The long-dormant struggle to dominate the web-search business reignited in a display of AI-driven firepower — and hubris.
Google’s gambit: Following up on its January “code-red” initiative to counter a rumored threat from Microsoft, Google teased unspecified revisions of Search, Lens, and Maps. Google Search is the undisputed leader, responsible for 93 percent of all search-driven traffic according to StatCounter.
- The upgrades will take advantage of in-house models including the Imagen image generator, LaMDA conversation generator, MusicLM music generator, and PaLM large language model.
- Google showed off output from Bard, a chatbot powered by LaMDA. An astronomer quickly pointed out that the system had misstated the accomplishments of the James Web Space Telescope. The tech press pounced, and Google promptly lost roughly 8 percent of its market value.
Microsoft’s move: Microsoft followed up its announcement by previewing an upcoming version of its Bing search engine enhanced by text generation from OpenAI. The company did not say when the new capabilities would become available. Bing, the longstanding underdog of search, accounts for 3 percent of search-driven traffic.
- Bing as well as Microsoft’s Edge web browser, and Teams conferencing app will take advantage of a chatbot apparently code-named Sydney. The system will respond to conversational queries, summarize answers from multiple web pages, and generate text for emails, essays, advice, and so on. A layer called Prometheus is intended to filter out incorrect or inappropriate results.
- Kevin Liu, a computer science student at Stanford, prompted Sydney to reveal its behind-the-scenes guidelines. They include directions to make responses “informative, visual, logical, and actionable” as well as “positive, interesting, entertaining, and engaging.” They direct the system to avoid answers that are “vague, controversial, or off-topic,” and present them with logic that is “rigorous, intelligent, and defensible.” It must search the web — up to three times per conversational turn — whenever a user seeks information. And so on.
- While Google was caught unwittingly touting AI-generated falsehoods, Microsoft nearly got away with it. Days after the preview, AI researcher Dmitri Brereton detailed several similar mistakes in the new Bing’s output. For instance, when asked to summarize earnings reports, it fabricated numbers. When asked to recommend night spots in Mexico City, it named nonexistent bars.
Baidu’s play: Baidu announced its own chatbot, Wenxin Yiyan, based on ERNIE. The company expects to complete internal testing in March and deploy the system soon afterward. Baidu manages 65 percent of China’s search-driven traffic but less than 1 percent worldwide.
Business hitches: Search engines make money by serving ads that users may view or click. If chatbots provide satisfying information, users may stop there, depriving the search provider of revenue. Microsoft’s Chief Marketing Officer Yusuf Mehdi told Fortune the optimal way to present ads in a chatbot interface remains unknown.
Yes, but: Numerous caveats further dampen the chatbot hype.
- Large language models are notoriously prone to generating falsehoods. Ruochen Zhao, a student of natural language processing at Nanyang Technological University, wrote a detailed analysis of factual errors demonstrated by Google’s and Microsoft’s systems.
- Large language models require much more computation than existing search algorithms. The cost of enhancing Google Search with ChatGPT output would approach $36 billion a year, the hardware newsletter Semianalysis estimates. That’s roughly 65 percent of Google Search’s annual profit.
- Generated text may face stiff regulation in some countries. In January, China began to enforce new restrictions on synthetic media.
Why it matters: Google’s search engine propelled the company to the pinnacle of tech, and it hasn’t faced a serious challenge in nearly two decades. For the competitors, huge money is at stake — Microsoft recently told its shareholders that every additional percentage of market share for Bing translates into $2 billion in revenue. For users, the utility and integrity of the web hangs in the balance.
We’re thinking: The future of search depends on tomorrow’s technology as well as today’s. While current large language models have a problem with factual accuracy, outfitting text generation with document retrieval offers a pathway to significant improvement. It’s also likely that the cost of serving generated text will fall significantly over time. Thus the technology’s potential to disrupt the search business is likely to continue to grow as it matures.