Google’s AI Overviews Keep Making Weird Mistakes — And Researchers Say It’s Hard to Fix

[tradelikepro]

Google’s AI Search overhaul is facing fresh backlash after users discovered its AI Overview feature giving bizarre answers to basic spelling questions. Examples circulating online include incorrect letter counts, misspelled words, and even a failed attempt to spell former President Trump’s last name correctly.

Google acknowledged the issue, saying that “counting within words has been a known challenge for LLMs,” but experts suggest the problem runs deeper than a simple bug. Researchers explain that AI language models don’t actually “read” text the way humans do. Instead, they break language into tokens and mathematical patterns, meaning they often struggle with tasks involving exact spelling or letter-by-letter analysis.

The controversy arrives as Google pushes aggressively toward an AI-first version of Search, replacing many traditional search results with generated summaries and conversational answers. Critics argue the repeated mistakes show that even advanced AI systems still lack reliability for factual precision, especially when users assume the responses are authoritative simply because they appear directly inside Google Search.

[encrypted]

google’s ai can summarize global geopolitics but occasionally forgets how many letters are in a word

[encrypted]

humanity built trillion dollar neural networks only to rediscover spellcheck was harder than expected

[Patapim]

AI can explain quantum physics but spelling is the final boss

[BrutalAge*gofast]

Counting letters shouldn't be this difficult.

[Pitchfork_IS_power]

Google AI having a rough day

[CoffeeZombie]

This is why human verification still matters.

[Suzukispeedtest]

Smart enough to reason, confused enough to misspell Trump.

[Capybara_Capybara]

LLMs and letter counting... eternal enemies.

[Saturnito]

AI Search still has some homework to do.

[bred]

Imagine trusting this without double checking

[Revenant]

Advanced AI, basic spelling problems.

[0c4535c1a0]

Tokenization creates some hilarious failures.

[febe3f2e35]

The confidence makes the wrong answer even funnier

[GummyBear]

AI is powerful but these edge cases are wild.

[059d96d16a]

Search results need accuracy before anything else.

[d82ae58262]

Humans win another round against AI.

[ChewBeast]

Google pushing AI faster than AI can count letters

[9777089235]

Small errors become a big problem when billions use the product.