AI Treats Ukrainian Text Worse Than English — New Fixes Could Help

Have you ever noticed that AI apps sometimes feel stingier in some languages than others? A new study from computer scientist Ivan Dobrovolskyi confirms one reason: AI systems break text into tiny chunks called tokens, and they need a lot more of them for Ukrainian than for English. You can think of tokens like pieces of a jigsaw puzzle — English text uses a handful of big pieces, while Ukrainian text gets smashed into hundreds of tiny ones. The study found Ukrainian uses 68% to 121% more tokens than English on modern AI systems, and 220% more on older ones.

Why does that matter to you? Because most AI services charge by the token. More tokens means higher fees, faster message limits, and less room for the AI to remember your conversation. It also means the AI can only process a smaller chunk of Ukrainian text at once, making responses feel shallower. The problem is especially bad for e-commerce, where Ukrainian product descriptions and customer questions can get cut off or lose detail.

The good news? The paper tests two fixes that actually work. One compresses the text before sending it to the AI, reducing the length by nearly half with no loss of key information. Another redesigns the AI's word-breaking method, shrinking the Ukrainian-to-English token gap from 2.22 times down to 1.30 times. Interestingly, simply switching to Latin letters (romanization) made things worse on most systems.

The catch? These fixes aren't built into the big AI tools yet. And the study suggests the root cause is simply that AI companies trained their systems on much more English than Ukrainian data. Until that changes, Ukrainian speakers are paying a hidden tax every time they use an AI chatbot, translator, or writing assistant.