The Sora Leak Nobody Asked Permission For

The Sora Leak Nobody Asked Permission For

Tech News ai-video artist-labor openai sora

I was trying to get into Sora the same week everybody else was, and instead of a login screen I got a wall of text from a bunch of artists who apparently beat OpenAI to the punch.

Quick recap for anyone who missed it: OpenAI spent the first couple weeks of December doing this "12 days of shipmas" thing, dropping something new almost every weekday. Day one was the $200-a-month ChatGPT Pro tier, which is a genuinely wild price for a chat subscription when you say it out loud. A few days later they finally opened up Sora, their text-to-video model, to the public — Plus subscribers get a limited number of generations a month capped at 720p and five-second clips, Pro users get longer and higher-res output and fewer restrictions. Fine, expected, this is how OpenAI does everything now: announce it forever, then ship a metered version of it behind a paywall.

What I didn't expect was that a group of artists who'd been given early access to Sora as testers would post working login credentials to Hugging Face days before the official launch, along with an open letter basically accusing OpenAI of using them as unpaid focus-group labor. Their complaint, as I understood it, wasn't really about the tech. It was that they'd spent months generating footage and giving feedback for free, under an agreement that let OpenAI use their name and work to promote the tool, with no compensation and no real say in how any of it got used. So a handful of them just... gave everyone else the keys. OpenAI shut the access down within hours, obviously.

I don't think this is a simple "artists bad, OpenAI good" or vice versa situation, and I'm annoyed at how fast it got flattened into a hot take on social media. These were people who agreed to the terms, used the access, and still felt burned by the arrangement — that's not nothing, even if you think they should've known what they signed up for. Startups have been running "early access tester" programs as unpaid marketing for years, it's just usually developers instead of painters, and developers don't tend to write open letters about it.

Anyway, I did eventually get Sora working once the servers calmed down (it was clearly getting hammered on launch day, my first four attempts just spun forever and then errored out). And honestly? The actual output was underwhelming for how much noise surrounded it. I fed it a pretty simple prompt about a dog running through a wheat field at golden hour, and the legs did that thing where they sort of blur and multiply for a second, then the wheat behind the dog just isn't there anymore. Five seconds of video, and it still looked like it was compensating for something the model couldn't quite hold onto. Maybe I'm spoiled from watching demo reels that were obviously the best of five hundred attempts. Actual first-try output from a rate-limited free account is a different animal.

Small tangent because I can't help myself: I still think the funniest part of this whole "12 days" bit is that OpenAI is doing a countdown-calendar format that used to be reserved for candy companies and Steam sales, for a product that costs as much as a car payment at the top tier. There's something almost charming about it, in a deeply unserious way, like they know nobody's going to be mad about being marketed to if the marketing is cute enough.

Google put out Gemini 2.0 Flash this week too, and from what I've read the benchmarks are solid, but I haven't had a chance to actually poke at it myself yet, so I'll leave that for another post once I've used it for more than five minutes instead of just skimming a blog announcement like everyone else does and pretending they tested it.

The thing that's sticking with me isn't the leak or the video quality, it's how normal all of this has started to feel. A year and a half ago a leaked AI demo would've been the whole news cycle for a week. Now it's a Tuesday. I don't know if that says something about how fast this stuff moves or how fast we adjust to anything, but either way I caught myself scrolling past it after about ninety seconds, which felt worth writing down before I forgot I'd noticed it at all.