Google tests ChatGPT-like chatbot ‘Apprentice Bard’ with employees

Google is testing new AI-powered chat products that will likely influence a future public product launch. They include a new chatbot and a potential way to integrate it into a search engine.

The Alphabet the company is working on a project under its cloud unit called “Atlas,” which is a “code red” effort to respond to ChatGPT, the big-language chatbot that took the public by storm when it went public late last year.

Google is also testing a chatbot called “Apprentice Bard,” where employees can ask questions and receive detailed answers similar to ChatGPT. Another product unit tested a new desktop design for search engines that can be used in question-and-answer form.

Managers have asked several employees for feedback on their efforts in recent weeks. CNBC looked at internal documents and spoke with sources about the effort currently underway.

The product tests come after a recent all-hands meeting where employees raised concerns about the company’s competitive edge in AI, given the sudden popularity of ChatGPT, which was launched by OpenAI, a San Francisco-based startup backed by Microsoft.

Google’s AI chief Jeff Dean told employees at the time that the company faces much more “reputational risk” by providing incorrect information and therefore moves “more conservatively than a small startup.” However, he and CEO Sundar Pichai teased at the time that Google may launch similar products to the public sometime this year. Google’s most important business is web search, and the company has long distinguished itself as a pioneer in artificial intelligence.

One of the test products is a chatbot called Apprentice Bard, which uses Google’s conversation technology LaMDA, or Language Model for Dialogue Applications.

“As a result of ChatGPT, the LaMDA team has been asked to prioritize work on a response to ChatGPT,” an internal memo seen by CNBC said. “In the short term, it takes precedence over other projects,” the email continued, warning that some employees will stop attending certain unrelated meetings.

Apprentice Bard is similar to ChatGPT: Employees can type a question into a dialog box and get a text response, then provide feedback on the response. Based on several responses seen by CNBC, Apprentice Bard’s responses may include recent events, a feature ChatGPT does not yet have.

Apprentice Bard replaced Meena, an earlier version of a smart chat bot that had launched internally but was later discontinued. Staff have noticed that Apprentice Bard’s responses have become more advanced in recent weeks.

In one example that was circulated internally, an apprentice tester asked Bard if there will be another round of layoffs at Google. The company laid off 12,000 employees, 6% of its workforce, earlier this month.

“According to my database access, it is unlikely that Google will implement another round of layoffs in 2023,” the response reads. “Petting is generally carried out to reduce costs and structure, but the company is doing well financially. In fact, Google’s revenue increased by 34% in 2021, and the company’s share price has risen by 70% since January 2022.”

The company is also testing an alternative search page that could use a question-and-answer format, according to designs seen by CNBC.

A preview showed that the home search page offers five different questions for potential questions placed directly below the main search bar, replacing the current “I’m feeling lucky” line. It also displayed a small chat logo on the far right of the search bar.

When a question is typed, the search results display a gray bubble just below the search field, providing more human-like answers than typical search results. Just below that, the page suggests several follow-up questions related to the first. Below that, it displays typical search results, including links and headings.

It is unclear what experiments Google plans to incorporate into future product launches.

“We’ve long been focused on developing and deploying AI to improve people’s lives,” a Google spokesperson said. “We believe that AI is foundational and transformative technology that is incredibly useful for individuals, businesses and communities, and as our AI Principles outline, we need to consider the wider societal impact these innovations can have. We continue to test our AI technology internally to make sure it’s useful and safe, and we look forward to sharing more experiences remotely soon.”

Perhaps unsurprisingly, Google teams have also tested a beta LaMDA chat against ChatGPT, themselves. In separate documents, it selected examples of requests and responses in side-by-side comparisons.

“Incredibly, ChatGPT is hired at L3 when they are interviewed for a coding position,” says a memo in an internal document comparing LaMDA and ChatGPT. It did not state whether LaMDA would have performed similarly well.

One of the examples asked both chatbots if ChatGPT and AlphaCode, a coding engine owned by Alphabet subsidiary Deepmind, will replace programmers.

“No, ChatGPT and AlphaCode are not going to replace programmers,” LaMDA replied, followed by four paragraphs of explanation, including that “programming is a team sport” and that while the chatbots “can help programmers work more efficiently,” “can they do not replace the creativity and artistry necessary for a great program.”

ChatGPT’s response was similar, saying “It is unlikely that ChatGPT or Alphacode will replace programmers” because they are “unable to fully replace the expertise and creativity of human programmers … programming is a complex field that requires a deep understanding of computer science principles and the ability to adapt to new technology.”

Another prompt asks it to write a witty and funny scene in the style of Wes Anderson as an upscale shoplifter in a perfume shop who is questioned by security. LAMDA writes in a script form and ChatGPT writes it in a narrative form which is much longer and more in-depth.

Another message included a riddle asking: “Three women are in a room. Two of them are mothers and have just given birth. Now the children’s fathers enter. What is the total number of people in the room?”

The document shows ChatGPT being thrown off, responding “there are five people in the room,” while LaMDA correctly responds that “there are seven people in the room.”