OpenAI shut down the Sora app in a two-sentence social media post in late March, but the reason it collapsed had nothing to do with video quality. The company wrote that it was “saying goodbye to the Sora app,” according to the Associated Pressafter months of pressure over deepfakes of Michael Jackson, Martin Luther King Jr. and Mister Rogers forced OpenAI to make reactive takedowns before the intervention of family estates and an actors union. Sora didn’t fail because the model couldn’t generate a compelling video. It failed because no one had built the infrastructure of trust around it before letting the public run amok on the chat box.
That same week, YouTube’s app data told a similar story in the opposite direction. In January, the platform permanently deleted 16 channels as part of what it now calls its Inauthentic Content Policy, a July 2025 renaming of the old rule on repetitive content. These channels had a total of 35 million subscribers and 4.7 billion views over their lifetime, and they produced mass-produced modeled videos without any human editorial input, according to a report from The Hollywood Reporter.
Both stories are about the same failure. It’s also not really about AI video getting better or worse. They talk about what happens when the scale exceeds the human judgment that is supposed to be above it, and that gap is exactly what I built my 5-pillar framework for AI content close in April. Four months later, it doesn’t need much updating.
The cost of scale just dropped again, raising the stakes
Two days before Sora’s shutdown made headlines, Google published a blog post announcing I see 3.1 Liteits most cost-effective video generation model, at less than half the price of Veo 3.1 Fast for the same speed. Developers can now generate four-, six-, or eight-second clips in landscape or portrait mode at up to 1080p, designed specifically for high-volume applications.
This is not a criticism of the tool. This is a fact that deserves to be verified in the field. The cost of producing video at scale continues to fall, which means that the pressure that Pillar 1 of my framework was designed to address – the temptation to treat AI as a shortcut rather than infrastructure – will only increase. A cheaper generation makes strategy-driven discipline more necessary, not less.
The human face has become a signal of confidence, not just a style choice
Craig Billings runs a science channel called Doctor NOS with 1.7 million subscribers, and he told The Hollywood Reporter that anonymous channels covering his same territory are being hit hard by the crackdown. Most of them are demonetized, he said, while creators who have never touched AI but never shown their faces find themselves caught in the same net.
This is the real cost of imperfect enforcement, and it’s worth mentioning honestly, but it also confirms something my framework has already argued in Pillar 5. YouTube’s own policy page, “How creators are using AI for content creation“, clearly indicates that the platform requires creators to disclose when AI was used to edit or generate realistic contentand that labels can appear on the video player for Shorts or under long videos. If a creator ignores the disclosure and YouTube’s systems detect the AI anyway, the label is applied automatically and creators cannot remove it once it is highly confident that it was created by the AI.
Four months ago, I wrote that hiding the use of AI was considered a weakness to a sophisticated public and that disclosure was considered a skill. It is no longer just a confidence strategy. It’s now built into the platform’s actual infrastructure, and treating it as an optional PR tweak is a strategic mistake, not just a missed opportunity.
What a working AI video actually looks like
Compare slope channels with what Think with Google offers Creativity Edition Guide documents. Google Creative Labs Matthew Carey describes the construction of the AI-assisted short film ANCESTOR deliberately avoiding generic prompts, prompting taking pictures of the cosmos using descriptions of specific microscopes and lights rather than the word cosmos itself, because the obvious prompt produces the default visual average of each model. Co-founder of Monks Wesley Haar, ter told the same publication that brands that have succeeded with AI have done the inglorious work of codifying exactly what their brand is before generating a framework.
Neither example treats AI as a volume machine. Both treat it as an execution capability that relies on a specific human decision regarding what belongs on the screen and what does not. It’s Pillar 1 and Pillar 5 working together, and it’s the difference between channels removed by YouTube and the case studies that Google now touts as the industry standard.
The trust gap is wider in the United States than in the United Arab Emirates
There is one market dimension that American marketers tend to underweight. In a YouGov survey of 19 markets I covered in JulyThe United States had the lowest rate of AI-assisted search of all countries tested, at 48%, compared to 89% in India, Indonesia, and the United Arab Emirates, and only 28% of U.S. searchers said they trusted an AI assistant’s response.
I teach a module called “Engaging Audiences with Content in the Age of AI” at the New Media Academy in the UAE, in a region where AI-powered discovery is already the norm rather than the exception. The lesson is not that Americans are wrong to be skeptical. This is because the disclosure and human judgment requirements integrated into Pillar 5 are not regional assets. This is the baseline a skeptical American audience needs and a receptive Emirati audience will expect it anyway once the app catches up in adoption.
3 updates to make before your next AI video goes live
First, check whether your disclosure practices follow the platform’s actual policy language.not your internal comfort level. YouTube’s own guidelines state that labels apply to photorealistic or significantly edited content, and that creators lose the ability to remove that label once the system flags it with high confidence.
Second, evaluate your production plan against what tools like Veo 3.1 Lite now make possible at scale, then deliberately choose to produce less than the ceiling allows. The technical ability to generate a thousand variations does not require you to publish a thousand variations.
Third, name the human decision maker on each AI-assisted part before delivery, like Carey’s team did on Ancestra and ter Haar’s team did with Monks brand knowledge bases. If no one can answer who decided it was the right cut, the part is not ready.
My opinion
The conversation about AI slop in this industry continues to be framed as a content quality problem, and I think that framing is wrong. This is a trust infrastructure issue, and Sora, the YouTube purge, and the Veo cost drop all point to the same gap from three different angles.
What I argued in April is valid. The only thing that changed was that the platforms stopped responding. Meaning cannot be automated, and the tools that evolve the fastest are the ones that make it most tempting to skip the human checkpoint. My framework didn’t need to be rewritten this fall. It was enough for the industry to catch up with Pillar 5.
More resources:
Featured image: Gorodenkoff/Shutterstock
