Define 'fertility' as a tokenizer metric and explain why high fertility for a language is both a fairness and a model performance problem.

**Fertility** is the average number of tokens a tokenizer produces per word, measured over a representative corpus. Fertility 1.0 means one token per word; fertility 3.0 means three tokens per word on average. English on modern tokenizers is around 1.3; many other languages are much higher. There are two separate problems, and it helps to keep them apart. The first is fairness. Providers charge per token, so the same message in a high-fertility language costs more, sometimes three to five times more, and it eats more of the context window. A user writing Swahili pays a premium for an equally complex question, purely because of how the tokenizer chops their language. The second is model quality. Training runs on a fixed budget of tokens. If a language needs five tokens where English needs one, the model sees fewer complete words for the same compute, so it learns that language less well. Research on the token tax found that doubling fertility lines up with about four times the training cost to reach the same accuracy. So the high-fertility language is both more expensive to use and genuinely worse served by the model.

**Fertility** is the average number of tokens per word. $$\text{fertility} = \frac{\text{number of tokens}}{\text{number of words}}$$ English sits near 1.3. Underrepresented languages can reach 3 to 5. The key move in a strong answer is to split the harm into two distinct problems. The fairness problem is economic. Providers bill per token, so identical semantic content in a high-fertility language costs proportionally more, and it consumes more of the context window. A Swahili user asking an equally complex question can pay several times what an English user pays and fit less context. That penalty is structural, set by tokenizer design, not by question difficulty. The quality problem is about learning. Training budget is counted in tokens. At a fixed budget, a high-fertility language delivers fewer complete semantic units, and each concept is smeared across more fragments, so signal per token drops. Tokenizer-fairness research finds doubling fertility tracks roughly 4x the compute to match accuracy. Equivalently, at fixed scale and budget, the high-fertility language lands at lower accuracy. So the same design choice taxes the user twice: more cost and less capable output.