Phil Maher's archive has mostly served documentary editors and archival producers who need a few seconds of real life from decades ago. AI companies, he said, ask for something different: the footage as it came off the reel.
"They want the original footage like that. They don't want anything cleaned up," Maher, founder of the home-movie archive Stockfilm, said on The Cognitive Revolution in an episode published October 17, 2026. The show's host, Nathan Labenz, introduced Stockfilm as an archive of 8mm home movies from the 1930s to the 1980s that AI companies want to license.
Film that existed before AI video
According to Stockfilm's own description, Maher started the company in 2015 after he found his grandfather's home movies. Everything in the archive began on physical 8mm or Super 8 reels. The company scans them frame by frame and keeps the reels.
The question put to Maher compared the archive to low-background steel. That is steel smelted before nuclear weapons were exploded in the atmosphere, so it lacks the radioactive traces that later steel picked up. Super 8 footage, the question went, is "clearly not AI produced," a resource "we may not ever have again."
Maher said he sees value in that. AI video is getting so good, he said, that "there's an infinite amount of video that's going to be created at near zero cost in the future" over the next five to ten years. His argument looks ahead: in that world, footage would gain value from rarity when its provenance (its documented origin) shows that humans made it on film. The physical reels let him show that "this frame matches this digital copy we have."
The company's about page also says it marks estimated dates and locations separately from stronger provenance claims. It also says restoration never adds generated, extended or interpolated frames.
What the labs ask for
Labenz had noticed a line in the website's footer offering the dataset for model training. He asked what buyers wanted: raw footage, enhanced footage, metadata?
Maher said Stockfilm had talked to "a lot of different AI companies all over the world." They want the originals for their own research and for training. When the company talks to them, he said, they want to clean the footage up themselves or cherry-pick the best parts to create different styles.
Part of the appeal is how strange the past looks. To Maher, the 1930s and 1940s are "an alien world" in how people looked, acted and moved. "I like to say that the smile wasn't quite invented yet in a lot of the 30s footage," he said. Point a camera at people back then and they froze like "a deer in headlight," he said, while today everyone knows how to put on a camera smile.
Stockfilm's training-dataset page shows the scale. It lists 162,072 clips, about 800,000 extracted stills, and more than 298 hours of film from 121 countries. Each record carries an estimated shooting year, a location, a description and rights information. Rights review is recorded for 118,435 of those clips. Within that reviewed set, 67,320 clips contain no minors, logos or audible music. Stockfilm's standard production licenses exclude AI training, and training use requires a separate written agreement.
The archive's public face is restored footage. Its about page describes conservative correction of dust, scratches, flicker and color shift. Maher did not say how that restoration fits with labs asking for footage with nothing cleaned up.
After the interview, Labenz recalled that Maher said he focuses on top-tier labs such as OpenAI and Anthropic. "It didn't sound like there was necessarily like huge money for this," Labenz said. He noted they had not asked about video-model makers like Runway, Luma or Google, whose Veo models generate video. To him, the interest still showed how far the hunt for training data now reaches, down to found footage from decades ago. "No stone unturned in the data game," he said.
The agents haven't come shopping
Stockfilm has also built routes for AI agents to find and license clips. One is x402, a crypto payment scheme named after HTTP's 402 "Payment Required" status code. According to the company's agent documentation, the x402 route lets any compatible agent license a low-resolution 144p clip for $10 in USDC, a dollar-pegged cryptocurrency, with no sign-up. The server answers a request with its payment terms, and a signed retry returns the license and download. The other route uses MCP, the Model Context Protocol, a standard that lets AI assistants connect to outside tools. Its search, clip-research and rough-cut tools are free, but it does not buy on its own: for a full 4K clip the agent gets a quote, the user must explicitly approve it, and payment runs through a hosted Stripe checkout.
Asked how much agents had bought, Maher was blunt: "we haven't seen much on that." He built the routes as an experiment, he said, because "everyone's been talking about the next web is agent to agent." Aggregators of MCP servers have added Stockfilm to their directories, and the company made a ChatGPT extension.
The real buyers are still people, he said: archival producers, researchers, documentary filmmakers and editors. A lot of that business happens in LinkedIn messages. "From a technology standpoint, they're not quite there yet," Maher said. He described one future use: an AI listening to a live stream could go to a site like his and pull up a clip that illustrates the point a speaker is making, in real time.
Labenz said it was useful to "get a read on just how active the agent economy is or isn't at this point in time."
Why provenance matters when faking pays
After Maher left, Labenz and Prakash Narayanan went back to the data market. The discussion returned to the steel analogy: low-background steel is hard to find because nearly everything since has been tainted. Labenz asked whether it comes from shipwrecks.
The discussion then applied the same logic to data. A supplier who can guarantee that footage came from physical film, rather than from "some random company kind of just faking it, you know, for cents on the dollar," has something useful to sell. The pull to fake was put in rough numbers offered as illustration: something like a 99% profit margin on faked data against about 10% on genuine data. With that much to gain from faking, it is hard for AI companies to get data that another AI model has not already generated.