Google’s AI Missteps: A Rocky Start for Generative AI Search Features

To look at this article again, go to My Profile and then check out your saved stories.

Knight Will

Google Acknowledges Flaws in AI Summaries Due to Nature of AI Technology

Just a week following incidents where its algorithms suggested consuming stones and applying adhesive to pizza, Google conceded on Thursday that its innovative AI-powered search function required refinement. This situation underscores the dangers of Google's rapid push to market generative AI, as well as the inherent and critical shortcomings of the technology itself.

Google's AI Summaries functionality leverages Gemini, a vast language model similar to OpenAI's ChatGPT, to produce written responses to certain search inquiries by condensing information available on the internet. The ongoing surge in AI development is largely attributed to these large language models' remarkable ability to process and generate text, yet this capability also allows them to present inaccuracies or falsehoods in a persuasive manner. Employing this technology to condense information from the web could simplify the consumption of search results. However, it poses risks when the underlying sources contain discrepancies or when individuals rely on this summarized information for critical decision-making.

"According to Richard Socher, a prominent figure in the development of AI for language and the founder of the AI-focused search engine You.com in late 2021, while it's now easier to develop a rapid prototype using a Large Language Model (LLM), ensuring that this prototype doesn't produce nonsensical advice like suggesting someone eat rocks requires significant effort."

Socher points out that managing large language models (LLMs) is challenging due to their lack of genuine world comprehension and the prevalence of unreliable information on the internet. He suggests, “Sometimes, it’s more beneficial not to provide a direct answer, or to present a variety of perspectives instead.”

In a recent blog post by Google, Liz Reid, the executive responsible for search operations, highlighted that the company conducted thorough evaluations before introducing AI Overviews. However, she acknowledged that incidents such as the misleading information about consuming rocks and pizza with glue — mistakes arising from Google’s system mistakenly extracting data from a satirical piece and a humorous comment on Reddit — have led to further modifications. These adjustments aim to enhance the identification of illogical search requests and reduce the dependency on content created by users, as stated by Google.

According to Socher, You.com consistently steers clear of the mistakes commonly seen in Google’s AI summaries. This is achieved through the implementation of approximately twelve strategies designed to ensure large language models (LLMs) function correctly during search operations.

"Socher highlights their commitment to precision, noting, 'Our accuracy stems from the significant investment we make in it.' You.com distinguishes itself by employing a unique web index that aids large language models (LLMs) in avoiding misinformation. Additionally, it leverages a variety of LLMs for different inquiries and incorporates a citation feature to address discrepancies in sources. Nevertheless, perfecting AI-driven search remains a complex challenge. A recent examination by WIRED revealed that You.com incorrectly responded to a question that has historically stumped other AI systems, erroneously claiming, 'There are no countries in Africa starting with the letter 'K.' Despite previously passing this test with flying colors."

Google has updated its most popular and profitable service with generative AI, a move reflecting a broader shift in the technology industry sparked by OpenAI's introduction of ChatGPT in November 2022. Shortly after ChatGPT's launch, Microsoft, an important ally of OpenAI, incorporated its technology into its underdog search engine, Bing. Despite the update leading to AI-induced mistakes and strange responses, Microsoft's CEO, Satya Nadella, embraced the strategy as a competitive push against Google, expressing his desire to shake up the industry by stating, "I want people to know we made them dance."

Achieving perfection in every aspect is beyond the realm of possibility for AI, as that's simply how it operates.

Several specialists believe that Google hurried its artificial intelligence enhancement. "It's unexpected that they rolled it out in its current form for so many types of searches—including those related to health and finance. I anticipated they would exercise more caution," states Barry Schwartz, the news editor at Search Engine Land, a platform monitoring the search sector. He also mentioned that the organization ought to have predicted that individuals would deliberately attempt to confuse the AI Overviews. "Google needs to be cautious with this," Schwartz emphasizes, particularly since these results are presented by default on their most significant offering.

Lily Ray, an expert in optimizing web search rankings, spent a year testing an early version of what is now known as AI Overviews, a project initially named Search Generative Experience by Google. She expressed that the mistakes observed last week did not come as a shock to her, considering the tendency of the earlier iteration to malfunction. “Expecting it to be error-free at all times is practically unrealistic,” Ray comments. “That's just how AI operates.”

Authored by Tess

Authored by Reece Rogers

Authored by Will

Authored by Kim Zetter

Even though obvious mistakes like recommending the consumption of rocks might decrease, AI search can still encounter various issues. Ray has highlighted less noticeable issues with AI Overviews, such as instances where they rely on weak sources like websites from a different region or those that no longer exist. According to her, this could lead to users receiving less valuable information, especially when they're looking for product advice. Content creators who aim to improve their rankings in Google's Search algorithm are currently grappling with understanding these challenges. "The degree of uncertainty within our field is unprecedented," she states.

As industry specialists and users continue to understand the nuances of Google's latest search functionality, it's unrealistic to anticipate a halt in errors. Daniel Griffin, a researcher and search consultant working on creating tools for easy comparison of various AI-driven search platforms, points out that Google encountered comparable challenges with the introduction of Featured Snippets in 2014. Featured Snippets provided answers to search queries by quoting text from websites.

Griffin anticipates that Google will address some of the most noticeable issues with AI Overviews, yet he emphasizes the significance of acknowledging that the challenge of Large Language Models (LLMs) misunderstanding facts or generating false information remains unresolved. "This issue extends beyond AI," he notes. "It encompasses the internet, the global situation. There isn't an absolute truth, per se."

Suggested for You…

Delivered directly to your email: Dive into the future with Will Knight's Fast Forward, focusing on AI progress.

Indian electorates are facing an onslaught of countless deepfake videos

Incarcerated individuals acquired tablets behind bars, only to discover an unful

The single factor impeding the progress of heat pumps

Eternal Sunshine: Discover the Perfect Shades for Any Journey

Steven Levy

Lauren Goode

Reece Rogers

Lauren Goode

Dave Paresh

Knight Will

Knight Will

Knight Will

Additional Content from WIRED

Evaluations and Manuals

© 2024 Condé Nast. All rights reserved. Purchases made via our website may result in WIRED receiving a share of the revenue, as a component of our Affiliate Agreements with retail partners. Content from this site is prohibited from being copied, distributed, broadcast, stored in a retrieval system, or used in any form without explicit consent from Condé Nast. Advertising Choices

Choose a global website