Keboola Engineering: Kai Behaves Differently in Czech

Kai Has a Different Personality in Czech vs English

24 Jul 2026 · 4 min read

Key Takeaways

Anthropic recently published research showing that Claude behaves differently depending on the language you use. Not just translating differently. Actually expressing different values. In Arabic, Claude is warmer and more deferential. In English, it's more rigorous and cautious. In Russian, it challenges your assumptions more.

When we read that, we had one question: does Kai do this too?

So Jordan, our AI lead, analyzed around 3,000 internal Kai sessions. The answer: yes. And the differences are bigger than we expected.

What we found

In one sentence: in Czech, Kai behaves like a consultant. In English, it behaves like an operator.

Czech Kai is deeper, more proactive, more didactic, and more numerically explicit. It suggests next steps in 82% of conversations (vs 70% in English). It asks clarifying questions nearly twice as often (14.8% vs 8.6%). It names specific tables by name in 81% of conversations (vs 61%).

English Kai is faster, more direct, and more literal. In English, Kai is tagged 'direct, to the point' 24% more often (46% vs 37%).

Here's what that looks like in practice:

Figure 1: Conversational behavior tags. Share of conversations exhibiting each behavior (judge-tagged, all shown assistant text). Two-proportion z on CZ vs EN.

Figure 3: Data-work behavior tags. Same method; data-assistant-specific behaviors.

It's not just the language. It's the substance.

Here's the part that surprised us: these behavioral differences persisted even when the questions were translated. Take the same question, translate it from English to Czech, and Kai still responds with more depth in Czech.

This suggests it's not the words driving the change. Something about the substance, structure, or cultural framing of how the question is asked shifts Kai's response style.

Anthropic's research found the same pattern at scale. They identified four axes along which Claude's values shift by language: deference vs caution, warmth vs rigor, depth vs brevity, and candor vs execution. Czech wasn't in their top 20 languages studied, but the pattern holds: language meaningfully changes AI behavior.

Figure 2: Claude's behavioral profile changes by language. Anthropic found this across 300,000 conversations and 20 languages.

Source: Anthropic Research (July 2026).

The practical gotcha

One significant limitation we found: if users translate terminology that exists as English string literals in the database, Kai's accuracy drops.

For example, if your table has a column called order_status with values like completed, pending, cancelled. If you ask about "dokončené objednávky" (completed orders), Kai has to map that Czech term back to the English string completed in your data. Sometimes it gets this right. Sometimes it doesn't.

For bilingual users, the practical advice is: use Czech for exploratory questions and explanations where you want depth. Switch to English (or use the exact column/value names) when precision on specific data points matters.

One important note

This is not a quality claim. Depth and proactivity are value expressions, not correctness. Whether the consultant profile or the operator profile is "better" depends entirely on what you need in that moment. This research only establishes that the difference exists, is large, and holds up to the checks we could run.

So what are we doing about it?

Honestly? This is just a genuinely interesting finding. If you're bilingual, test it. Switch languages mid-conversation and see what happens. Have fun with it.

And if you find something we missed, tell us. We love this stuff.

📚 Anthropic's research: How Claude's values vary by model and language
📚 Building Kai (blog)