AI videos are flooding the internet – here’s what brands need to know before jumping in

Welcome to the era of video AI, where simulated clips dominate our feeds and leave us questioning what’s real and what’s fake.

Leading this new wave is Sora, a generative AI app rapidly emerging as one of the year’s most talked-about and hotly debated tech launches. Developed by OpenAI, the company behind ChatGPT, Sora can transform written prompts into lifelike, AI-generated videos within seconds.

It reached one million downloads in less than five days despite a limited release in North America on iOS only. Its instant popularity drowned out Vibes, a similar short-form AI-generated video feed that Meta introduced to its AI app around the same time. But undoubtedly, both have contributed to the uptick in AI videos appearing online, along with a broader proliferation of text-to-video generators.

The Sora app provides a way for mobile users to access OpenAI’s latest video-generation model, Sora 2, completely free. A more experimental model promising higher quality, Sora 2 Pro, is available via Sora.com for ChatGPT Pro users, who pay more than €200 per month, while those on the cheaper Plus plan can access Sora 1 Turbo. And those more serious about utilising OpenAI’s video-generation services can plug into the Sora 2 API, which costs a minimum of 10 cents per second of footage generated.

Google is now on the third generation of its text-to-video model Veo, and its Flow platform pulls in the capabilities of this model alongside image generation from its Imagen model, as well as the ability to edit and refine with the Nano Banana model (officially called Gemini 2.5 Flash Image, but that’s far less fun to say). This combination of models and capabilities is popular with AI video creators because of Nano Banana’s prowess in retaining consistent characters and objects, something generative video generally struggles with, seeing as each ‘scene’ is a new creation.

Audio is a fairly new addition to these video generators, and not all speech synthesisers are created equal, which is why specialist AI voice generator Eleven Labs is another popular tool in the kit of creators. And when you want to be specific about character movement, you can do a gen-AI version of motion capture with Runway’s Act 2.

Not to be left out, Chinese tech giant ByteDance has developed the Seedance model for text-to-video. ByteDance also owns the hugely popular video-editing app CapCut, which recently introduced Pippit. Targeted at small businesses, Pippit generates video marketing materials for a product based on a URL, image or description.

This is simply just a whistlestop tour along the current AI video high street, scratching the surface of the various tools available. Though it’s sold as a process as simple as text to video, the reality is you will need time, money and knowledge to bridge the gap between those two elements. PJ Ace, arguably one of the top AI video makers out there right now, frequently documents the process in detail for his followers. Ace leads Genre.ai, an advertising agency described as ‘AI native’. Genre’s contributions to video advertising don’t hit much differently than the general tone of US advertising, if a bit glossier around the edges and lacking a certain sense of tangibility, as is the case with most AI-generated content.

An ad the agency made for fintech start-up Kalshi, in particular, garnered plenty of attention – which was, of course, the whole point. Right now, it’s still a valid marketing stunt just to release something AI-generated and watch the headlines roll in. The Kalshi ad also capitalised on the modern audience’s appetite for unhinged brainrot, featuring a series of off-kilter scenarios such as a chicken farmer bathing in a paddling pool filled with eggs and an alien in a basketball vest chugging beer at a house party. This kind of chaotic content may resonate with online audiences, but is it good for your brand?

Brand marketing typically prioritises precision and consistency, but these are not the strongest strengths of generative AI. If you want any more than five to 10 seconds of video with a recurring character, you are likely going to run into some challenges getting that right.

These applications often promise the possibility of endless iterations as an opportunity, but it can also be an in-built liability. With every iteration, you run the risk of a hallucination polluting the output. Also, editing errant text from a chatbot is easy peasy compared to what it takes to tweak and finesse multimedia content. With generated content all you get is an end product – no raw files or production drafts to play around with in the edit.

You can’t bank on the model’s reliability, and you also can’t bank on costings or timeline estimations. Every new prompt and output adds an element of randomness to the equation. There’s no telling how the server load will impact the time it takes to produce an output, and you don’t know how many prompts and iterations it will take to get your final finished product. Generative video systems are like slot machines. You feed in a prompt like a coin, pull the lever and hope for the best, but you won’t get a winner every time. In fact, the odds are very likely stacked against you.

Ace said he generated 300 to 400 clips to get 15 usable shots for his Kalshi ad. That number, on any given day, could run lower or much higher. No one knows how many times you need to pull the lever for the slot machine to pay out, but it is going to cost time and money every time.

Some may be willing to take the chance, but buyer beware. It needn’t be said that business users should shy away from any video creation that will wade into these products’ copyright troubles. The Sora app, in just weeks of existence, has quickly had to rein in its free-for-all on generating real people’s likenesses and copyrighted characters and content. Those legal challenges are far from over for OpenAI and others in this space.

Businesses that previously partnered with creatives should also consider how a move like this might burn bridges with that community, which – despite constant reports of AI heralding its demise – won’t be going anywhere. After all, if these models didn’t have human-made work to train on, they’d be useless. These generators are also energy-intensive so their use will have a significant bearing on your environmental impact.

And there’s reputational damage to be considered here too. Over 90pc of all deepfake content is pornographic, the majority of which is non-consensual. The Sora feed is littered with a juvenile, edgelord sense of humour, and users have taken to creating fetish content that skirts the platform’s no-nudity rule but still provides sexual gratification. These are not things that scream ‘positive brand association’.

If you absolutely need to create a video that involves scenes you simply couldn’t afford to stage, then maybe this is an option for you. But it’s not a magic wand. Creating production-quality gen-AI is a skill in and of itself, and you’re still best placed partnering with someone who knows what they’re doing than thinking of this as a DIY job.