As AI chatbots take off, Helsinki-based Silo AI and other European startups race to improve responses in non-English languages, like German, Hebrew, and Arabic
Companies aim to improve response quality after criticism of reliance on small group of mostly US participants
Context & Ripple Effects
The capability gap is documented, not hypothetical: AI researchers have found that ChatGPT and its rivals perform significantly worse outside English, a shortfall the FT ties to reinforcement feedback drawn from a small group of mostly US participants. Silo AI's push into German, Hebrew, and Arabic is Helsinki's answer to a problem that has been building since the GPT-3 era, when [[a:969997|startups in China, South Korea, Israel, and Germany began building general-purpose language tools for their home markets]].
First-order effects
- Non-English users of ChatGPT-class chatbots get measurably weaker responses today, which is exactly the opening Silo AI and fellow European startups are targeting with localized tuning in German, Hebrew, and Arabic.
- The criticism of US-skewed participant pools puts direct reputational pressure on OpenAI and its peers, whose English-first feedback loops are now a named competitive weakness.
Second-order effects
- US model builders face a forced choice: invest in multilingual feedback pipelines or cede non-English markets to regional specialists — the same dynamic driving South Korean startups to train on Korean language and culture to hold their home ground.
- Localized quality becomes a procurement criterion for European enterprises and public services, giving Silo AI-type firms pricing power as the credible domestic alternative.
Third-order effects
- If regional specialists keep winning on language quality while US labs own the general frontier, the industry splits into a two-tier structure: English-first global models plus nationally tuned layers, with regulation like city-level AI rules for public services reinforcing local preference.
The trend: AI chatbots are internationalizing along two tracks — English-first global models from US labs and locally tuned challengers from Europe and Asia competing on native-language quality.