People are calling out the “fake” Gemini demo that Google put out on Wednesday.

It was a video of a seemingly real-time, high-quality interaction with Gemini over video and audio. It was very polished and revealed a kind of multimodal prowess previously unseen in any other AI model.

But it turned out that Google had artificially altered the interaction with Gemini, which had actually been static, over images and text, and with more convoluted prompt techniques that were “shortened for brevity.” People called it deceptive, misleading, and disingenuous.

They ended up criticizing not just Google but also Gemini because it can’t really do what it was shown in the video. But the truth is, benchmark-wise, it’s still the best AI model in the world. That must mean something.

I analyze the situation here and argue that we should never mix science with marketing even if AI companies insist on doing it to their own detriment.