On 31 July 2026, Snapchat decided that wholly AI-generated video is no longer eligible for recommendation on Spotlight.
Search Engine Journal reported the change, and the wording Snap used is the important detail. The content is not banned. It is not eligible for recommendation.
Those are different sentences with very different consequences. You can still post it. The platform simply will not carry it.
Demotion is a smarter weapon than removal
Removal creates a fight. Somebody appeals, somebody screenshots the takedown, and the platform spends a week defending a judgment call about a single video.
Demotion creates silence. Your reach declines, no notification arrives, and there is nothing to screenshot. Most creators will conclude the algorithm is being difficult this month.
This is why I expect demotion to become the standard mechanism rather than the exception. It is cheaper to operate, harder to contest, and it lets a platform express a preference without publishing a rulebook that gets gamed within a week.
Snap framed the reasoning around authenticity, saying it wants Spotlight to stay a place where people discover creativity from real people, and pointing at original perspectives and personal storytelling as what it intends to reward.
The audience data behind the decision
Platforms do not make decisions like this out of principle. They make them out of retention data.
Snapchat’s core demographic is Gen Z, and roughly 42 percent of that group reports anxiety about AI. When your most valuable cohort has a negative reaction to a category of content, promoting that category is a retention cost.
The timing is not accidental either. The change lands days before the EU transparency obligations applied on 2 August, which push artificial origin toward being machine-readable.
Put those two facts side by side and the sequence becomes obvious. Regulation makes provenance detectable. Detectability makes provenance rankable. Audience sentiment decides which direction the ranking goes.
Assisted and generated are now separate categories
The policy targets wholly AI-generated content. Human-directed work that uses AI tools sits in a different bucket, and that distinction is where your content operation should be paying attention.
For anyone producing volume, the practical implication is a change in method rather than a change in tooling. The winning approach is not less AI. It is more identifiable human presence inside the work. AI can write, but it cannot know your client, and that gap is now a distribution advantage.
A real face. A real voice. A location, a scene, an opinion that a model could not have generated because it required somebody to have been somewhere and decided something.
We run content operations at agency volume, and the shift is already visible in how we scope projects. Generation is cheap and getting cheaper. Evidence of authorship is now the expensive input, which means it is the one worth budgeting for.
The word wholly is also going to be contested, and platforms will not publish the threshold. Nobody will tell you whether an AI-generated background behind a real presenter counts, or whether a synthetic voiceover on human-shot footage crosses the line.
That ambiguity is deliberate. A published threshold becomes a specification for getting as close to it as possible, which is precisely what the policy is trying to prevent.
So the safe operating assumption is the strict one. If a piece of content contains no moment that required a person to be somewhere, decide something, or say something in their own voice, treat it as ineligible for organic reach and plan its distribution accordingly.
What this costs you if you ignore it
The risk is not a penalty. It is a slow decline nobody attributes correctly.
A team ships fully synthetic video, reach softens over a few weeks, and the internal explanation becomes a creative problem or a seasonality problem. Nobody connects it to a policy change on a platform that never sent a notification. It is the bill for cheap AI content, arriving as a reach decline instead of an invoice.
So build the diagnostic now. Segment your output by how it was made, track reach and completion separately for each segment, and you will see a divergence before you see a crisis.
Tag every asset at publication with one of three labels: human-shot, AI-assisted, fully generated. It costs a dropdown in your workflow and it is the only way to attribute a reach decline to a cause rather than a mood.
Do it for a full quarter before drawing conclusions, since platform recommendation systems move slowly and a two-week sample will tell you whatever you want to hear.
Cold data beats a theory about the algorithm every time, and it settles the internal argument faster than any opinion in the room.
The teams that get hurt here are the ones that scaled synthetic output hardest, because their entire cost structure assumes the reach was free.
The platforms just started paying for something they cannot generate. Make sure you are producing it.