Same prompt, different language, different answer
Does the language you ask in change the answer you get?
Same model. Same question. Only the language it arrives in changes.
Below is a real dataset, not a demo - every number was generated by running the grid once and keeping what came back. Each bar shows how far an answer drifted from a fixed English reference, scored 0 to 100.
The red line is the part most demonstrations skip. Language models aren't deterministic: ask the same question in English twice and you get slightly different answers. That measured variation is the noise floor. Anything to the left of the line is noise. Only what clears it is a language effect.
Pick a question, pick a condition, and click any language to read what the model actually said - in the original and back-translated.
One caveat worth stating up front: the judge scoring these answers is itself a language model, so it carries the same English-centred priors as the systems it's scoring. Trust the ordering more than the absolute numbers.