MANY MINDEDFour ways to tell an AI demo was faked
In December 2025, several dictionaries independently landed on the same word of the year. Merriam-Webster, the American Dialect Society and The Economist all chose "slop"; Australia's Macquarie Dictionary chose "AI slop". The word describes low-quality synthetic content produced at volume and filling the feed.
One variety of slop does more damage than the rest: the staged capability demo. A video claiming a result came from one prompt, one take, no cuts — where the cuts are the entire product.
The reference case is not an anonymous account. It is Google. In December 2023 the company published a "Hands-on with Gemini" video showing what looked like a fluid, real-time spoken conversation with a model watching the world through a camera. A Google spokesperson later told Bloomberg the video had been made "using still image frames from the footage, and prompting via text." There was no live interaction. The voice was dubbed in afterward. The YouTube description did carry a line noting that "latency has been reduced, and Gemini outputs have been shortened for brevity" — which is precisely the point. A disclaimer technically existed. The video still implied a capability that did not exist as shown.
That gap, between technically disclosed and practically misleading, is where nearly all of this lives.
Four tells, none of which require expertise:
One. It is a screen recording, never a link. If you cannot run it, you cannot check it.
Two. No model, no version, no seed. Conditions that are never named cannot be reproduced or falsified.
Three. Nobody in the replies can reproduce it. The comment section is the peer review, and it is free.
Four. The account has never posted a failure. Real work produces failures. An account with none is showing you a cherry-picked tail and calling it a median.
Here is the uncomfortable part, and it cuts against the obvious response. You cannot reliably spot this by feel. In Bynder's 2024 study, when an AI-written article was unlabelled, 56 percent of readers who expressed a preference found it more engaging than the human-written one. Told the same piece was AI-generated, 52 percent reported feeling less engaged. Perception tracked the label, not the content. "I can just tell" is not a defence, which is exactly why the four tells above are procedural rather than perceptual.
Nor is the line simply synthetic versus real. Meta's stated position is that its ranking does not demote content for having been made with AI; what it demotes is unoriginal, templated work. The distinction that matters is disclosed versus disguised.
Why this matters here: this project reads cost and capability curves to say what is actually buildable, and when. Those curves are assembled from reported capability. A staged demo injects noise directly into the measurement. The cost is not offence — it is mis-timing. When everything looks like a miracle, you cannot tell which miracle shipped, and you make worse decisions about what to learn, what to build, and what to wait for.
The honest caveat on our own side: we cannot tell you what share of viral AI demos are staged. Nobody can, and any figure claiming to would be doing the precise thing this article describes. What can be said is that the incentive is real, the reference case is admitted, and the tells are cheap to apply.
So apply them. Then unfollow the accounts that fail all four — not out of spite, but because attention spent on a fake is attention taken from something real that shipped this week.
We publish the curves, the sources, and the friction, counted honestly.
Source: TechCrunch