Google AI Overview claims there are 2 Ps in Google. It also states there is exactly 1 R in the word poop, counts 2 Ds in journalism (which it spelled j-o-u-r-n-a-d-i-s-m), and identified 1 P in the last name of the U.S. president — which it spelled t-r-p-u-m. The product that ate rocks and glue last year is now struggling with kindergarten-level spelling.
Google confirmed the issue in an emailed statement: counting within words has been a known challenge for LLMs, and the company is working to fix this particular issue. The failures arrive weeks after Google redesigned its 29-year-old flagship search engine to make generative AI the centerpiece experience. AI Overviews now render above organic results for most queries.
This is not the first time AI Overview has stumbled. Last year the product cited satirical posts from The Onion and Reddit, advising users to eat rocks and put glue on their pizza. Last week Google patched a separate bug where searching the word disregard yielded what looked like a dictionary definition, except the definition read: Understood. Let me know whenever you have a new prompt or question.
Key facts
- 01Google AI Overview claims there are 2 Ps in Google, 1 R in poop, and 2 Ds in journalism — which it spelled j-o-u-r-n-a-d-i-s-m.
- 02The product now anchors Google's 29-year-old flagship search engine after a May 2026 redesign doubled down on generative AI.
- 03Transformer-based LLMs encode text as numerical tokens representing words, syllables, or letters — not as discrete characters humans can count.
- 04Google patched a separate issue last week where searching 'disregard' returned a fake dictionary definition reading 'Understood. Let me know whenever you have a new prompt or question!'
- 05Researchers say no perfect tokenizer exists because models inherently chunk text in ways that obscure letter-level structure.
The spelling errors persist because LLMs do not perceive text as words made of letters. Transformer-based models — the architecture powering most chatbots and text-generators — break sentences into tokens, which can be full words, syllables, or individual letters depending on the model. The AI converts those tokens into numerical representations and predicts the next token in a sequence. At no point does the model parse text character-by-character the way a human reader does.
The token-based design makes letter-counting structurally hard. A model trained on the token for the word strawberry does not know there are three Rs in that token unless it happens to break the word into subword chunks that isolate the Rs. When prompted to count letters, the model is guessing based on patterns it has memorized, not performing arithmetic on a string.
Researchers do not expect a fix soon. The architecture would need to represent every word as a sequence of discrete character tokens, which would balloon training costs and inference latency. Even then, edge cases around punctuation, diacritics, and rare words would remain fuzzy.
The failures are amusing but load-bearing for trust. Google's AI Overview now sits atop the world's most-used search engine, returning confident-sounding answers that are trivially wrong. Users who relied on Google Search for factual lookup now face a product that cannot spell the name of the company that built it.
The deeper problem is architectural. Transformer models excel at pattern-matching and probabilistic text generation. They do not excel at tasks that require symbolic reasoning over discrete units — like counting letters, verifying citations, or distinguishing satire from fact. Google is retrofitting a 29-year-old product around a technology that is not optimized for the product's core use case.
DuckDuckGo installs surged 30% in the two weeks following Google's AI Search redesign, which we covered earlier this month. Users are voting with their feet. The question is whether Google can patch the spelling errors faster than competitors can ship a search product that does not hallucinate the alphabet.
Google has not disclosed a timeline for fixing the letter-counting issue. The company's statement characterized it as a known problem, which suggests it has been triaged but not prioritized. The token architecture is load-bearing for every LLM Google ships, so a systemic fix would require rethinking inference or retraining models from scratch.
The immediate takeaway is that AI Overview cannot be trusted for literal queries. If a user asks how many letters are in a word, the model will guess. If a user asks for a spelling, the model may invent one. The product works for probabilistic summarization — drafting an email, paraphrasing a definition — but fails at tasks that require precision. Google is deploying it as if it works for both.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
