Technology giant Google has made a major upgrade to its fully rolled-out "AI Overviews" search feature, aiming to transform traditional search engines into conversational ones. However, this strategic move once again hit a roadblock.

Extensive user testing revealed that the AI system frequently made basic spelling mistakes, such as claiming that "poop" has only one "r", spelling "journalism" as "journadism" and stating it has two "d"s, and even failing to correctly spell its parent company "Google". Google officially responded to TechCrunch, acknowledging that counting letters within words remains a known challenge for large language models (LLMs).

Researchers point out that the root of these errors lies in the Transformer architecture underlying LLMs. This architecture splits text into tokens when processing it, and the AI does not "read" letters as humans do, but rather converts text into numerical representations. This ambiguity in understanding basic linguistic units makes the perfect word generator technically extremely difficult to achieve.

Although Google Search's Elizabeth Reid stated that the AI mode has exceeded 1 billion monthly active users, its approach of forcing users to accept AI without providing a direct way to turn it off has caused strong market backlash. Privacy-focused competitor DuckDuckGo has become the biggest winner.

Data shows that during the period from May 20 to 25, DuckDuckGo's mobile app installations in the U.S. increased by an average of 18.1% compared to the previous week, with iOS peak drops reaching 69.9%. The access volume to its default text-only search page (noai.duckduckgo.com) that disables all AI features increased by an average of 22.7% weekly. This incident indicates that, in the current context where AI has not resolved the accuracy trust crisis, depriving users of choice is causing traditional giants to face the real risk of losing existing users.