Cohere details cultural riddle tests that drew confident, wrong AI answers

An AI answer can sound convincing while missing the cultural clue needed to get it right. In an October 22 account, Cohere described a community research project that used culturally specific riddles to probe that gap. Contributors reported confident wrong answers and, in tests involving Buginese, virtually no knowledge of the language.

A puzzle with a cultural answer key

The project began with Hungarian diplomat Boczka Károly, who wrote Hungarian trivia-style riddles to test whether models recognized cultural references. He found answers that sounded plausible but missed the relevant facts; some contained invented information. After he shared those observations in a Cohere Labs community discussion, staff research scientist Julia Kreutzer worked with him to take the effort international.

The community built benchmarks—sets of questions used to evaluate models—across subjects including sports, history and politics. Each question took the form of a riddle. The exercise examined fluency, logic and reasoning in models that Cohere described as trained mostly on English data, with attention to the cultural knowledge behind an answer.

Riddles gave the work a specific interpretive task: ambiguous wording could still have a particular intended solution. Cohere illustrated the idea with a question about a house one enters blind and leaves seeing, whose answer is a school. Its account says models can struggle with that same riddle when it is asked in another language.

Cultural knowledge was the entry requirement

Recruitment ran through social media and existing Cohere Labs research communities, without formal entry criteria. Participants needed to identify with a culture and commit to writing and evaluating riddles, rather than bring technical AI expertise. They gathered on Discord. According to Boczka, the effort attracted more than 100 contributors from more than 50 countries.

Among them were Henry Gagnier, a 15-year-old American high school student, and Jane Smith, a retired Canadian teacher. Gagnier described cultural knowledge, not AI skills, as the useful qualification for constructing the benchmark. He said the group welcomed him despite his initial hesitation about joining as a high school student and seeking a team leadership role.

Smith brought a classroom habit to evaluation: working backward from a wrong answer to ask what might have led to it. She described using the same approach when assessing children’s work. Her contribution focused on understanding the route to an error, rather than simply recording that the final answer was wrong.

The scary part to me is that it sounds so authoritative and people are looking for authoritative answers.

Jane Smith, project contributor and retired teacher, in Cohere’s account

Language gaps complicate the cultural test

Gagnier’s observations about Buginese pointed to a more basic obstacle than interpreting a cultural reference. He said the tested models showed virtually no knowledge of the language, which also translated into no cultural knowledge under the team’s evaluation. His description concerns those tests, not a measured result for every AI model or every Buginese task.

Smith identified a separate problem in how the models presented their answers. She was surprised by how rarely they acknowledged not knowing, even when their responses were wrong. That observation puts the concern on the answer’s apparent authority as well as its accuracy: a confident tone could make an incorrect response seem trustworthy.

Cohere’s account offers qualitative observations rather than a model-by-model scorecard. It does not identify the tested models, give a question count or publish accuracy rates. The findings therefore illustrate failures contributors encountered, without establishing how often those failures occur or which systems handle culturally grounded questions best.