Google’s AI-powered search experience is facing fresh ridicule online after users discovered it repeatedly fails at one of the simplest language tasks imaginable: counting letters in words.

Ask Google’s AI Overview how many “Ls” are in the word “Google,” and it may confidently answer “2.” Ask how many “Ms” are in “Gemini,” and it might also say “2.” Even stranger, some users report that the tool invents letters that do not exist in a word at all.

One example making the rounds online involved the query:
“How many Ps are there in Google?”

Google AI Overview reportedly answered:
“There is only 1 ‘p’ in the word ‘Google.’”

There are, of course, zero Ps in “Google.”

The mistakes have become viral fodder at a time when public skepticism around artificial intelligence is already growing. Critics argue that when one of the world’s largest technology companies cannot get elementary-school spelling questions right, it raises broader concerns about the reliability of AI-generated answers.

Why Google AI Overview Is Struggling With Simple Words

The irony is that modern AI systems can summarize research papers, write software code, and generate realistic images, yet stumble over counting letters in a five-letter word.

The reason comes down to how large language models (LLMs) process text.

Contrary to how humans read language, AI systems like Gemini do not naturally “see” words as sequences of individual letters. Instead, they break language into “tokens,” which are chunks of text that may represent entire words, syllables, prefixes, or commonly used character patterns.

For example, an AI model may process the word “apple” as a single unit rather than recognizing A-P-P-L-E separately.

That creates a surprising weakness:
The model develops a statistical understanding of language without true character-level awareness.

When asked to count letters, the system often attempts to “reason” probabilistically rather than literally inspecting each character one by one.

That is why AI can produce highly sophisticated writing while simultaneously failing at tasks most humans solve instantly.

Even Gemini Admits This Is a Known Problem

When users asked Google Gemini why AI tools struggle with spelling and letter counting, the chatbot itself acknowledged the issue.

Gemini reportedly explained:

“Rather than reading text like a human, these models dissect language into tokens.”

It added that because words are often processed as single token IDs rather than individual characters, AI systems lack an inherent understanding of spelling patterns.

The chatbot eventually summarized the problem with unusual honesty:

“An AI can write a brilliant essay on Shakespeare but may still stumble if you ask it to count the number of ‘n’s’ in the word ‘banana.’”

That explanation aligns with longstanding limitations in transformer-based language models.

The “Google Has Two Ls” Mistake Has an Odd Backstory

One particularly bizarre response involved Google AI Overview insisting that “Google” contains two-letter “Ls.”

In at least one case, the AI appears to have pulled information from an old 2009 article discussing Google’s 11th birthday doodle, stylized as “Goog11e.”

Instead of understanding the actual word being queried, the system seemingly relied on contextual web information connected to the topic.

That illustrates another weakness of generative AI search systems:
They often prioritize contextual associations over literal correctness.

In traditional search, users would receive links to webpages containing the answer. AI Overviews instead synthesize responses directly, which increases the risk of confidently presenting incorrect information.

Google’s AI Search Is Also Confusing Words for Commands

The spelling issue is not the only strange behavior users have uncovered.

Some users discovered that entering words like “disregard,” “ignore,” or “forget” into Google Search caused AI Overview to interpret them as instructions rather than dictionary queries.

Instead of defining the words, the AI reportedly responded with messages like:

“Understood. Let me know whenever you have a new prompt or question!”

That behaviour resembles prompt injection confusion, a phenomenon where AI systems misinterpret user input as operational commands.

The issue highlights a broader challenge facing AI-powered search products:
Natural language systems often struggle to distinguish between content and instructions.

For users, that creates moments that feel less like intelligent search and more like accidental comedy.

Why These Errors Matter More Than They Seem

On the surface, these mistakes are funny.

But for Google, they also represent a credibility problem.

The company has aggressively integrated AI into search in response to mounting competition from AI chatbots and generative search platforms. Google’s AI Overview is designed to keep users inside the Google ecosystem by delivering instant, synthesized answers instead of simply showing blue links.

That strategy only works if users trust the answers.

When AI confidently produces incorrect spellings, invented letters, or bizarre command interpretations, it reinforces a growing criticism of generative AI:
These systems often sound authoritative even when they are wrong.

Researchers sometimes refer to this phenomenon as “hallucination,” though critics argue the term understates the seriousness of fabricated information.

And while spelling errors are relatively harmless, the same underlying flaws can affect:

• Medical information
• Financial advice
• Legal summaries
• News interpretation
• Scientific explanations

That is why seemingly silly mistakes attract outsized attention. They expose the gap between AI fluency and AI reliability.

Why AI Still Struggles With Precision Tasks

Large language models are fundamentally prediction engines.

They excel at generating likely language sequences based on patterns learned from enormous datasets. But they are not inherently designed for precise symbolic reasoning.

Tasks involving:

• Exact counting
• Letter-by-letter analysis
• Arithmetic precision
• Structured logic
• Deterministic outputs

can still trip up systems that otherwise appear highly intelligent.

Some AI companies are now integrating hybrid systems that combine language models with symbolic tools, calculators, or external verification layers to reduce these errors.

But the fact that such fixes are necessary reveals an important truth:
Fluent language generation is not the same thing as understanding.

The Bigger Risk for Google: Public Trust

Google’s AI rollout has already faced criticism over factual inaccuracies, bizarre recommendations, and manipulated search summaries.

Every viral mistake compounds a larger perception issue.

Users are increasingly asking whether AI-generated answers are replacing traditional search quality with speed and spectacle.

For Google, the challenge is no longer simply building more advanced AI systems.

It is convincing users that those systems can reliably handle basic facts without inventing information along the way.

And right now, being unable to count the letters in “Google” is not helping that case.

TL;DR

• Google AI Overview is being mocked online for repeatedly miscounting letters in simple words.
• Users reported incorrect answers for words like “Google,” “Gemini,” and “journalism.”
• AI systems process text as tokens rather than individual letters, causing character-level mistakes.
• Gemini itself acknowledged that letter counting remains difficult for large language models.
• Users also found cases where Google AI Overview treated words like “ignore” and “disregard” as commands.
• The incidents highlight broader concerns about AI reliability and hallucinations in search results.