Google had already developed several conversational AI models when ChatGPT drew attention to chatbots as a way to answer questions. Deepmind’s Sparrow could give the company another route into the market: a chatbot with internet access and sources for its responses, planned for private beta later in 2023.
From research to a possible product
Google had shown dialogue-focused language models including LaMDA and Flamingo. It had also developed Meena, which could hold credible conversations with people by the end of 2020. Yet Google had not turned this research into a consumer chatbot.
The company pointed to reliability and safety concerns. Those concerns carry particular weight for a company operating at Google’s scale: a chatbot that spreads falsehoods or hate speech could create reputational risks. A system that sounds convincing could also lead people to attribute consciousness to it.
Meanwhile, ChatGPT’s rapid growth, and Microsoft’s extensive involvement in spreading OpenAI’s work, put pressure on Google to show it could compete. Sparrow offered a potential way to make that response more visible.
What Sparrow was designed to do
Deepmind introduced Sparrow in September 2022. Like ChatGPT, it was trained using human feedback, which Deepmind said was intended to make it more helpful, accurate and harmless. Sparrow was also designed to access the internet through Google, bringing current information into its answers.
The planned beta was expected to provide sources that matched the chatbot’s responses, according to Time. Showing sources could help readers check where an answer came from, while internet access could help the system draw on information beyond its training.
Sparrow was based on Deepmind’s Chinchilla language model. Chinchilla had fewer parameters than OpenAI’s largest models but was trained on a large amount of data. Introduced in April 2022, it outperformed GPT-3 in common language benchmarks. ChatGPT, however, was based on the more advanced GPT version 3.5.
Those details suggested Sparrow could be competitive, but they did not establish how it would perform against ChatGPT in practice. The article also noted that Google had more powerful language models, including PaLM, in reserve.
The business problem behind the technology
Google’s hesitation may have reflected more than technical and safety questions. Its search business was growing and highly profitable, with revenue generated by search results pages. A chatbot that answers questions directly could become an alternative to that format, but its monetization strategy was unclear.
That creates a difficult trade-off. Google could introduce a successful chatbot and still face a financial challenge if it earned less than traditional search. The company would need to respond to a possible shift in how people find information while weighing that new product against a strong existing business.
This tension is often described as the “Innovator’s Dilemma.” Clayton Christensen coined the term in 1997 for the difficulty established companies can face when new technologies or business models disrupt their traditional markets. A company may have substantial resources and capabilities, yet its existing customers and processes can make it harder to embrace changes that threaten what already works.
Why Deepmind could matter
One proposed route was for Deepmind to launch a commercially oriented chatbot as a sister company to Google. That could give Google a competitor to both its own search experience and OpenAI while providing a way to act if chatbots became a serious search alternative.
In that scenario, Sparrow’s planned beta would serve as more than a research milestone. It could help Google test a conversational product while showing shareholders that the company had a response in development. Whether that approach could bridge the gap between chatbot answers and search’s established business model remained uncertain.
Sparrow therefore represented both a technical opportunity and a business question. Internet access, human feedback and sources could shape how it answered users, but the larger challenge was deciding how a chatbot might fit alongside Google’s profitable search business.