Ask a large language model to write a sonnet about quantum physics and it will deliver something passable within seconds. Ask it how many r's appear in the word "strawberry" and there is a reasonable chance it will get it wrong. This asymmetry is not a bug to be patched but a window into the fundamental nature of these systems — and understanding it matters more than any benchmark score.

The confusion stems from a category error most users make instinctively. We call them artificial intelligence, so we assume they work like minds, or at least like calculators. They do neither. A large language model is a pattern-completion engine trained on text, predicting the statistically likely next token in a sequence. When you ask it to multiply 847 by 293, it is not performing arithmetic. It is asking itself: given all the text I have seen where numbers like these appeared near multiplication signs, what digits typically followed?

The tokenization problem

The trouble begins before the model even starts thinking. Text enters these systems not as letters or words but as tokens — chunks that the model learned to recognize during training. The word "strawberry" might be split into "straw" and "berry," or into "str," "aw," and "berry," depending on the tokenizer. The model never sees the individual letters laid out for counting. It processes statistical representations of text fragments, which is efficient for language but catastrophic for tasks requiring character-level precision.

This is why a system that can explain the French Revolution with reasonable accuracy struggles to reverse a word reliably. Reversing "computation" requires knowing exactly which letters are present and in what order — information the model accesses only indirectly, through probabilistic inference about token patterns.

Reasoning versus retrieval

When a language model produces a correct answer to a math problem, it is often retrieving a pattern rather than reasoning through steps. Ask for 12 times 12 and the answer comes instantly and correctly — because "144" appeared near "12 × 12" countless times in training data. Ask for 12 times 137 and accuracy drops, because that specific combination appeared less frequently. The model must interpolate, and interpolation in high-dimensional statistical space is not the same as carrying the one.

Recent architectures have improved matters through chain-of-thought prompting and extended reasoning, which essentially force the model to show its work, generating intermediate tokens that constrain subsequent predictions. This helps, but it remains a workaround rather than a solution. The model is still predicting plausible-looking reasoning steps, not executing a mathematical algorithm.

What this means for users

None of this makes language models useless — they remain genuinely transformative tools for drafting, summarizing, coding assistance, and creative exploration. But the counting problem is a useful heuristic for calibrating expectations. Any task that requires precise enumeration, exact sequence manipulation, or guaranteed logical consistency falls outside what these systems do natively. They can often get such tasks right, especially with careful prompting, but they cannot guarantee it the way a pocket calculator guarantees 2 + 2 = 4.

The vendors know this, which is why production systems increasingly route mathematical queries to separate code interpreters rather than relying on the language model alone. The hybrid approach works, but it is an acknowledgment of limitation, not a transcendence of it.

Our take

The strawberry test is not a gotcha designed to embarrass AI companies. It is a diagnostic that reveals something important: these systems are mirrors trained on human text, reflecting our patterns back at us with eerie fluency. They are not minds, not calculators, not oracles. They are extraordinarily sophisticated autocomplete, and that framing — less flattering but more accurate — is the starting point for using them well. The users who get the most from language models are those who understand what the models actually are, not those who mistake statistical brilliance for general intelligence.