Change comes at you fast.
In late May 2025, Google announced its Veo 3 video generation model at its I/O conference. Early testers were stunned by the photorealistic video, audio, and dialogue that the model produced.
Just a month later, a single director used Veo 3 to create a hyper-realistic AI-generated ad for betting platform Kalshi. The ad took three days to create and cost $400 in Veo 3 credits. It aired during the NBA Finals beside ads that took months to create and cost hundreds of thousands of dollars.
Veo 3 isn’t the first AI video generation platform. But its power has given rise to a sobering fact: in the right (or wrong) hands, generative AI tools can now produce images and videos that are indistinguishable from reality.
These tools are cheap and are usable by anyone who can type a prompt into a chat window. That may lead to an unprecedented wave of AI-generated content across social media platforms and digital channels that is so realistic it’s hard to tell if it’s real or not. In fact, AI slop already has a strong presence on platforms like YouTube, Pinterest, and Instagram, as John Oliver explained on a recent episode of Last Week Tonight.
This begs the question: What, if anything, can be done about this?
A Big (Synthetic) Problem
How big of a problem is the proliferation of hyper-realistic synthetic images and video generated by AI? According to experts, it’s enormous.
“It is now almost impossible to tell, in the digital world, what is real and what is artificial,” said Paolo Giudici, a professor of statistics at Italy’s University of Pavia.
AI image and video models can now produce content for distribution on social media that most users “would not question as fake,” said Mike Perkins, a researcher at the British University Vietnam who has done work on synthetic content.
In a sense, Pandora’s box has been opened. The AI tools that generate hyper-realistic synthetic content are not going away. In fact, as Veo 3 proves, they’re getting better, faster.
So, the first, and biggest, effort to address what is digitally real and what is not starts at the tool level. And the primary way that is being tackled right now is through watermarking.
Some AI labs are using, or participating in, digital watermarking efforts to indicate that an image is AI-generated by adding data to the image file itself the moment it is generated by an AI tool. The image essentially carries a digital “badge” that indicates it is AI-generated.
One of the top lab-run watermarking initiatives is SynthID from Google. SynthID embeds into content machine-readable watermarks that can’t be seen by the human eye. It is now automatically added to all content produced by Google’s generative AI models, including Veo. Google is reported to have already watermarked more than 10 billion pieces of content with SynthID.
Other labs, like OpenAI and Microsoft, participate in C2PA, the Coalition for Content Provenance and Authority, an open standard for watermarking AI-generated content. The C2PA initiative seeks to create a standardized way to track the origin of digital content. It allows cryptographically signed metadata to be attached to digital assets identifying the tools that created it.
Watermarking shows plenty of promise, said Giudici. Efforts like SynthID and C2PA are becoming more sophisticated and cost-effective.
But there’s still an obvious problem.
Watermarking works, but it requires universal, consistent application at scale to fully address the problem. And we are nowhere near that.
Policing Falls Short
To fill the gaps in watermarking’s coverage, some social media platforms are also taking steps to combat AI-generated content.
Instagram automatically attaches a “Made with AI” tag to AI-generated content flagged by its systems or by the user uploading it. YouTube requires creators to disclose when a video they upload contains AI-generated content. TikTok requires users to label AI-generated content.
But a familiar problem quickly rears its ugly head: for platform-led policing to be effective, every platform needs to do it consistently. Based on the flood of AI-generated content already available across these platforms, this is something that is decidedly not happening today.
That leaves a burden on users to address the issue. The problem, said Perkins, is that many users are unable to identify AI-generated content.
“I believe we are now at the tipping point where the signs of fakes are becoming so small that we need to rely on critical thinking of viewers rather than any visible artifacts or problems with the content,” he said. That raises the importance of verifying the source of consumed information, a skill in short supply if the proliferation of online misinformation is an indicator.
The best bet, Perkins said, may be more education at the user level. When social media users know how good AI-generated content has become, they can be better prepared to handle their consumption.
“Having some form of AI literacy is important so that users realize what is possible, and then know what to watch out for,” he said.
Then there’s the final, perhaps largest, elephant in the room.
Watermarking is inconsistently applied. Platform- and user-led policing is inconsistently applied. But even if both somehow worked perfectly, that wouldn’t eliminate hyper-realistic AI-generated content that’s designed to mislead, due in part to open-source AI.
Open-source AI allow users to run models that generate images and video locally on a robust machine without restrictions or guardrails imposed by the model creator.
That means, said Perkins, that malicious actors can find open-source models that don’t have watermarking and use them to generate non-watermarked content. Even if an open-source model creator adds watermarking later, bad actors can simply run a version of the model—or another one—without any safeguards.
While Perkins recommends that platforms implement more robust policies and technical checks, and that users become more AI-aware and literate, at the end of the day open-source models challenge a perfect solution to distinguishing what’s real and what’s not.
And that’s a future for which users are not ready.
“It’s my opinion,” said Perkins, “that the current methods for detection of AI-generated image and video content is entirely insufficient to prevent the average user from being fooled by the new generation of text-to-video diffusion models.”
Logan Kugler is a freelance technology writer based in Tampa, FL. He is a regular contributor to Communications and has written for nearly 100 major publications.
