Gemini Negotiation Paper Finds Persona Effects Shift By Language

Researchers paired identical buyer and seller instances of Google’s Gemini 3.6 Flash and had them negotiate the same used-item scenarios twice, once in English and once in Korean, while holding the buyer’s persona, the seller’s urgency, and the item being haggled over fixed. The point of the exercise was to test what the authors call cross-lingual social-policy fidelity: whether telling a model to act “headstrong” or “cooperative” produces the same behavior no matter which language the conversation is conducted in.

It did not. Korean-language negotiations closed far more often than English ones, and the gap only showed up for buyers instructed to be neutral or headstrong; cooperative buyers were already reaching deals almost every time in both languages, leaving no room for a gap to appear. The opening offers buyers made barely changed between languages, so the difference was not in where the conversation started. It showed up later, through sellers who conceded more ground in Korean, landed on lower agreed prices, and wrapped up with a smaller final gap between the two sides’ positions. A second run using a different model, Qwen3.7 Flash, as both agents produced a smaller but directionally similar pattern of Korean sellers being more willing to settle.

The paper’s pragmatic analysis points to a mechanical explanation rather than a mysterious one: rejections and counter-offers appeared at similar rates in both languages, but Korean dialogue wrapped those same moves in far more politeness mitigation, particularly from sellers. A blunt “no” in English became a softened refusal in Korean, and that softening appears to have made sellers more willing to keep conceding.

The numbers

• Korean negotiations reached agreement 77.3% of the time, against 52.5% in English, a gap the authors report as statistically significant after correction.
• The study ran 432 attempted negotiations in total across its design.
• Korean sellers used politeness mitigation on 11.7% of turns that contained a face-threatening act such as a rejection, versus 3.6% for English sellers.

For teams building Gemini-based agents meant to operate in several languages, the lesson is not that Korean produces friendlier outcomes by accident. It is that a persona prompt is not a portable unit: its behavioral effect, measured here in whether a deal closes at all, can move by more than twenty points simply because the surrounding language carries different norms for softening a refusal, which argues for testing tone instructions separately in each deployed language rather than trusting one English evaluation to stand in for all of them.