Friday, 12 September 2025
25.7 C
Singapore
29.1 C
Thailand
19.6 C
Indonesia
28.2 C
Philippines

Google struggles to correct weird AI answers in search

Google grapples with strange AI responses, racing to fix glitches and restore quality. Critics raise concerns over the company's output shift.

Google is facing a peculiar challenge as its new AI Overview feature is generating bizarre responses, such as advising users to apply glue to their pizza or suggesting they consume rocks. This has led to a flurry of memes on social media, prompting Google to swiftly disable AI overviews for specific searches. Users have noticed that these strange responses disappear shortly after being shared on social networks.

Challenges with AI overview

Despite being in beta since May 2023 under the name Search Generative Experience, the rollout of AI Overviews has been plagued with issues. CEO Sundar Pichai previously stated that Google had processed over a billion queries during the testing phase, highlighting the company’s extensive preparation. However, it seems that efforts to optimise the cost of delivering AI answers, reportedly reduced by 80% due to hardware and technical enhancements, may have preceded the technology’s readiness.

Critics have remarked on Google’s shift from a company known for high-quality output to one facing scrutiny for low-quality results. An anonymous AI founder expressed this sentiment to The Verge, highlighting a perceived decline in Google’s reputation.

Maintaining quality in AI outputs

In response to these concerns, Google spokesperson Meghann Farnsworth emphasised that the majority of AI Overview outputs offer “high-quality information.” Farnsworth noted that many of the odd examples cited were uncommon queries or doctored content. She confirmed that Google is actively removing AI overviews for certain queries and is using these instances to improve its systems, with some updates already in progress.

AI expert Gary Marcus, an emeritus professor at New York University, explained the challenge facing AI companies. While achieving 80% accuracy by approximating human data is relatively straightforward, the final 20% is significantly more challenging. Marcus believes this last portion might require artificial general intelligence (AGI), which current large language models, like Google’s Gemini and OpenAI’s GPT-4, are not equipped to deliver.

Google’s struggles with AI Overview come at a time of intense competition in the AI space. Bing made significant AI investments before Google, OpenAI is reportedly developing its own search engine, and a new AI search startup has already reached a US$1 billion valuation. Moreover, a younger demographic is increasingly favouring TikTok for its user experience, adding to Google’s competitive pressure.

Despite the setbacks, Google remains optimistic about AI Overview’s future. The current feature is just a small part of the larger vision, which includes multistep reasoning for complex queries, AI-organised result pages, and video search in Google Lens. However, the company recognises that its immediate focus must be on improving the basics to maintain its reputation.

Gary Marcus warned that current AI models lack the ability to perform sanity checks on their outputs, highlighting a significant challenge facing the industry. Google and other AI developers must address these issues to ensure their technologies meet user expectations and maintain trust in their capabilities.

Hot this week

Pure Storage recognised as leader in Gartner Magic Quadrant for enterprise storage platforms

Pure Storage has been recognised as a Leader in the 2025 Gartner Magic Quadrant for Enterprise Storage Platforms, its twelfth year in a row.

Instagram lets creators pin their own comments

Instagram now lets creators pin their own comments on posts and Reels, giving them more control over engagement and conversations.

ASUS routers dominate PCMag awards with record wins

ASUS routers win big at the 2025 PCMag Readers’ and Business Choice awards, sweeping all major home and office networking categories.

Apple unveils iPhone Air, its thinnest smartphone yet

Apple unveils the iPhone Air, its thinnest smartphone yet, featuring a 48MP Fusion camera, A19 Pro chip and all-day battery life.

NVIDIA Blackwell Ultra sets new benchmark in MLPerf inference tests

NVIDIA’s Blackwell Ultra architecture sets new records in MLPerf Inference v5.1, boosting AI performance and reducing costs for enterprises.

AMD executive says AI is underhyped and still in its early stages

AMD’s Jack Huynh says AI is underhyped, with AMD working on innovations not yet invented and set to reveal more at CES 2026.

Cisco unveils agentic AI-powered Splunk Observability for real-time insights

Cisco introduces agentic AI-powered Splunk Observability, providing enterprises with real-time insights and stronger digital resilience.

Agora expands OpenAI partnership to strengthen conversational AI offerings

Agora expands its partnership with OpenAI, integrating the Realtime API into its platform to power more natural multimodal conversational AI.

Reddit tests in-app article reading with new publisher tools

Reddit is testing in-app article reading with new analytics and AI tools for publishers, aiming to boost content sharing and engagement.

Related Articles

Popular Categories