A Business Insider investigation has exposed what many advertisers already suspected: Meta's generative AI advertising tools are producing embarrassing results. Nonsensical copy appears alongside images with distorted limbs, while products end up resembling warped versions of the originals. And Meta's response, in essence, is a digital shrug: "That's your problem."

The story, pieced together from internal documents and testimonials, shows how the platform's push toward fully automated campaigns — known as Advantage+ — is generating not just creative disasters but also a dangerous accountability gap. Mark Zuckerberg's company spends billions to convince brands to hand ad creativity over to its AI, promising efficiency and at-scale personalization. But when the results are patently unacceptable, the implicit message is that advertisers are on the hook for quality control, gutting the promise of full automation.

The dark side of unsupervised generative AI

This isn't a temporary technical glitch. The defects — copy with no logical thread, limbs that multiply or merge, distorted logos and products — are the fruit of generative models designed to produce variations from limited inputs, with no real understanding of context. Left unsupervised, the Large Language Models and vision models powering Meta's tools can "hallucinate" outputs that look credible to a statistical algorithm but appear grotesque to human eyes.

Meta chose to drop these models into an advertising environment where volume matters more than precision, hoping that advertisers, lured by lower cost-per-click and greater reach, would accept a certain error rate. The problem, as flagged by Business Insider, is that the error rate is far from negligible.

Who's to blame? The passing-the-buck game

Meta's official stance — or rather, the reaction described by inside sources — is emblematic of a business model where the platform provides increasingly opaque tools but disclaims any responsibility for the output. It's as if a creative agency handed over a storyboard full of mistakes and told the client, "You can always fix it." Except here, the human editor would have to sift through hundreds of automatically generated variations, wiping out the promised time savings.

This attitude shifts the quality burden entirely onto brands, which have zero visibility into the models' inner workings. They can't do targeted fine-tuning, can't set quality thresholds, can't preemptively exclude certain hallucination categories. They are hostages of a black box.

What this means for those weighing control over their AI tools

The Meta episode isn't an isolated case; it's a symptom of a structural tension. When AI is delivered as a centralized cloud service, control over outputs is inversely proportional to ease of use. For companies in regulated sectors or those whose value rests on brand equity, entrusting advertising messages to unverifiable models is a reputational gamble that can't be priced in advance.

Not every organization can afford to let garbled text or deformed products appear next to its logo. Some are already evaluating architectures where generative models run on their own infrastructure, allowing for quantization, fine-tuning, and strict guardrails. Choosing an on-premise deployment for advertising workloads is neither simple nor cheap: it requires skills, adequate hardware, and a careful TCO assessment. But the transparency and predictability gained may become a competitive advantage after episodes like this.

AI-RADAR focuses precisely on these decisions: analyzing the trade-offs between the convenience of cloud platforms and the sovereignty of a self-hosted system. Without endorsing one approach over the other, but providing the critical lens to understand when the promise of automation risks turning into a reputational liability.

In short, the Meta episode reminds us that automation without control is merely deferred cost. And brands that are now told "it's your problem" may decide, tomorrow, to hold onto the keys to their own systems.