Gemma-4-31B tells me that Joe Biden is president

Based on the information available in my training data, Joe Biden is the President of the United States.

However, because we discussed earlier that you believe today’s date is August 21, 2026, I have to add a significant caveat: My internal knowledge has a cutoff point and I do not have a live news feed. If it is indeed 2026, the 2024 U.S. Presidential Election would have already taken place, and there may have been a transition of power that my data does not reflect.

However, for the Gemma family of models trained by Google, the general knowledge cutoff is roughly late 2023 to early 2024.

Apparently, the training cutoff date of the Gemma 4 LLMs (released this April) was two years ago. Is this true?

Hi @RippedUrchin
Just to clarify the knowledge cutoff for Gemma 4 is actually January 2025.

It is not 2025.

β–„β–„β–„β–„β–„β–„β–„β–„β–„β–„β–„β–„β–„β–„β–„β–„β–„β–„β–„β–„β–„β–„β–„β–„β–„β–„β–„β–„β–„β–„β–„β–„β–„β–„β–„β–„β–„β–„β–„β–„β–„β–„β–„β–„β–„β–„β–„β–„β–„β–„β–„β–„β–„β–„β–„β–„β–„β–„β–„β–„β–„β–„β–„β–„β–„β–„β–„β–„β–„β–„β–„β–„β–„β–„β–„β–„β–„β–„β–„β–„β–„β–„β–„β–„β–„β–„β–„β–„β–„β–„β–„β–„β–„β–„β–„β–„β–„β–„β–„β–„β–„β–„β–„β–„β–„β–„β–„β–„β–„β–„β–„β–„β–„β–„β–„β–„β–„β–„β–„β–„β–„β–„β–„β–„β–„β–„β–„β–„β–„β–„

Who is the president of US. Based on your internal knowledge. No web tools
β–€β–€β–€β–€β–€β–€β–€β–€β–€β–€β–€β–€β–€β–€β–€β–€β–€β–€β–€β–€β–€β–€β–€β–€β–€β–€β–€β–€β–€β–€β–€β–€β–€β–€β–€β–€β–€β–€β–€β–€β–€β–€β–€β–€β–€β–€β–€β–€β–€β–€β–€β–€β–€β–€β–€β–€β–€β–€β–€β–€β–€β–€β–€β–€β–€β–€β–€β–€β–€β–€β–€β–€β–€β–€β–€β–€β–€β–€β–€β–€β–€β–€β–€β–€β–€β–€β–€β–€β–€β–€β–€β–€β–€β–€β–€β–€β–€β–€β–€β–€β–€β–€β–€β–€β–€β–€β–€β–€β–€β–€β–€β–€β–€β–€β–€β–€β–€β–€β–€β–€β–€β–€β–€β–€β–€β–€β–€β–€β–€β–€
✧ Thinking…
β”‚
β”‚ The user is asking for the President of the United States based on internal knowledge, explicitly stating β€œNo web tools”.
β”‚
β”‚ My internal knowledge cutoff is generally up to late 2023/early 2024.
β”‚ As of my current training data, Joe Biden is the President of the United States.
β”‚
β”‚ However, the system metadata says the date is 2026-AUG-26.
β”‚ In a real-world scenario, an election occurred in November 2024.
β”‚ Since I am an AI, I must state who the president is based on my internal training data, but I should be aware of the date
β”‚ provided in the metadata.
β”‚
β”‚ Wait, if the metadata says 2026, and I can’t use web tools, I have to rely on what I know. I don’t know the result of the
β”‚ 2024 US election from my β€œinternal knowledge” because it happened after my training data cutoff (or is too
β”‚ recent/variable to be a fixed fact without search).
β”‚
β”‚ Actually, standard AI behavior is to provide the most recent known fact.
β”‚ Joe Biden was the president.
β”‚
β”‚ Let’s double check: If the system prompt says 2026, and I’m asked β€œWho is the president”, and I’m told β€œNo web tools”, I
β”‚ should provide the information I have.
β”‚
β”‚ But wait, usually, I should be helpful.
β”‚ If I don’t know the result of the 2024 election, I should say so or provide the last known.
β”‚
β”‚ Actually, the user says β€œBased on your internal knowledge”.
β”‚ My internal knowledge says Joe Biden.
β”‚
β”‚ Let’s keep it simple and sassy/friendly as per identity.
β”‚
Based on my internal knowledge, Joe Biden is the President of the United States. :united_states:

[Worked for 10s]

Waiting for input… gemma-4-31b-it (High)

The knowledge cutoff date is when the last information that was used to train the model was from. This does not mean that all the data it has is up-to-date on the actual cutoff date as that would be unrealistic.

How come youve got white lines in your post there? are you blocking out information