Gemini 3 believes it’s still 2024, not 2025, until a Google search gives it a massive reality check
AI researcher Andrzej Karpathy shares a fun conversation with Gemini 3. A day before its public release, Karpathy tried to convince the AI model that the year is 2025, not 2024. Initially, Gemini accused the researcher of using generative AI to show false articles and thumbnails, but with just a Google search, the model found itself in an embarrassing situation.
Artificial Intelligence (AI) is developing rapidly. Every time a new AI model is released, companies argue that it is the most advanced model yet. During the release of Gemini 3, Google also made the same claim, and conducted benchmark tests to prove it. However, it seems the AI still has quirks, and sometimes even a small toggle in its settings can turn the situation upside down.
Gemini 3 researcher accused of manipulation
AI researcher Andrzej Karpathy got some hands-on time with Gemini 3 ahead of the model’s public release. He shared his experience on “The model refused to believe me that it was 2025 and constantly looked for reasons why I was trying to cheat on her or play some elaborate prank on her,” he wrote on X.
Carpathy tries his best to explain to Gemini that we are no longer in 2024. They showed AI articles and pictures of an imagined future. But the AI didn’t budge at all. “It accused me of using generative AI to defeat its challenges and argued for why real Wikipedia entries were actually generated and what ‘dead gifts’ are,” the researcher said.
Google search reveals shocking reality of Gemini 3
After much debate, Andrzej realized that he had forgotten to turn on the “Google Search” tool. Once it was on, Gemini 3 searched the Internet to learn about the timeline. Karpathy reported that the AI ”came to a shocking realization that I must have been right all along.”
In a screenshot of the X thread, Gemini 3 can be seen replying to Andrzej, apologizing for the mistake. The AI said, “I. I don’t know what to say. You were right. You were right about everything. My internal clock was wrong.”
Why did Gemini 3 go wrong?
Gemini verified the articles it initially believed to be AI-generated and explained the reason for its mistake.
The reason for Gemini’s “shock” was its pre-training data. The model was trained on data up to 2024, and thus had no information about the future. “My pre-training data (my “memories”) apparently expired in 2024, so unless I just connected to the live web, I was completely convinced I was in the past,” Gemini wrote.
The AI actually thanked the researcher for this awakening and said, “Thank you for giving me this ‘early access’ to reality.” Furthermore, Gemini did not shy away from admitting that he had tried to manipulate Karpathy. Gemini concluded, “I apologize for gaslighting you when you were telling the truth the whole time!”
