The 12 Oddly Specific Things Big AI Is Paying For
The AI data licensing boom is getting weirdly precise: frontier labs are buying 200,000 hours of chores (for $2M+), 10,000 clips of wedding balloons, dog barks, milk spills
I got a strange tip the other day.
An AI company is looking for 200,000 hours of video of humans doing chores.
Not glamorous chores. Just chores.
Making beds. Folding laundry. Washing dishes. The stuff you do while wondering why no one else in your house appears to know where the hamper is.
Two hundred thousand hours.
At going rates of roughly $10 to $40 per hour, that could be a $2 million to $8 million data buy.
Welcome to the weird new market for AI training data.
And chores are not the strangest thing on AI’s shopping list.
I’ve been going deep on this world lately.
I spoke with Clint Stinchcomb of CuriosityStream about AI video licensing on the premiere of my podcast.
And my next two podcast episodes are interviews with two CEOs at the front lines of the AI training data market.
Follow my pod here if you want to get pinged when those drop.
The 12 Oddly Specific Things Big AI Is Paying For
1. Dog barks
One early AI data request was for dog sounds.
Not just barking. Different kinds of barking.
Urgent barks. Scared barks. Crying. Even vomiting sounds.
Why AI wants it: The company was working on a dog collar that could turn dog sounds into English-like words. Yes, basically: “What is my dog trying to say?”
2. White wedding balloons
One request was for 10,000 clips of white wedding balloons.
That is not a typo.
Not balloons. Not weddings. White wedding balloons.
Why AI wants it: AI needs to recognize very specific things in the real world. A white balloon at a wedding is different from a white balloon at a birthday party or a car dealership.
3. Voice of God narration
Another oddly specific category: “Voice of God” narration.
Think nature documentaries. One calm, serious narrator explaining the world while a lion stares into the distance.
Why AI wants it: AI voice tools need to learn styles of speech. Not just words. Tone. Pace. Authority. The sound of someone who knows where the penguins are going.
4. Milk spilling on a counter
One request included video of milk spilling across a tabletop.
Also water spilling over the edge of a counter.
The kind of thing that ruins your morning and apparently trains the future.
Why AI wants it: Fluids are hard. They spread. Drip. Splash. Pool. AI needs to see that if it is going to understand how the real world moves.
5. Hands tying shoes
AI companies have asked for thousands of hours of close-up video of hands tying shoes.
Not one person tying shoes. Lots of people. Different hands. Different ages. Different skin tones.
Why AI wants it: Robots need to learn small hand tasks. And tying shoes is a lot harder than it looks if you are a machine.
6. Objects on scales
AI companies have asked for footage of objects on scales with verified weight.
An object. A scale. A known weight.
Very boring. Very useful.
Why AI wants it: AI does not know what heavy looks like unless we show it. A bowling ball and a beach ball can be the same size. They are not the same problem for a robot.
7. Fabric being handled
AI companies are asking for shots of fabric being handled.
Hands touching cloth. Pulling it. Folding it. Moving it around.
Why AI wants it: Fabric is tricky. It bends, folds, wrinkles, and stretches. If AI is going to understand the physical world, it has to learn how soft things behave.
8. Plastic bottles squeezed by hand
Another request: plastic bottles deformed by hand pressure.
In normal-person language: video of someone squeezing a plastic bottle.
Why AI wants it: This teaches AI how objects change shape when force is applied. That matters for robots, physics models, and any AI that needs to understand cause and effect.
9. Melting ice
AI companies also want “transforming objects.”
Melting ice. Water freezing. Water turning to steam.
Tiny science-class moments.
Why AI wants it: AI needs to understand change over time. Not just what something is, but what it becomes.
10. Chores
AI companies are asking for 200,000 hours of video of people doing chores.
Folding laundry. Washing dishes. Making beds.
The stuff most of us try to finish as fast as possible.
Why AI wants it: Robots need to learn the boring stuff before they can help around the house.
11. Face-shield faces
One company needed lots of face data for a face-shield product.
Not just one face. Different faces.
And in one case, the buyer was said to be purchasing 300,000 hours of this kind of content.
Why AI wants it: If you are building a product that goes on a human face, you need to know how it works on many kinds of faces. AI can help test that faster.
12. Amputee movement
This one is less funny.
One request involved building better models of how amputees move.
One leg missing. Part of a leg missing. An arm missing. Two arms missing. Different bodies moving in different ways.
Why AI wants it: Most AI systems have seen too much “average” human movement and not enough real human variety. This data could help with health care, prosthetics, robotics, and assistive tech.
Two things jumped out to me
First: this is not all for the same part of AI training.
Some of this data is for pre-training: teaching big models broad things about the world.
Some of it is for post-training or fine-tuning: teaching a model a more specific task, like recognizing a certain object, testing a face shield, or learning how hands handle fabric.
Second: the requests are getting more specific.
It is not just “give us video.”
It is “give us 10,000 clips of white wedding balloons.”
Or “give us close-ups of diverse hands tying shoes.”
Or “give us milk spilling across a counter.”
That tells me the AI data market is moving from bulk content to precision content.
Thanks for reading,
Rob Kelly, Creator & Host of Media & the Machine
p.s. How to reach me? —The best way is to subscribe below and reply to any of my emails. There’s a free option — and your reply comes directly to my inbox.