According to tokenizer fairness research, if a language's fertility doubles, what typically happens to downstream model accuracy?

The correct answer is that accuracy degrades. Here is the intuition. **Fertility** is the average number of tokens a tokenizer uses per word. English usually sits near 1.3 tokens per word. Many other languages, especially ones written in non-Latin scripts or that were underrepresented when the tokenizer was built, sit much higher, sometimes 3 to 5 tokens per word. Now think about training. A model learns from a fixed budget of tokens. If a language needs four tokens to say what English says in one, then for the same number of words the model burns through four times as many tokens. Most of that extra budget goes to gluing fragments back together rather than learning meaning. Research on tokenizer fairness found that doubling fertility lines up with roughly four times the training cost to reach the same accuracy. So at a fixed budget, the high-fertility language ends up with a real accuracy gap. The wrong answers assume more tokens help or that the effect cancels out; it does not.

The trap in this question is assuming more tokens means more signal. It is the reverse: when a language's **fertility** doubles, downstream accuracy degrades, and the effect is large. Fertility is tokens per word: $$\text{fertility} = \frac{\text{number of tokens}}{\text{number of words}}$$ A model trains on a fixed token budget. If a language packs only a quarter as much meaning into each token, then per unit of compute it sees far fewer complete semantic units. Tokenizer-fairness work quantifies this: doubling fertility correlates with roughly 4x the training cost to hit the same accuracy. At fixed compute, that shows up directly as an accuracy gap. The mechanism has two sides. At training time, high-fertility languages get less signal per token about their structure. At inference time, the same prompt fills more of the context window, leaving less room for content. Both push quality down. This is the core of the token tax. It is not just that non-English users pay more per API call. It is that the model is genuinely worse at their language, for a structural reason that has nothing to do with question difficulty.