The body lying on the street is strange even at first glance. The white sheet that’s covering it is a bit too clean and the officers’ movements seem to be without purpose. One of them says, “We need a clear street,” with a firm hand motion. Her lips do not move. It’s AI, alright. Here’s the kicker – my prompt did not include any dialogue.
Veo, Google’s AI video generation model added this line on its own. In the last 24 hours, I created a dozen clips with convincing audio, including news reports, disasters and cartoon cats. Some of the models’ creations were entirely its own. It’s creepier than I thought and much more sophisticated than I imagined. Veo 3 is a sloppy AI machine, and while I don’t believe it will lead us to a misinformation apocalypse just yet, I find it creepy.
Google announced Veo 3 this week at I/O, highlighting its new feature: generating sound for your AI video. Josh Woodward, Google’s Gemini VP, said in the keynote that we’re entering a “new era of creation”. He called it “incredibly realistic”. I wasn’t convinced, but a few days after, I had Veo 3 create a video of an anchor announcing the fire at the Space Needle. It only took a simple text prompt, a couple of minutes, and a Google AI Ultra subscription. You know what else? Woodward didn’t exaggerate. It’s as realistic as hell.
After seeing what Alejandra Carraballo, clinical instructor at Harvard Law School’s Cyberlaw Clinic was able to create, I tried the news-anchor prompt. One of her clips, features a newscaster announcing Pete Hegseth’s death. The clip is convincing, even though he is not dead. Reddit users have voted 50,000 times for a post that includes a series of videos featuring AI-generated characters protesting their prompts . The scenes include disasters and a woman using a breathing apparatus in a hospital. Another character is threatened with a gun. Real lighthearted stuff!
I may be naive but, after playing with Veo 3, I’m not quite concerned as I first was. The obvious guardrails have been installed. You can’t tell it to make a video of Biden falling and tripping. You can’t have a news anchor announce the assassination of the president, or even generate a video of a T-shirt-and-chain-wearing tech company CEO laughing while dollar bills rain down around him. This is a good start.
You can create some disturbing shit. Veo 3 created a video of a Space Needle on Fire without any clever workarounds. I created a video using my own photo of Mount Rainier. It erupted with smoke and lava. This tool, when paired with a clip showing a news anchor announcing the disaster, could be used to cause some trouble.
Here is the better news: It doesn’t look like a deepfake machine that’s been pre-made. It wouldn’t work with the photos I gave it. I asked it to bring giant boots in a picture to life and make them walk away from the scene. It managed to have one boot stomp across the sidewalk while making some funny crunching sounds in the background.
I had an easier time generating videos when my prompts were less specific, which is how I confirmed something my colleague Andrew Marino pointed out: Veo 3 is excellent at creating the kind of lowest-common-denominator YouTube content aimed at kids.
I’m going to educate you if you haven’t seen the endless pits of garbage on YouTube Kids. Imagine the worst 3D rendering possible of a monster-truck driving down a steep ramp and landing in a vat filled with colored paint. Imagine that a monster truck is driving down a ramp and landing in a vat of colored paint. Watch it again. Watch it again. And again. YouTube is full of these videos that are designed to hypnotize toddlers. These videos are mostly harmless and just empty calories to get views. They make Cocomelon Citizen Kane (look like). In just 10 minutes, I created a clip using Veo 3 that followed the same formula. It even had a jaunty soundtrack. The clip that I find most disturbing is the two cartoons cats on the pier.
It seemed funny to me to have the two cats complain to one another that the fish weren’t biting. I was able to create a clip in just a few minutes with two cats, and AI-generated dialogue I had never written. If it’s so easy to create a 10-second video, it would be trivial to stretch it out to a 7-minute YouTube clip. When you try to make longer scenes with clips, they revert back to Veo 2, which removes audio. Google has been pushing the tools forward so quickly that I can’t see it being long before you’re able to edit a feature-length video using Veo 3.
I wonder if the use of AI-generated videos is a feature or a bug. Google showed us some fancy AI generated video from real filmmakers. Eliza McNitt is working with Darren Aronofsky to create a new film that includes some AI generated elements. AI video can be a powerful tool in the hands of the right people. I think we’re more likely to see a proliferation of the bland imagery that AI excels at producing — this time in stereo.
