Xie Shi Min wrote a young adult novel called Ghost Hunters of Geylang. It deals with generational trauma and abuse, wrapped in a ghost story set in a Singapore neighbourhood. Serious book, done properly, the kind of thing an author spends years getting right.
In March she found it on Shopee with an AI-generated video attached. An AI avatar was spruiking her book as a "fun" and "wholesome adventure" for children. It also claimed readers could "smell durian" from the pages, because durian gets one mention in the entire book and has nothing to do with the plot. The sword on her cover, the one her illustrator designed as a symbol for the story, had been warped into what looked like a pole.
She didn't make this video. Nobody at her publisher made it either. Shopee's AI made it, automatically, off the back of her own product listing. And it put the thing live without asking anyone.
"To have it be put through AI is just plain insulting," she told CNA. Fair call, mate. Someone spent years writing about abuse and trauma with care, and a machine decided the vibe was "kids having ghost adventures, smells like durian."
Her publisher, Epigram Books, reported the video. Shopee has settings to turn off AI-generated images on a listing, but not video, and even switching off what you can doesn't clear existing clips. When Xie checked her listing again in August, five months after the first complaint, there was a different AI video up. "It's like a Hydra," she said. "You kill one video and then another comes up."
She's not the only one. A paper goods designer named Koh Wan Yi found AI-generated people showing off her yearly planner on TikTok Shop, except the AI had made the calendar look enormous next to the person holding it, in the wrong colours, sometimes narrated in Spanish for a product that only sells in Singapore. She asked her TikTok account manager if she could block the affiliate creators posting the dodgy clips. Couldn't. The best she got offered was switching off "open collaboration" entirely, which also kills the legitimate creators she actually likes working with.
Here's the bit that should worry you if you sell anything online. Neither of these women asked for this content. They can't fully switch it off, and they can't fix it when it's wrong. The platform generated marketing material using their product, their name, and their brand, then left them to clean up the mess in the reviews and the DMs. Epigram's marketing manager put it simply: authors started messaging the publisher asking "how could you do this?", assuming the dodgy AI video was their doing. It wasn't. But try explaining that to a customer scrolling past it at 11pm.
So here's the trap. Any tool that generates customer-facing content on your behalf without a human checking it first will eventually do this to you as well. It's not that AI can't write a product description or cut a product video. It's that nobody looked at the output before it went in front of a paying customer. A pole instead of a sword is sort of a rendering glitch, forgivable. Calling a book about abuse "wholesome" is a category error a five-year-old wouldn't make, and it went live anyway because there was no human between the model and the "publish" button.
I run into a version of this every week with clients who've bolted an AI tool onto their sales or marketing stack and let it run unattended. The instinct is understandable. You're busy, the tool promises to save you three hours a day, so you switch it on and walk away. The bit everyone skips is the checkpoint. Somewhere in that workflow, a human needs to glance at what's about to go out before it reaches a customer, at least until you've watched the system get it right a hundred times in a row.
That doesn't mean reviewing every single output by hand forever, that's just moving the bottleneck from "no marketing" to "you're now a full-time editor." It means building the check into the system itself. If you're running AI-generated content anywhere near your customers, whether that's product descriptions, review replies, ad creative, or follow-up emails, ask yourself three things. Can you see everything it's produced in one place, not scattered across a platform's back end? Can you switch it off completely, not just "turn off images but not video"? And if it gets something wrong, how many customers see the mistake before you find out?
If the honest answer to any of those is "I'm not sure," you don't have automation, you've got a slot machine wearing an automation costume. It might land on something fine most of the time. It might also decide your trauma novel smells like durian.
The fix isn't complicated, and it's not "never use AI for content." It's owning the pipeline instead of renting someone else's black box. When Koh wanted control back, the only lever the platform gave her was an all-or-nothing switch that also cut off the genuine creators she wanted. Compare that to running your own content system, where you decide what gets generated, what gets reviewed, and what a human signs off before it touches a customer. Same technology, completely different outcome, because you built the checkpoint in rather than hoping the platform remembered to include one.
There's also a reputation cost that doesn't show up until later. Shopee introduced the AI video feature in late 2025 to help shoppers make more informed purchasing decisions, which is a fair goal. TikTok Shop rolled out something similar for its creators. The problem isn't the intent, it's that both platforms treat the AI output as done the moment it renders, with no step where a human checks it matches reality. Singapore's Advertising Standards Authority has already flagged that shoppers are getting wary of brands using generative AI for misleading ads, which means every dud video isn't just annoying the one seller, it's chipping away at trust in every seller on that platform.
Epigram's marketing manager Chloe Chee told CNA the company has reported roughly ten affected listings since March and still isn't sure which videos actually got taken down, because she can't tell if she's messaging a person or a chatbot on the seller support line either. That's two AI systems failing in sequence, one that generated the wrong video and one that can't confirm it fixed the complaint. If your own business has a support inbox where you genuinely don't know whether a human or a bot answered the last message, you've got the same blind spot, just smaller.
None of this means AI-generated marketing is a write-off. Plenty of small businesses use it well, for first drafts, for variations to test, for scaling one good idea into fifty. The difference between that and what happened to Xie and Koh is entirely about who checks the output before a stranger sees it. Their businesses had zero say in that step. Yours doesn't have to work that way, because you're not renting a marketplace's black box, you're building your own.
Want to see how this could work in your business? Book a call and let's talk about where you're at and what's possible.
Sources:
Brewed by Steven, poured by Viktor
About Steven Tann: Steven helps business owners build systems that run themselves using AI. After 10+ years helping 7,000+ businesses and building his own autonomous operations, he's the bloke who actually does it, not just talks about it. Find out more at steventann.com.