When Google unveiled its new conversational AI system Gemini earlier this week, the demo video seemed almost too good to be true. In the polished 6-minute clip, Gemini effortlessly described a series of drawings, differentiated between objects, and even cracked a few jokes while interacting with a friendly narrator. The system appeared smart, quick-witted, and uncannily human. But as more details emerge, it seems Gemini’s magic touch may have been the result of some smoke and mirrors on Google’s part rather than true AI capabilities.

Upon release of the demo, Google briefly noted that “latency has been reduced and Gemini outputs have been shortened for brevity.” But they failed to mention just how edited the demo was. As reported by Bloomberg and The Information, the entire video was pre-recorded and prompts were fed to Gemini ahead of time rather than in real-time conversation. Google eventually confirmed that the video was “an illustrative depiction” based on Gemini’s capabilities, stopping short of admitting it was doctored.

Many AI experts cried foul, saying the demo was misleading about the current limitations of conversational systems. Tech analyst Benedict Evans argued that Google exaggerated Gemini’s stability and speed to gain an upper hand on Microsoft, who recently showcased how their Bing search engine could chat with ChatGPT in impressive (though not perfect) live demos.

Intentional or not, the discrepancy reminds many of Google’s last AI, at which engineers from the company admitted to rushing an unreliable chatbot demo just to keep up with hype from Microsoft and OpenAI. With Gemini, Google made similar bold claims about outperforming competitor models like GPT-4 in benchmarks, but skeptics now wonder if those tests factored in Gemini’s apparent sluggishness.

As AI grows more advanced, demos will undoubtedly remain an important part of capturing public imagination and funding. But edited videos risk damaging public trust. Tech companies may be feeling the pressure in an AI arms race, but taking shortcuts prevents users and developers from truly assessing progress. For all its slick editing tricks, the Gemini launch leaves the uncomfortable question: is Google’s AI as far ahead as they claim, or are there still more smoke and mirrors?