Google was supposed to be the leader in the chatbot revolution, but OpenAI beat it to the punch with its ChatGPT. How did Google develop AI chatbots Bard and Gemini? Forbes recounts the highlights from Wired.
Get insights that give you an edge: subscribe to Forbes Digital and always stay one step ahead in the business world.
From 49 UAH per month!
In December 2022, Google gave CEO Cissy Shao 100 days to build a competitor to the popular large-scale language model ChatGPT, developed by Sam Altman’s OpenAI. The task was no easy one – the public ChatGPT was gaining millions of users, while Google, despite having its own language model LaMDA, kept it behind closed doors.
Google was losing ground: advertising sales were stalling, innovation was slowing, and the Assistant voice assistant was not enthusiastic even among timer fans. By the end of 2022, the share price of Google's parent company, Alphabet, was 39% lower than at the end of 2021.
Popular Category Innovations Date Today “We have not recorded attacks of this level.” The head of Ukrzaliznytsia on the restoration of IT systems after a cyberattack, assistance to Kyivstar and the problem of tariffs. Blitz interview
The response to the crisis was radical: to merge the new products of DeepMind in London and Google Brain in Mountain View, to create a next-generation language model codenamed Gemini, and to restore Google's status as a leader in AI. The decision provoked internal disputes, night shifts, and a deep restructuring. Engineers, marketers, and lawyers worked overtime. The key motto of Shao's team was: “Quality over speed, but still fast.”
To Bard in 100 days
The priority was clear. In 100 days, Bard, a future competitor to ChatGPT, was created, which became an internal project of the company. The team for this project was selected for strategic thinking, emotional intelligence and flexibility. They often performed several roles at once. The 100-day deadline forced a lowering of the security bar. Some employees who had previously checked content for child abuse were transferred to test Bard.
At the same time as Bard launched, the company was experiencing its first mass layoffs in its history – 12,000 jobs, or 7% of its workforce. Some employees understood that if they wanted to stay in the company, they had to work overtime. The infrastructure team was freeing up servers, engineers were shifting resources, burning data and energy. However, Bard was “hallucinating” from the very beginning – sometimes confusing information, sometimes issuing phrases that contained racial stereotypes.
In February 2023, as Google was preparing to introduce Bard to the world, the company released a promotional video: the chatbot, like an “intelligent assistant,” answered user questions. One of the queries was: “What new discoveries from the James Webb Telescope can I tell my nine-year-old about?” In its response, Bard claimed that JWST had taken the first photo of an exoplanet —and that it was a factual error.
In fact, the first photo of an exoplanet was not taken by JWST, but by the Very Large Telescope . The mistake was quickly spotted by Reuters journalists. It went viral, and as a result, Alphabet's stock price fell by 9% (about $100 billion in market capitalization).
Google was developing two AI products in parallel: DeepMind in London and Google Brain in Mountain View, but decided to merge them into one – first Bard, and later Gemini. Photo by Getty Images
Gemini as the new Bard replacement
After the failure with Bard, Google intensified its efforts around the conversational chatbot Gemini . In April 2023, the company announced the creation of a new division of Google DeepMind, led by Demis Hassabis. He received the green light for an ambitious project – to create a visual-auditory agent with a long memory, capable of accompanying the user in everything. The technical team was united in a new office Gradient Canopy, with strict access controls even for other Google employees.
The work continued nonstop—across eight time zones, in hundreds of chat rooms, with a focus on speed and control. At the same time, Google was curtailing access to researchers, limiting publications about the product. Gemini’s training formula was too valuable. But as it grew in scale, new problems emerged—the model would give strange medical advice or draw erroneous conclusions based on images. The ethics team had little time to detect potentially toxic scenarios.
Despite all the risks, in December 2023, Google presented Gemini. It really looked like a comeback: the model beat GPT-4 in most tests, demonstrating its power in analyzing scientific texts, videos, and legal documents. But the celebration did not last long – attempts to scale the functionality immediately began: long memory, podcast generation, PDF processing. The team began to integrate Gemini into all the company's products.
Demis Hassabis and Google Brain project manager Jeff Dean now had to juggle requests from different teams: one needed Japanese translation, the other needed video analysis. All this was done against the backdrop of external competition, as ChatGPT continued to outpace Gemini in downloads: 600 million to 140 million, according to Sensor Tower.
But the main challenge was not in the number of functions, but in proving to investors and users that AI could generate profit.
ChatGPT has 600 million app downloads, while Gemini has 140 million, according to Sensor Tower. Photo: Getty Images
Ethical issues
When Google launched its image generator as part of Gemini in February 2024, it didn’t produce the racist or sexist images that were feared. Instead, when asked “US senator from the 1800s,” Gemini returned black women, Asian men, or a Native American woman in feathers, but no white men. There were also more disturbing examples, such as depictions of Nazi soldiers as people of a different skin color.
Elon Musk, who launched his chatbot Grok 3 in February 2025, has repeatedly written in X about Gemini’s shortcomings, calling the chatbot “racist and sexist” and singling out a Gemini team member he believes is responsible. According to colleagues, the employee has closed his social media accounts, fearing for his safety. Google has halted Gemini’s ability to generate images of people, and Alphabet’s stock has fallen again.
The company quickly hired new people to create a trust and safety department, and Shao ramped up internal support and developed a set of public principles for Gemini that were framed around “you,” the user. A key point was the emphasis that “answers do not necessarily reflect Google’s beliefs or views,” according to the principles.
This allowed the responsibility for the erroneous result of the chatbot's response to be shifted: “Gemini's results depend largely on what you ask it to do. Gemini is what you do.”
However, employees from ethics departments had doubts: the time for verification was reduced, and the launch of new features was accelerated.
The next step was the AI Overviews feature in Google search, which was supposed to synthesize results and immediately give a short answer. However, the system started offering suggestions like “add glue to pizza sauce.” The company reduced the number of such “suggestions,” but did not abandon them. Users could not manually disable the feature.
Google insisted that most users were happy, but there was a struggle within the team between those eager to catch up with OpenAI and those urging caution. The search engine, once associated with truthfulness, now risked falling victim to its own experiment in generating nonsense.
Gemini had problems at launch when the chatbot started generating nonsense in its advice to users. Photo illustration generated using Meta.ai AI
Quadrillions instead of trillions
Despite all the crises, there was a sense of excitement inside Gradient Canopy . When the Google Labs team presented a podcast generated by Gemini, Google Labs head Josh Woodward couldn’t contain his excitement. The NotebookLM Audio Overviews text-to-podcast tool was part of the Google I/O presentation at the conference. While other projects, such as Astra (an AI assistant that sees the real world) and the search update, attracted attention, this did not diminish the team’s faith in their work.
Software engineer Noam Shazir, one of the creators of the “Transformers,” also believed in this team again. In 2021, he left Google because the company did not want to make the chatbot publicly available, but later returned to the Gemini project. “Organizing information is a trillion-dollar opportunity. But a trillion dollars is no longer cool,” he explained his decision. “Cool is a quadrillion dollars.”
For Shazir and Hassabis, the ultimate goal is to create an artificial general intelligence that can think, plan, and help humanity. But Google also needs to find a business model. For now, the company is pushing a classic formula: your data in exchange for a free tool.
Competition is growing, the energy consumption of AI models is colossal, but profits are not yet guaranteed. Products with long context, video generation, voice podcasts – all this requires billions of dollars of investment and no less trust. It is this that Google is afraid of losing by launching unfinished products.
Two years after the “code red” inside Google have passed like one long day: crises, cuts, breakthroughs, new features, hopes and fears. The company is forced to balance between speed of development and control, between caution and market pressure. And although Gemini has indeed brought Google back into the game, victory is not yet guaranteed. In the coming years, Google could lose up to a quarter of its advertising revenue due to antitrust decisions, warns JPMorgan analyst Doug Anmuth.
