---
title: "Google AI counted two 'P's in its own name: why neural networks can't count"
description: "Google's AI search caused a real embarrassment: the algorithm counted two 'P's in the word 'Google' 🤯. It turns out that neural networks cannot count letters due to architectural peculiarities. We analyze why the 'omnipotent' AI cannot handle the alphabet and why you cannot blindly trust its answers 🤖❌"
date: 2026-05-28T11:55:08.000Z
lang: en
url: https://xab.info/en/posts/google-ai-counted-two-ps-in-its-own-name-why-neural-networks-cant-count
tags: []
publisher: "XAB.info"
---

# Google AI counted two 'P's in its own name: why neural networks can't count

![Google logo with a spelling error where the letters G and O are replaced by P, illustrating AI's inability to count characters](https://xab.info/media/2026/05/28/ii-google-ne-smog-poschitat-bukvy-v-slove-google/logo-google-s-oshibkoi-bukv.webp)

It might seem like a trivial task for an AI of Google's caliber: count the letters in a word. However, the company's latest AI search feature demonstrated a surprising inability for such an "omnipotent" machine. Instead of the expected result, the algorithm confidently stated that the word "Google" contains two letters "P". This is not an isolated glitch but a systemic feature that has once again reminded users of the fragility of modern neural networks.

### From advice to eat stones to alphabet errors

The public has not encountered updates from the search giant with irony for the first time. Not long ago, "AI Overviews" were notorious for absurd recommendations, suggesting users grease pizza with glue or eat stones. Now the focus has shifted to basic literacy. Google itself acknowledges the problem, calling character counting a known challenge for large language models, and promises to work on fixing it.

### Why doesn't AI see letters?

The paradox lies in the architecture of neural networks. Large Language Models (LLMs), which underpin chatbots, do not "read" text the way we do. For them, a sentence is not a set of letters and words, but a stream of tokens. Transformers break information into fragments (syllables, parts of words, symbols) and translate them into numerical vectors, analyzing context to generate a logical response.

That is why AI can solve a complex mathematical problem that scientists struggle with, but stumble on spelling, producing a result worthy of kindergarten. Researchers agree: this is a fundamental architectural limitation, not a temporary bug. The priority for such tasks is low, as the value of the model lies in generating meaning, not in perfect spelling.

### A lesson in distrust

These errors, though they look funny, serve an important function: they shatter the illusion of omnipotence. AI is a powerful tool, but not a perfect oracle. As recent incidents show, when an algorithm interprets the word "ignore" as a command, or confuses letters in its own logo, blindly trusting its answers is dangerous. A neural network may seem all-knowing, but in matters of accuracy and factuality, it remains only a probabilistic generator.