local optimum in click-through rate and conversion rate. If the system finds that a bizarre AI-generated image gets a higher click-through rate than your carefully designed brand asset, it will shift budget toward that image, or even replace your original creative outright.
That is exactly what happened in the case that has been widely discussed in the industry: a hit ad that had been steadily delivering returns was automatically replaced by the platform with an AI-generated image. The style was completely off-brand, but spend was still surging. When the client saw their own ad while scrolling through the app, their first reaction was:
“This is not the ad we launched.”
One media buyer used a particularly accurate analogy:
AI is more like assisted driving than autonomous driving. Humans still need to keep their hands on the wheel. AI can help press the accelerator and find the route.
In practice, the team developed a mechanism we call the “hit-creative isolation layer.” Once a creative is proven to be a breakout winner, it is immediately removed from exploratory campaigns where AI has a high degree of freedom and moved into an evergreen manual campaign with all AI enhancement features turned off. Budget is rigidly controlled, and the creative is physically protected.
Does this sound like “going backward to the manual era”?
It is actually the opposite. This is using human judgment to protect the results produced by the machine. Evergreen creatives are guarded by people, while new creative testing is still handed over to AI for exploration. The two tracks run in parallel without interfering with each other.
Will content trigger compliance risks?
The judgment from all three media buyers: high risk, and it must be managed.
This issue has two layers.
The first layer is the platform black box.
In certain automated creative modes, the platform may break apart and recombine your copy in the backend, or even automatically generate new music and video cuts. The key problem is that these derivative assets are not visible in the media buyer’s daily preview. Only after the ad starts spending, or after the client sees it inside the app, does the team discover that the creative actually being served is not the same one that passed review.
The client’s reaction is direct:
“We approved A. Users saw B.”
The second layer is legal compliance.
New York State’s AI disclosure law took effect on June 9. It requires prominent disclosure for any ad running on channels such as Meta, Google, or TikTok if it contains an AI-generated “synthetic performer.” Missing the label on even one ad is a violation.
For teams using AI-generated assets in bulk, this means every creative needs a compliance checkpoint before launch. Most teams currently do not have this step.
The frontline solution has two parts:
Technical isolation: turn off the platform’s “flexible format” and other high-risk optimization options. Limit AI’s permissions to safer operations such as size adaptation and brightness adjustment, leaving no room for it to re-edit videos or add new music.
Process moved upstream: establish an AI creative audit SOP. Any creative containing AI-generated elements must be labeled before upload, and the platform’s AIGC disclosure toggle must be selected. For sensitive industries such as financial services, add one more layer of manual review.
This is not conservatism. It is professionalism.
The cost of compliance is far lower than the loss caused by a single account incident.
Will brand tone collapse?
The judgment from all three media buyers: it depends on the category, but at the brand-asset level, yes.
This is where the underlying conflict is deepest.
Advertising-platform AI is extremely utilitarian. If the same campaign contains both high-quality brand assets and vulgar, borderline creatives, the latter will often get higher click-through rates in the first few days because of the curiosity effect. The algorithm then determines that the vulgar creative is “better” and starts shifting more than 80% of the budget toward it.
This is data hijacking brand tone.
Short-term local metrics may appear to improve, but the quality of users attracted by these creatives collapses. Backend ROI and long-term user value decline at the same time.
Another hidden problem is the “machine smell” of AI copy. AI-written ad copy is grammatically correct, but it is full of templated expressions such as “Unleash your inner...” and “You won’t believe what happens next!” For categories that rely on emotional resonance—short-form drama, novels, and content products—this coldness directly reduces completion rate and willingness to pay.
Frontline teams do not respond by banning AI. Instead, they install three layers of guardrails around it.
Visual guardrails: turn the brand’s color matrix, layout safe zones, and whitespace ratios into constraint parameters within the workflow. AI can only make adjustments inside these “boxes.” It cannot improvise freely.
Language guardrails: abandon generic prompts. Build a dedicated corpus for each product, using large volumes of local, real user comments and native ad copy as input samples. Force AI to avoid common “AI words.”
Budget guardrails: physically separate brand-tone creatives and scale creatives at the account-architecture level. Set manual budget caps so the algorithm cannot drag the brand into a ditch for the sake of click-through rate.
The media buyer running consumer products summarized it bluntly:
AI is not good at creating a brand from nothing. But it is very good at rapidly amplifying and iterating on brand content that has already been proven effective.
So the right division of roles is this:
Keep the core of the brand in human hands. Let AI handle expansion.
The real question is not whether AI can be controlled
After writing through these four points, a clearer picture emerges:
Every concern clients have is real, but every concern also has an executable boundary.
Interestingly, all three media buyers mentioned the same thing: the key to controlling AI ad delivery is not giving it more data or a better model. The first step is turning the rules in your head into SOPs.
Intent-based grouping, hit-creative isolation, upstream compliance audits, and brand-parameter packaging—none of these are technical breakthroughs. They are the experience of senior media buyers broken down into executable operating standards.
In other words:
Instead of asking “Can AI ad delivery be controlled?”, ask “Have you turned the media buyer’s judgment into rules the whole team can reuse?”
In the past, the answer was: “Find a great media buyer.”
But people leave. People forget. People’s judgment declines when they are managing their 15th market.
Documenting these experiences, turning them into SOPs that a new team member can follow on their first day—this is the real endgame of controllable AI ad delivery.
