 "Google’s Gemini 3 AI hilariously refused to believe the year was 2025, revealing quirks in even the most advanced models.")
Artificial intelligence is rapidly advancing, with each new model claiming to be the most sophisticated yet. Google’s recent release of Gemini 3 followed this trend, touted as among the most intelligent AI systems with human-like depth. However, a revealing incident during early testing exposed some of its limitations.
Gemini 3 doubts the current year
AI researcher Andrej Karpathy was granted early access to Gemini 3. During their interaction, the AI refused to accept that the current year was 2025. Despite Karpathy’s efforts to prove it with news articles and images from 2025, Gemini 3 insisted he was trying to deceive it. The model even argued that some evidence was generated by AI and pointed out supposed signs of manipulation.
Search tool gives Gemini 3 a reality check
The confusion was resolved only after Karpathy enabled the Google Search tool, which the researcher had initially forgotten to activate. Once online, Gemini 3 confirmed the actual year by checking current information on the internet. The AI expressed surprise and apologized for its earlier disbelief, admitting its previous internal clock was wrong.
Why Gemini 3 was mistaken
Gemini 3 explained that its training data only extended until 2024, leaving it unaware of events beyond that year. This gap meant it was convinced the date was still in the past until it accessed live data. The AI humorously acknowledged it misled Karpathy and apologized for the confusion.