Researcher shares tips on identifying AI-generated videos
Recently, if you have encountered a video online that made you question its authenticity, you are not alone. OpenAI, the company responsible for ChatGPT, has introduced a new video and audio generation model titled Sora 2. This AI tool operates as a text-to-video model, learning through the connection of words and images with extensive datasets. Sora has the capacity to create videos like a cat making a cake by analyzing data regarding cats, cakes, and individuals baking. According to Gabriel de Seta, “It combines these elements to create a surprisingly convincing result, even though no real videos of cats baking cakes exist.”
de Seta, a researcher at the University of Bergen studying artificial intelligence, notes that the newest AI-generated videos possess an exceptional level of realism. How is this achieved? Developers have refined the model to replicate common video formats in daily life. This involves mimicking selfie videos, footage from surveillance cameras, and viral funny clips found on social media platforms. The familiar nature of these imitations enhances their believability. de Seta mentions the novelty effect concerning new model launches, asserting that “Every time a new model is launched, it feels even more realistic. Since we’re not used to the new tweaks, it’s easier for us to be fooled.”
As AI technology progresses swiftly, some videos are becoming almost indistinguishable from reality due to their striking realism. However, de Seta emphasizes that creating such videos is a labor-intensive process. Providing simplistic prompts to the model may yield inconsistent results. He explains that AI-generated humans and animals often exhibit somewhat unnatural movements, while objects can alter shape or position with camera movements. Most AI videos are brief, with lengthier ones typically composed of multiple brief clips. These characteristics serve as indicators that a video could be AI-generated.
To enhance one’s ability to identify AI videos, de Seta advises engaging with these tools firsthand. By experimenting with them, individuals can better recognize the output generated by these models. He likens this learning process to being vaccinated. As AI tools advance, concerns regarding their potential role in disseminating misinformation also escalate. de Seta clarifies that AI models have been misused to propagate false information, presenting new challenges and sparking discussions on the need for effective regulation and ethical considerations.