Human, meet loop

A human-in-the-loop (HITL) workflow means more than someone glancing at an AI-generated asset before launching it into the world.

In marketing, a robust and defined process is key. And yes, that likely means counting the appropriate number of fingers in images, or making sure that your brand’s tagline hasn’t been contorted into textual mush for a quick social campaign.  

The media loves a whimsical AI fail, and the news remains full of people using AI poorly—often with mediocre or hilarious results. While it can be easy for observers to chalk this up to the underlying unreliability of current AI systems, it should instead simply be a reminder for the importance of human-in-the-loop processes.

Let’s dive in.

Two fresh “don’t-let-this-be-you” moments

• A New York Times exposé follows Northeastern University senior Ella Stapleton, who opened her teacher’s lecture notes to find the professor’s raw prompt—“expand on all areas. Be more detailed and specific”—sitting mid-document. A slide deck for the same class flaunted AI-generated stock photos of office workers with extra limbs, garbled gibberish on AI-created infographics, and bullet-point typos. Group-chat panic ensued, and Stapleton demanded an $8,000 tuition refund.
  
  
• A new Coca-Cola AI-assisted campaign highlighted authors who name-dropped the soda brand in their work. It included a quote attributed to JG Ballard, screwing up the fact that the words in question were from an interview the author gave—not one of his novels. What’s more, the name of the city “Shanghai” was misspelled on-screen in the spot.

Neither scenario above would have made waves if basic HITL protocols were in place to add guardrails to AI-created assets. 

Let’s talk about text

GenAI finally nailed five-fingered hands, yet it can still turn WELCOME HOME into WLCØM HØM. As technologist Markus Brinsa explains, “The model sees letters as pixel shapes...[rather than characters meant to convey meaning]...so kerning and spelling crumble.”

Until hybrid pipelines separate typography from imagery, every AI-rendered asset needs both a designer’s eye and a copy-editor’s brain.

Consider one prominent fail in this regard: a Skechers ad in Vogue circa December, 2024. The spread showed two glam models… wearing high heels, not Skechers, for some reason. Close-ups revealed gibberish signage and melted faces; a TikTok roast of the asset hit 2 million views. All press is not good press.

If you think you can use GenAI tools like Propellers to create one-click, production-ready assets with subtly rendered text, you’re risking a high-profile flop. We don't expect our human co-workers to be perfect 100% of the time, and we shouldn't assume GenAI is, either.

Consider the assets you spin up to be solid first drafts that still need human review and polish, especially when letterforms are in the mix.

Set expectations & fix the workflow

Why do we shrug when baristas misspell our name, but meme an AI for one typo?

AI bloopers aren’t proof the field is snake oil; they’re proof someone skipped QC. And HITL isn’t a concession to AI’s weakness—it’s the professional standard we already apply to human work. 

Brands that combine speed-boosting automation with disciplined oversight will thrive. Everyone else? Prepare for your cameo in the next “AI-fail” thread.