Google caught faking part of Gemini AI demo

Google has admitted it staged a video during the demo of its new Gemini generative AI model. Gemini is Google’s direct response to OpenAI’s GPT-4 model, which has taken the internet by storm since it was introduced last November. Up until the launch of GPT-4, Google had generally been considered to be the leader of AI software but suddenly found itself behind in the generative AI race. 

One year after the launch of GPT-4, Google is back with Gemini, which it claims outperforms GPT-4, but is also the first AI model to outperform human experts on Massive Multitask Language Understanding (MMLU). Gemini “Ultra” apparently scores 90.04% on the MMLU beating out GPT-4, which scores 87.29%. It is one of three new flavors of Gemini which also comes in “Pro” and “Nano” sizes - the latter made small enough for the Android devices like the Pixel 8 Pro.

However, all may not be what it seems to be as far as Google’s first ostensibly impressive demo (embedded below) of Gemini. In the video, Gemini is heard initially interacting with a person asking it what it sees being drawn in real-time as a squiggle morphs into a more complex object. The demo continues through a number of additional scenarios and interactions, each increasingly more impressive. 

When Bloomberg contacted Google to ask how the video was created, Google admitted that the video was edited and that the Gemini voice was not responding to images in real-time, but each was a separate still image. Moreover, the prompting was not made by the person heard speaking in the video using natural voice recognition (as it seems) but that the voice prompts were in fact made by text inputs. According to Google, the video was merely made to “inspire developers.” 

In the disclaimer accompanying the video on YouTube, Google only acknowledges that “For the purposes of the demo, latency has been reduced, and Gemini outputs have been shortened for brevity.” This markedly different from the comment Google provided to Bloomberg which stated the video was created, “using still image frames from the footage, and prompting via text.” With OpenAI’s GPT-5 on the way, it looks like Google is desperate to reclaim the king of AI crown - very desperate. 

We asked Google’s Bard AI what it had to say about Google’s Gemini AI snafu - it's response is embedded below for interest.