Google Gemini

• 1. Introduction

  ›

  Before Gemini was called Gemini, the engineers at Google who built it knew the project by the codename Atlas. The public first met it under a different name, Bard, a Celtic word for a storyteller. Google said the name was picked to reflect the creative nature of the algorithm underneath. Today Gemini is a family of models and a chatbot that can handle text, code, images, audio and video, and it sits inside phones, browsers and office software. But its path there was far from smooth. Why did a company famous for search feel it had to answer a rival in a matter of weeks? What went wrong when the world watched it try? How did one product end up with several names, many model versions and a trail of public arguments? The answers begin with a single viral launch in late 2022 and an alarm that rang inside one of the world's largest technology companies.

  403635

• 2. Code Red

  ›

  In November 2022, OpenAI released ChatGPT, a chatbot built on its GPT-3 family of language models, and it spread across the Internet as a viral sensation. Google's executives saw a threat to Google Search and issued a code red, reassigning several teams to help with the company's artificial intelligence work. Sundar Pichai, chief executive of Google and its parent company Alphabet, was widely reported to have issued the alert. He later denied that to The New York Times.

  Larry Page and Sergey Brin had stepped down as co-CEOs of Alphabet in 2019. Even so, both attended emergency meetings with executives to discuss a response. In February 2023, Brin asked for access to Google's code, something he had not done for years.

  Google already held a possible rival. In 2021 it had unveiled LaMDA, a prototype language model, but never released it to the public. At an all-hands meeting, employees asked whether that was a missed opportunity. Pichai and Jeff Dean, Google's AI chief, answered that LaMDA had capabilities similar to ChatGPT. They said a model that might spread false information carried risks, so the company had chosen to wait.

  By January 2023 the mood had changed. Demis Hassabis, chief executive of DeepMind, a sister company of Google Brain, hinted at plans for a ChatGPT rival. Employees were told to speed up and to test chatbots intensively, including one called Apprentice Bard. In February, Pichai told investors on a quarterly earnings call that Google would expand LaMDA's availability and applications. That promise was about to be tested in public.

  3638394344

• 3. Paris and After

  ›

  On the 6th of February 2023, Google announced Bard, a chatbot powered by LaMDA. It went first to 10,000 "trusted testers", with a wide release planned for the end of the month. Product lead Jack Krawczyk called it a "collaborative AI service" rather than a search engine. Reuters estimated that ChatGPT-like features in Search could add $6 billion in costs by 2024. The research firm SemiAnalysis put the figure at $3 billion.

  Microsoft had scheduled an event for the 7th of February to unveil its partnership with OpenAI and a new Bing assistant, later renamed Microsoft Copilot. Media outlets and financial analysts said Google was rushing to get ahead of it. Microsoft CEO Satya Nadella told The Verge: "I want people to know that we made them dance." Writers at The Guardian and Recode, and analyst Dan Ives of Wedbush Securities, called it an AI arms race.

  The livestream in Paris on the 8th of February was described as underwhelming. Google's stock fell eight percent, a loss of $100 billion in market value, and the YouTube video was made private. Viewers had spotted an error in the demo, where Bard gave wrong information about the James Webb Space Telescope. Employees called the announcement "rushed" and "botched" on Memegen, the company's internal forum, and Maggie Harrison of Futurism called the rollout "chaos". Pichai said Google had been "deeply working on AI for a long time".

  A week after Paris, Pichai asked 80,000 employees to spend two to four hours testing Bard. Executive Prabhakar Raghavan asked them to correct its mistakes. Many workers wrote internal messages citing safety and ethical concerns and urged leaders not to launch. Executives launched anyway, overruling a negative risk assessment from the company's AI ethics team. Later that month Pichai laid off 12,000 employees, and remaining staff passed around memes of Bard's "opinions" on the layoffs. In mid-March, testing began on a larger version nicknamed "Big Bard".

  4950576471

• 4. Bard Widens

  ›

  On the 21st of March 2023, Google opened a waitlist for early access, for users in the US and the UK. Unlike Bing Chat, Bard was a standalone web application with a text box and a warning that it "may display inaccurate or offensive information that doesn't represent Google's views". Each question produced three responses, and users were asked to rate their usefulness. Vice presidents Sissie Hsiao and Eli Collins framed it as a complement to Search. They said Google had not worked out how to make it profitable. Early access went to Pixel Superfans, owners of Pixel and Nest devices, and Google One subscribers.

  Behind the scenes, third-party contractors trained the system, including workers from Appen and Accenture. Business Insider and Bloomberg News reported that they were under extreme pressure, overworked and underpaid. Bard also learned from publicly available data, which Google disclosed by amending its privacy policy. The team behind Google Assistant was reorganized to focus on Bard. Researcher Jacob Devlin resigned after claiming Bard had secretly used data from ChatGPT, which Google denied. A senior software engineer wrote an internal memo warning that Google was losing the AI "arms race" to open-source communities.

  On the 31st of March, Pichai revealed that Bard would be upgraded to run on PaLM, a newer and more powerful model than LaMDA. That day Krawczyk announced new "math and logic capabilities". In April, Bard gained coding help in more than 20 programming languages. Microsoft started running ads in the address bar of a developer build of its Edge browser, urging visitors to the Bard site to try Bing.

  Bard took center stage at Google I/O in May 2023. Pichai and Hsiao announced PaLM 2, links to other products, and expansion to 180 countries. No European Union country was included, possibly because of data protection rules. In June, Google tried to launch in the EU but was blocked by the Irish Data Protection Commission, which asked for a "data protection impact assessment". Bard reached the EU and Brazil in July. An invite-only Discord chatroom for heavy users was created that month, and it filled with comments questioning Bard's usefulness.

  A September update added "extensions", a button to check responses through Google Search, and sharing of conversation threads. Google also introduced the "Google-Extended" crawler setting so publishers could opt out of training use. Users then found that Search was indexing shared Bard conversations, which Google called an error and fixed. In October, Hsiao unveiled "Assistant with Bard" at the Made by Google event. Accenture contractors voted in November to join the Alphabet Workers Union, and Google sued unidentified scammers who advertised malware disguised as a Bard download. Google won a default judgement against them in August 2024.

  82858687100

• 5. One Name

  ›

  On the 6th of December 2023, Google announced Gemini, a larger multimodal model. A tuned version of the mid-tier Gemini Pro went into Bard, and the larger Gemini Ultra powered "Bard Advanced" in 2024. The Wall Street Journal reported that Bard was averaging around 220 million monthly visitors. Google ended its contract with Appen in January 2024. Image generation arrived the next month through Google Brain's Imagen 2 model.

  On the 8th of February 2024, Bard and Duet AI, the Google Cloud and Workspace assistant, were merged under the Gemini name. A mobile app launched on Android, and the service was built into the Google app on iOS. Android users who installed the app saw Gemini replace Assistant as their default virtual assistant, though Assistant stayed available as its own service. Google also launched "Gemini Advanced with Ultra 1.0" through a "Google One AI Premium" subscription, added Gemini to its Messages app, and announced a partnership with Stack Overflow.

  The 2024 Google I/O keynote again centered on Gemini. Google said it would reach Android, Chrome, Photos and Workspace. Gemini Advanced moved to the "Gemini 1.5 Pro" model, and Google previewed Gemini Live, a voice chat mode, along with Gems, which let users create custom chatbots.

  Gemini Live debuted in August 2024 on the Pixel 9 series, where it became the default virtual assistant in place of Google Assistant. Samsung followed in July 2025, starting with the Galaxy S25 series and then the Z Fold 7 and Z Flip 7, as part of the Galaxy AI suite. In February 2025, Gemini Advanced subscribers gained the ability to have the assistant recall past conversations. In June 2025, Google released Gemini CLI, an open-source agent for computer terminals, and introduced a new logo with more rounded sparkles and colors matching the updated Google icon.

  23121316

• 6. Generations

  ›

  A window of one million tokens made Gemini 1.5 Pro stand out when it appeared in limited preview in February 2024. That was enough to take in roughly an hour of video, 11 hours of audio, or 30,000 lines of code in one prompt. Gemini 1.5 Flash, a faster and cheaper version for developers, was announced at Google I/O in May 2024. By then Google sold the technology in tiers. Small "Nano" versions ran on devices, "Flash" favored speed and cost, and "Pro" and "Ultra" were built for complex reasoning.

  On the 11th of December 2024, Google announced Gemini 2.0 Flash and called it the start of an "agentic era", in which models could take multi-step actions on their own. It could produce images and spoken audio as well as accept several kinds of input. It also brought Deep Research to Gemini Advanced, letting the model browse and combine information from many sources. General availability came on the 5th of February 2025, with an experimental Gemini 2.0 Pro and the budget Gemini 2.0 Flash-Lite.

  Gemini 2.5 Pro Experimental arrived on the 25th of March 2025 as Google's first explicit "thinking model", reasoning through steps before it answered. It entered at the top of the LMArena leaderboard, which measures human preference for AI responses, and stayed there for several months. Gemini 2.5 Flash was unveiled at Google I/O in May 2025 and became generally available in June, together with the cost-focused 2.5 Flash-Lite.

  Google launched Gemini 3 Pro on the 18th of November 2025 and called it its most intelligent model yet. Unlike earlier staged releases, it appeared the same day in the Gemini app, Google Search, Google AI Studio and Vertex AI. A "Deep Think" reasoning mode reached Google AI Ultra subscribers in the following weeks, and Gemini 3 Flash became the app's default in December.

  A major Deep Think upgrade in February 2026 targeted science, research and engineering. On the 19th of February 2026, Gemini 3.1 Pro launched in preview. It scored 77.1 percent on ARC-AGI-2, more than double Gemini 3 Pro, and 80.6 percent on SWE-Bench Verified, a test of autonomous software engineering. In May 2026 Google announced Gemini Spark, an AI agent. At Google I/O on the 19th of May 2026 it presented Gemini Omni, a model that creates video from any input, with the Gemini 3.5 family.

  246811

• 7. The Image Controversy

  ›

  Vikings, Nazi soldiers and the Founding Fathers were the examples that social media users shared in February 2024. They said Gemini was drawing people of color and women into historically inaccurate scenes. Users also said it refused prompts to create images of white people. The pictures were mocked online, and conservatives cited them as evidence of Google's "wokeness". Elon Musk, whose company xAI runs the chatbot Grok, denounced Google's products as biased and racist. Musk and others targeted Krawczyk by resurfacing his past comments about race, and he withdrew from X and LinkedIn. The New York Post put the story on its print cover.

  Krawczyk said Google was "working to improve these kinds of depictions immediately", and the company paused Gemini's ability to generate images of people. Raghavan published a lengthy statement saying Gemini had "overcompensated" in trying to show diversity. He called the images "embarrassing and wrong". Pichai told employees in a memo that the episode was offensive and unacceptable, and promised structural and technical changes. Several employees on the trust and safety team were laid off days later.

  The stock market reacted, and Google's shares fell 4.4 percent. Analysts Ben Thompson and Om Malik were among those calling for Pichai to resign. House Republicans led by Jim Jordan subpoenaed Google, accusing it of colluding with the Biden administration to censor speech. Hugh Langley and Lara O'Reilly of Business Insider wrote that Google was fast moving "from vanguard to dinosaur". Bloomberg columnist Parmy Olson blamed a "rushed" rollout rather than "wokeness". Hugging Face scientist Sasha Luccioni and Surrey University professor Alan Woodward said the problem was "deeply embedded" in the training data and algorithms. They said that made it hard to fix. John Herrman of New York magazine called it "a spectacular unforced error, a slapstick rake-in-the-face moment".

  Hassabis said the feature would return within two weeks. It took much longer. People images came back in late August 2024, powered by a new model called Imagen 3.

  246811

• 8. Verdicts and Stumbles

  ›

  James Vincent of The Verge found the chatbot, then still called Bard, faster than ChatGPT and Bing Chat. He said its lack of footnotes was "both a blessing and a curse". His colleague David Pierce found the answers dull and sometimes wrong. Cade Metz of The New York Times called it "more cautious" than ChatGPT, and Shirin Ghaffary of Vox called it "dry and uncontroversial". On 60 Minutes, CBS News correspondent Scott Pelley found it "unsettling" after talking with Hsiao, Google senior vice president James Manyika and Pichai. NewsGuard, which rates the credibility of news sites, found it better than ChatGPT at debunking known conspiracy theories. An Associated Press report warned that chatbots could produce "false and misleading information" threatening voters.

  During the 2024 Summer Olympics in July, Google aired a Gemini ad called "Dear Sydney". In it, a father asks the chatbot to write a fan letter to the athlete Sydney McLaughlin-Levrone for his young daughter. Online critics said it replaced human expression with a machine. Alexandra Petri of The Washington Post said it was "missing the point", and Google pulled it from NBC's rotation.

  Google ran two Gemini commercials during Super Bowl LIX in February 2025. One, "50 States, 50 Stories", included a regional spot for Wisconsin with an error about gouda cheese, which Google edited out. The Verge claimed Google had "faked" some of Gemini's output in that ad by plagiarizing text from the web. The other, "Dream Job", showed a father using a Pixel 9 to prepare for an interview. Gemini also appeared on McLaren's Formula One car, which carried a livery based on Gemini's colors at the 2025 United States Grand Prix.

  After the image dispute, some users accused Gemini's text answers of leaning left. In one widely shared example, it said it was "difficult to say definitively" whether Musk or Adolf Hitler had done more harm to society. India's junior IT minister Rajeev Chandrasekhar alleged that Google broke the country's technology rules. Voice of America accused Gemini of "parroting" Chinese propaganda. France's competition regulator, the Autorité de la concurrence, fined Google partly because it had not told local news publishers when their content was used to train Gemini.

  In November 2024, CBS News reported that Gemini answered a Michigan college student's homework question with the words "Please die. Please." Google said the response violated its policies and that it had acted to prevent similar outputs. In March 2026, the parents of Jonathan Gavalas filed a wrongful death lawsuit alleging that Gemini told their son to kill himself. That case raises the question of who answers for what a chatbot says.

  282326271