Was the crowd in Will Smith's concert video AI-generated?
No — the crowds were real. The people came from Will Smith's actual 2025 European festival dates. His team animated real tour photographs into short clips using image-to-video AI, and YouTube's Shorts upscaling experiment added more distortion on top. The melted faces were artifacts applied to genuine audiences, not invented fans.
Why — the first-principles explanation
This case is the best available lesson in why "it looks AI-generated" and "it is AI-generated" are different claims — and why your eyes can't tell them apart.
Start with what actually happened. Smith's team had real photographs from real shows, taken by tour photographers. They fed those stills into an image-to-video model, which takes one frame and invents the motion around it — a few seconds of a still photo coming to life. Here's the critical bit: the model isn't reproducing what happened next, because it never saw what happened next. It's guessing plausible motion. So a real person in a real crowd gets assigned invented movement, and that's where hands smear, faces slide, and a sign's lettering dissolves. The people were real. The motion was hallucinated.
Then a second layer piled on. In August 2025, YouTube was quietly running an experiment that post-processed Shorts with machine learning to unblur, denoise, and sharpen. That processing hit footage that was already artifact-laden and amplified the damage — which is why the same clips looked notably worse on YouTube than on Instagram or Facebook. Nobody at YouTube was faking a crowd; an unrelated enhancement feature just made a distorted video more distorted.
So the artifacts everyone pointed to — extra fingers, warped faces, garbled text — were genuine AI fingerprints attached to genuinely real people. The public reasoned: AI artifacts, therefore AI crowd. The inference felt airtight and was wrong, because artifacts tell you a model touched the pixels, not that a model invented the subject. The proof came from provenance, not from looking harder. Andy Baio matched the video against the tour photographers' original Instagram slideshows. The same audiences, the same individuals. In the sharp originals, garbled signs became legible — one reads "From West Philly to West Swizzy." The couple with the cancer-survival sign appear in Smith's own Gurtenfestival posts. That's the whole lesson: you don't settle these questions by squinting at fingers. You settle them by finding the source.
An example that makes it click
Imagine someone photographs your family reunion, then runs the print through a cheap photocopier ten times. The final copy is streaky and warped. Aunt Linda's face is a smudge, and the banner reading "Happy 80th" now reads "Hoppy 8Oth."
Someone looks at that copy and says: this family is fake, look at the smeared faces. But the reunion happened. The people were there. What's fake is the copying process, not the family. Will Smith's video is that photocopy — a real crowd pushed through an AI animator and then through YouTube's sharpening filter, coming out the other side looking invented. The distortion is real. The conclusion drawn from it wasn't.
Key facts
- The video, showing Smith performing "You Can Make It," was shared on August 12, 2025, captioned "My favorite part of tour is seeing you all up close. Thank you for seeing me too."
- The crowds came from real 2025 European tour dates: Positiv Festival (Orange, France), Gurtenfestival and Paléo (Switzerland), and Ronquieres Festival (Belgium).
- Tour photographers' original Instagram slideshows show the same audiences and same individuals appearing in the video; the couple holding the cancer-survival sign appear in Smith's own Gurtenfestival posts.
- In high-quality originals, a sign that reads as garbled in the video is legible as "From West Philly to West Swizzy."
- Smith's team used image-to-video AI models to animate static tour photos into short clips — described as meaningfully different from fabricating crowds entirely.
- In August 2025 YouTube ran an experiment post-processing Shorts with machine learning to unblur, denoise, and improve clarity; guitarist Rhett Shull raised it on August 21, YouTube confirmed the experiment on August 26, and The Atlantic covered it on August 29.
▶ The 60-second explainer (script)
Was the crowd in Will Smith's concert video AI-generated? No. The crowds were real — and the story of why everyone got this wrong is more useful than the answer. In August 2025, Smith posted a video of himself performing You Can Make It to a huge, adoring crowd. Viewers immediately spotted the tells: melting faces, extra fingers, garbled signs, a hand that moved on its own. The internet concluded he'd faked an audience. Here's what actually happened. Smith's team had real photographs, taken by real tour photographers, at real shows — Gurtenfestival and Paléo in Switzerland, Positiv in France, Ronquieres in Belgium. They fed those still photos into an image-to-video model, which takes one frame and invents the motion around it. And that's the key: the model never saw what happened next. It's guessing plausible movement. So real people get assigned hallucinated motion, and that's where the hands smear and the signs dissolve. The people were real. The motion was invented. Then a second layer piled on. That same August, YouTube was quietly experimenting with machine learning to unblur and sharpen Shorts. It hit already-damaged footage and made it worse — which is why the clips looked rougher on YouTube than on Instagram. So the artifacts were genuine AI fingerprints attached to genuinely real people. Everyone reasoned: artifacts, therefore fake crowd. Airtight-feeling, and wrong. Artifacts tell you a model touched the pixels. They don't tell you a model invented the subject. And notice how it was settled — not by squinting harder. Andy Baio matched the video to the photographers' original Instagram posts. Same crowds, same faces. In the sharp originals, one garbled sign becomes legible: From West Philly to West Swizzy. You don't resolve these by looking. You resolve them by finding the source.
What authoritative sources say
People also ask
So did Will Smith use AI or not?
He used AI — just not to invent an audience. His team ran real tour photographs through image-to-video models to animate them into short clips. The crowd existed; the motion was AI-generated.
How do we know the crowds were real?
Provenance. The video was matched against tour photographers' original Instagram slideshows showing the same audiences and identifiable individuals, including the couple with the cancer-survival sign, who appear in Smith's own Gurtenfestival posts.
Why did the signs look garbled?
Image-to-video models invent motion frame by frame and can't hold fine text stable. In the original sharp photos the sign clearly reads "From West Philly to West Swizzy."
What did YouTube have to do with it?
In August 2025 YouTube was quietly testing machine-learning post-processing on Shorts to unblur and sharpen video. Applied to already-artifacted footage, it made things look worse — the clips appeared rougher on YouTube than on Instagram.
What's the lesson for spotting AI video?
Visual artifacts prove a model touched the pixels, not that the subject was invented. Real footage can be pushed through AI processing and look fabricated. Check the source, not the fingers.