I don’t mean to be alarmist, but I do think it’s time to start assuming everything you see online is fake.
The internet is full of content produced by real people, of course (this article included). But AI-generated media is getting so realistic, that it almost puts you at a disadvantage to presume the content you’re scrolling past on your feeds is legitimate.
Don’t skip this article because you know what AI content looks like—the current stuff your algorithm delivers to your social media feeds is easy to spot if you know what you’re looking for. But even if you can identify AI slop the second it hits your eyeballs, you need to know you’re not ready for the next wave of AI-generated videos. That wave isn’t just on its way—it’s already here.
AI content is already fooling people
Most of us are acutely aware of the “AI video” look: This “tragic” video of a cat parent saving their kitten by throwing it out of a burning airplane is obvious AI slop to most who watch it. You probably know Trump isn’t working this construction site, and you most assuredly can understand this family of cat farmers is, in fact, AI-generated.
But there are the videos that aren’t so obvious, especially to those of us not quite so in tune with AI, or technology in general. You might know this video of babies dancing in a circle is AI, but plenty of the people in the comments didn’t (assuming they aren’t bots, either). You might also be able to tell that this family of pets isn’t really watching a bird investigate a toy alligator, but, again, plenty can’t. And there is no end to the America’s Got Talent videos that feature “realistic” yet impossible visuals—that still capture the hearts of hundreds of thousands, if not millions of people. (I weep.)
But I’m not writing this piece today because I’m concerned about how many of these “believable” AI videos are tricking way too many people into thinking they’re real. I am worried about that, but those worries pale in comparison to my new fears.
So far, most of the AI videos taking over social media feeds rely mainly on their visuals and background sounds to sell their alleged authenticity. You’ll notice none of the characters in any of these videos actually speak. If they do, it’s immediately off-putting, with out of sync lip movements and, typically, robotic voices. It’s been easier for AI creators to put the emphasis on the realism of the people and animals in their videos, and hope you’re wowed enough by a baby dancing with a lion to not think, “this is bullshit, right?”
Even OpenAI’s Sora video model, which shocked me with its quality in February of last year, was working off of its realistic visuals. A video of woman “filming” her reflection through a train window too real for comfort, but Sora wasn’t spitting out fully-rendered conversations. If you see such a scene on your feeds, you probably assume, of course, it’s a real video—or at least one generated by humans.
AI video is about to change completely
Something happened this week that only made me more pessimistic about the future of truth on the internet. During this week’s Google I/O event, Google unveiled Veo 3, its latest AI video model. Like other competitive models out there, Veo 3 can generate highly realistic sequences, which Google showed off throughout the presentation. Sure, not great, but also, nothing really new there.
But Veo 3 isn’t just capable of generating video that might trick your eye into thinking its real: Veo 3 can also generate audio to go alongside the video. That includes sound effects, but also dialogue—lip-synced dialogue.
In order to demonstrate Veo 3’s audio/video capabilities, Google showed off a clip of an old sailor at sea. The video quality is sharp and realistic, and the words the man speaks are synced to his lip movements. Of course, knowing the video is AI, you notice quirks that give away the game (to my eye, this looks like a high quality animation more than a live action shot) but I am quite confident this video would fool a lot of fans of fake AGT videos.
But even this clip wasn’t what inspired my newfound fears—it was the videos that users started making once they got their hands on Veo 3. PetaPixel has a great roundup of some of the “best” Veo 3 videos people have made so far, but I’ll highlight some of the ones that should scare you most.
This clip shows a streamer playing Fortnite. Everything, including the game footage, was generated with Google’s AI:
This Tweet is currently unavailable. It might be loading or has been removed.
This clip shows three concerts that never happened, featuring musicians and crowds that do not exist. The music isn’t good, but that’s not the point. The music, from the vocals to the instrumentals, was generated entirely by the AI, and then synced to lips, drums, guitars, and strings:
What do you think so far?
This Tweet is currently unavailable. It might be loading or has been removed.
But this clip is, without a doubt, the one that should sound the alarm for each and every one of us. Someone generated a fake video of a fake car show, featuring fake interviews with fake attendees. It’s far from perfect, but any AI quirks are totally overshadowed by the surface-level realism here. Not only would the AI’s Got Talent fans buy this, I would buy this, especially if I wasn’t looking out for it:
This Tweet is currently unavailable. It might be loading or has been removed.
It’s the visuals; it’s the dialogue; it’s the crowds; it’s the lighting; it’s the candid laughter at “mistakes;” it’s the sound of the mic being “bumped” into. Congratulations on noticing the dialogue often doesn’t make sense, or that the people in the background defy the laws of physics—you won’t notice it when it hits mid-scroll on TikTok or Instagram.
Even Veo 2, which isn’t as powerful as Veo 3, now offers tools for realism, like the ability to dictate how you want the camera to move. And both models are available in Flow, Google’s AI video editor of sorts. Creators now have the ability to generate highly realistic AI content that feels like it was filmed in-person, and the tech is only getting better.
This Tweet is currently unavailable. It might be loading or has been removed.
Google’s best AI video generator tools cost $250 a month through its new AI Ultra subscription plan. That’s expensive, but not out of reach for plenty of people interested in making AI-generated content. But the $20 per month plan, AI Pro, still comes with Veo 2 and Flow access. The rate limits are lower, but I wouldn’t be shocked to see some realistic slop come out of those limitations, too.
It’s time to be a full-time skeptic
None of this tech is perfect. I’m not here to tell you that everything Veo 3 spits out is indistinguishable from real content, or that the videos are absent any of the usual AI tells. In fact, there’s clearly something up with Veo 3’s training data: As 404 Media reports, the model continuously generates the same weird “dad joke” whenever you ask for a generation of a comedian performing standup.
What I’m saying is, it’s time to turn on your bullshit detectors and keep them active full time. When engaging with videos on the internet—especially short-form algorithmic clips—you might be safer operating under the assumption the content is fake from the jump, and require proof beyond a reasonable doubt that what you’re seeing wasn’t generated with a simple prompt and a $250 budget. That feels extreme, but after what I’ve seen this week, I don’t really see another way to engage with this content going forward.
We’re in scary territory now. Today, it’s demos of musicians and streamers. Tomorrow, it’s a politician saying something they didn’t; a suspect committing the crime they’re accused of; a “reporter” feeding you lies through the “news.”
I hope this is as good as the technology gets. I hope AI companies run out of training data to improve their models, and that governments take some action to regulate this technology. But seeing as the Republicans in the United States passed a bill that included a ban on state-enforced AI regulations for ten years, I’m pretty pessimistic on that latter point.
In all likelihood, this tech is going to get better, with zero guardrails to ensure it advances safely. I’m left wondering how many of those politicians who voted yes on that bill watched an AI-generated video on their phone this week and thought nothing of it.