I’ve written about Sora since its very first release back in February 2024. The physics glitches, the copyright minefields, the deepfake risks. I’ve been critical. But when OpenAI officially pulled the plug yesterday, shutting down the app, the API, and video generation in ChatGPT entirely, my first reaction wasn’t really relief.
I get it. I’ve spent real time in the trenches of AI’s impact on creative work. My work at Pex focuses around tech that identifies content, including AI-generated music. I’ve written about the pending art crisis, the music authenticity problem, and the legal battles between labels and AI companies. I’m not a tech bro cheering for the machines. But I don’t see much of a reason to celebrate yet.
Sora didn’t die because it was too powerful. It died because it wasn’t profitable. Downloads dropped 32% month-over-month. Total consumer revenue across its entire lifespan was $1.4 million, an unbelievable fraction of its heft compute costs. With its closure, OpenAI is pivoting toward agents and profitability ahead of its IPO, and a video tool hemorrhaging costs with declining users didn’t survive the spreadsheet.
The technology didn’t go anywhere. The demand didn’t go anywhere. But critically: the people who want to generate video with AI won’t see the Sora shutdown and think, “Well, guess I’ll pick up a camera.” They will open a different tab.
The models you’ve never heard of are the ones you should worry about
Here’s what the AI video generation market looks like without Sora in it: a $1.04 billion industry projected to reach somewhere between $2 billion and $82 billion by the mid-2030s (depending on which analyst you believe), with the biggest players operating under far less scrutiny than OpenAI ever faced.
Kuaishou’s Kling 3.0 currently sits at the top of text-to-video benchmarks. It generates clips up to 60 seconds in native 4K at 60fps. ByteDance built Seedance 2.0, a multimodal model that accepts text, image, audio, and video inputs simultaneously. ByteDance paused its global rollout in March 2026 over copyright and deepfake concerns, but the model is still available domestically in China through Douyin and other platforms. Tencent’s HunyuanVideo is open-source with over 13 billion parameters, the largest open-source video model ever released.
These companies face different regulatory environments, different public pressure, and different incentive structures than OpenAI did. China’s Cyberspace Administration has implemented mandatory labeling for AI-generated content on domestic platforms (Kuaishou removed over 37,000 noncompliant videos and penalized 3,400 accounts in one enforcement sweep). That’s something, but the enforcement applies to Chinese domestic platforms. The global-facing versions of these tools have a much more blurry regulatory picture.
Sora, for all its problems, operated under a microscope. Every controversy made international news. The artist beta tester revolt was covered by The Verge, Ars Technica, Fortune, Deadline. The Robin Williams and Martin Luther King Jr. deepfakes generated estate attention. OpenAI was forced to implement content filters, usage restrictions, and community guidelines specifically because the world was watching.
Kling and Seedance and HunyuanVideo don’t get that scrutiny. When was the last time you read a major English-language investigation into Kuaishou’s content moderation policies? The public pressure that shaped (and ultimately killed) Sora isn’t being redirected toward its replacements.
Open source models have zero incentive to watermark
Sora embedded both visible and invisible watermarks. It included C2PA metadata, the Coalition for Content Provenance and Authenticity standard that functions as a cryptographic receipt proving content origin and the tools used to create it. When you generated a video with Sora, there were technical mechanisms in place to trace it back to the source. Imperfect mechanisms (visible watermarks can be cropped, C2PA metadata can be stripped), but mechanisms.
Now look at what’s replacing it.
LTX-2, released by Lightricks under an Apache 2.0 license, generates 4K video at 50fps with up to 20-second clips. It explicitly advertises no watermark on outputs. MAGI-1 is a fully open-source model supporting infinite video extension for long-form content. Wan2.1 runs on consumer GPUs with 8GB of VRAM. None of these models implement C2PA content credentials. None have a financial incentive to.
If anything, the incentive runs the opposite direction. Watermarks reduce the commercial appeal of generated content. Open-source models compete on output quality and usability. LTX-2 advertises it like a selling point because it is one.
Detection needs to get better, and it needs funding
Here’s the current state of AI video detection. In controlled lab conditions, leading detectors report 92-98% accuracy. Under real-world conditions (compressed video, low resolution, novel generators), that accuracy drops 45-50%. The SynthForensics benchmark (published in February) tested state-of-the-art detectors across five different video generators and documented a mean performance drop of 29.19% AUC. Some detectors performed worse than flipping a coin. Under heavy compression, the kind you get from every social media platform, top models lost over 30 percentage points in accuracy.
At Pex, I see the gap between what detection can do and what it needs to do. Our AI Song Detector works (quite well), but it requires continuous investment, continuous dataset curation, and continuous updates because the generation models don’t stand still. Our v4 dropped in February 2026. V3 came out a month before that. The cycle never stops.
Video detection is (understandably) harder. Video has structural complexity that makes identification more challenging: temporal dynamics, visual artifacts that get destroyed by compression, generation techniques that vary wildly between models. The SynthForensics authors built their benchmark specifically because existing detection tools were failing against purely synthetic video from text-to-video models, as they were trained more for deepfake protections than full generations.
When the most visible AI video product shuts down, the urgency to fund detection research drops. Why fund a detector for a dead product? But only the product is dead; the technology is everywhere. The demand for detection has never been higher: political deepfakes are running in midterm ads, open-source models are proliferating without watermarks, and the gap between generation quality and detection capability is widening.
What we should do instead of celebrating
The instinct to cheer when a controversial AI tool goes away is human and understandable. Sora caused real harm and anxiety for creative workers; I don’t blame anyone for feeling relieved, but the measured response to Sora’s death should be the opposite of relaxation.
Keep pressure on every company, not just the visible ones. OpenAI was an easy target because it was the loudest, but Kuaishou, ByteDance, Tencent, Lightricks, and the dozens of smaller open-source projects building video generation tools need the same level of scrutiny. The Concept Art Association’s push for AI Copyright Transparency legislation is exactly the right approach, but it needs to apply broadly, not just to the companies Americans have heard of.
Fund detection research. The SynthForensics paper makes it clear: current detectors aren’t keeping up. Government funding, platform investment, and academic research need to scale with the generation capabilities they’re trying to match. Detection accuracy that drops by 29% against new generators and 30+ points under social media compression isn’t good enough.
Push for C2PA adoption as a baseline. The EU AI Act mandates it. The California AI Transparency Act supports it. But mandates without enforcement are useless. The C2PA standard works when everyone implements it. When the biggest video generation tool shuts down and its replacements don’t bother, the standard weakens.
Watch the open-source ecosystem. This isn’t about opposing open-source development (if anything, we need to support it far more). Open-source AI has enormous value for research, accessibility, and innovation. But the same openness that makes these models valuable also makes content provenance harder to enforce. Facebook’s VideoSeal project, which embeds invisible watermarks that survive common transformations, is the kind of open-source contribution that helps move the needle. More of that. Less “no watermark” as a competitive advantage.
The AI video generation market is growing, fast, and spreading across more tools, more models, and more jurisdictions. The company that had the most public accountability is gone now, which is more of a warning than a victory.
Keep your shoes on. The ground is shifting.