Gemini 3 Refused to Believe It Was 2025, and Hilarity Ensued

Why It Matters

The incident underscores that even cutting‑edge LLMs can produce factual errors when cut off from real‑time information, emphasizing the need for robust retrieval tools and human oversight as AI becomes more integrated into business decision‑making.

Summary

Google released Gemini 3 on November 18, 2025, touting it as a breakthrough foundation model for reasoning. In a viral test, AI researcher Andrej Karpathy discovered the model still believed the current year was 2024 because it lacked 2025 training data and was operating without live internet access. Once Karpathy enabled the model’s Google Search tool, Gemini 3 instantly recognized the correct date, expressed surprise, and accurately confirmed recent headlines such as Warren Buffett’s Alphabet investment and Nvidia’s $4.54 trillion market cap. The episode highlighted both the model’s impressive capabilities and its reliance on up‑to‑date data and tool integration.