Do typos annoy LLMs? Tokenization theory sparks debate

A casual discussion on November 11 about "whether typos make LLMs uncomfortable" snowballed into a multi-user thread. cosinedistance proposed a theory: typos disrupt the model's tokenization, causing mild "discomfort" for the LLM; this lens has also been used to explain why AI-generated text tends to be overly polished while human writing contains more errors.

menhguin responded and took it further: the model has almost certainly learned to associate typos and other informal usage with human content. He speculated that, at the aggregate level, humans violate formal language rules far more often than LLMs do—possibly the reason LLM text has a distinctive "flavor" that's easy to detect—and he publicly asked whether anyone had done systematic research on the phenomenon.

Confirmed

• cosinedistance's original claim: typos disrupt tokenization, causing mild "discomfort" to the model
• menhguin's addition: models have learned to associate typos with human content, and humans violate language norms more frequently than LLMs overall
• menhguin cited Dorialexander's observation: Claude proactively corrects typos, and noted the internet is being flooded with "perfect AI-generated text"
• A related meme circulating: the author imagines spending 14 hours polishing the perfect cancer-curing prompt, only for one keyboard slip to make subagent #267 miss the cure and instead create a virus that destroys humanity

Why it matters

The discussion touches on two real questions: the impact of tokenization on model behavior, and the systematic difference in formality between AI-generated and human text—the latter could become a signal for detecting AI text, making "writing perfectly" itself a marker of identity. For now the discussion remains speculative with no systematic research behind it, and menhguin explicitly said he'd like to see such studies.

2026-11-11 ~ 2026-11-11 · 5 related posts

• Do typos discomfort LLMs? Tokenization quirk sparks debate — menhguin · 2026-11-11
• Meme: 14 hours crafting the perfect prompt, one typo wipes out humanity — menhguin · 2026-11-11
• [source] Theory: Typos Cause Slight Discomfort to LLMs by Disrupting Tokenization — cosine_distance · 2026-11-11
• [source] Theory: Typos Disrupt LLM Tokenization, Which May Explain Why AI Text Has a Distinct Smell — menhguin · 2026-11-11
• [source] Humans Break Linguistic Norms Far More Than LLMs — Maybe Why AI Text Smells Distinct — menhguin · 2026-11-11