Bard
Last reviewed
May 17, 2026
Sources
No citations yet
Review status
Needs citations
Revision
v6 ยท 4,289 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
May 17, 2026
Sources
No citations yet
Review status
Needs citations
Revision
v6 ยท 4,289 words
Add missing citations, update stale details, or suggest a clearer explanation.
Bard was the conversational AI chatbot developed by Google, launched as an experimental service on February 6, 2023, opened to a public waitlist on March 21, 2023, and rebranded to Gemini on February 8, 2024.[^1][^2][^3] Built initially on a lightweight version of Google's LaMDA (Language Model for Dialogue Applications), Bard was Google's direct response to OpenAI's ChatGPT, which had become a cultural phenomenon after its November 2022 release.[^4] During its roughly twelve months under the Bard name, the chatbot cycled through three different model backbones (LaMDA, PaLM 2, and Gemini Pro), expanded from two countries to more than 230 countries and territories in over 40 languages, picked up integrations with Gmail, Drive, Maps, and YouTube, and produced two of the most-discussed LLM controversies of the early generative AI era: a botched launch demo that erased roughly $100 billion in Alphabet's market value, and an image generation system that depicted historical figures in racially incongruent ways.[^5][^6][^7]
Bard never fully recovered from its first impression. By the time Google retired the name in 2024, the product was widely seen as the slower, more cautious of the major chatbots, even as the underlying models caught up to and in some benchmarks surpassed competitors. The retirement was less a death than a renaming: the chat interface, the URL, the team, and the product lead's remit all carried over into Gemini.[^7]
The immediate trigger for Bard was ChatGPT. OpenAI released the chatbot on November 30, 2022, and within two months it had reached 100 million monthly active users, the fastest consumer software adoption ever recorded at the time.[^4] The New York Times reported in December 2022 that Sundar Pichai had declared a "code red," reassigning teams across Google to accelerate AI products.[^4] Pichai later pushed back on the characterization during the March 31, 2023 episode of the Times' Hard Fork podcast, saying he had not personally issued a code-red declaration, although he conceded that the company had reorganized around the new competitive reality.[^17] Whether or not the phrase originated with Pichai, it captured what was happening internally.
Larry Page and Sergey Brin, both retired from operational roles since 2019, returned for a series of emergency meetings with executives in January 2023. Brin requested access to Google's source code repository in February 2023 for the first time in years and would later co-author a small Gemini-era bug fix that he highlighted publicly.[^9] The internal effort to ship a ChatGPT competitor was given the codename Project Atlas. A specific test product known as Apprentice Bard, built on a fine-tuned version of LaMDA, was distributed to Google employees in late January 2023 and previewed the conversational format that the public product would eventually take.[^18]
The conflict at Google was partly cultural. The company had invented the transformer architecture in the 2017 paper "Attention Is All You Need" and had trained large language models internally for years, including LaMDA. But it had repeatedly declined to release them, citing safety, factuality, and reputational risk. Within weeks of ChatGPT's launch, the same caution that had kept Google out of the consumer chatbot market began to look like a strategic error, and the pressure to ship reversed several years of public positioning.
The name Bard is a reference to the medieval term for a poet or storyteller, an allusion Google never officially confirmed but which Pichai's launch post strongly implied by framing the product around creativity, learning, and curiosity.[^1][^8] Where ChatGPT signaled a generic chat tool and Microsoft's Bing Chat leaned on a search engine brand, Bard reached for a literary association meant to set expectations for prose and creative writing rather than raw question answering. The internal codename Atlas, by contrast, gestured at the scope of the engineering effort rather than the user-facing experience.[^18]
LaMDA, the model that would power the first version of Bard, was first publicly demonstrated by Pichai at Google I/O on May 18, 2021, with a conversation in which the model spoke as the planet Pluto and as a paper airplane.[^9] The largest LaMDA contained 137 billion parameters and had been trained on roughly 1.56 trillion words of public dialog data and web text.[^9] Despite the demo, Google declined to ship a consumer chatbot in 2021 or 2022, with executives publicly emphasizing the risk of factual errors and harmful outputs.
The LaMDA name re-entered public consciousness in June 2022 when Blake Lemoine, a senior software engineer in Google's Responsible AI division, told The Washington Post that the model had become sentient based on his conversations with it about self-identity, religion, and Isaac Asimov's Three Laws of Robotics.[^10] Lemoine was placed on paid administrative leave on June 11, 2022, and fired on July 22, 2022, after he hired an attorney to represent the model and shared internal documents publicly.[^10][^11] The mainstream AI research community, including Gary Marcus and Yann LeCun, dismissed the claims, but the episode lodged LaMDA in public memory as the system that had supposedly fooled an engineer into thinking it was a person.[^10]
On February 6, 2023, Pichai published a blog post titled "An important next step on our AI journey" announcing Bard as "an experimental conversational AI service, powered by LaMDA."[^1] The post said Bard would first be opened to trusted testers and then made available to the public "in the coming weeks," framing the product as a way to combine Google's search infrastructure with the response style of a large language model.[^1] Pichai's example prompt asked Bard to explain new discoveries from the James Webb Space Telescope to a 9-year-old, a choice that would become unfortunate within 48 hours. Microsoft held its own press event on February 7 to unveil its OpenAI-powered Bing Chat, and Google followed on February 8 with the "Live from Paris" livestream pitched as a broader showcase of AI in Search and Maps.[^4][^5]
In the marketing GIF embedded in Pichai's blog post and in subsequent promotional material, Bard answered the question "What new discoveries from the James Webb Space Telescope can I tell my 9-year-old about?" One bullet claimed that the telescope had "taken the very first pictures of a planet outside of our own solar system."[^5][^12] That claim was incorrect. The first direct image of an exoplanet was captured in 2004 by the European Southern Observatory's Very Large Telescope, nearly two decades before Webb began operating. Astronomers, including Grant Tremblay of the Center for Astrophysics, flagged the error on Twitter within hours.[^12]
Reuters reported the error on the morning of February 8, 2023. By the close of trading that day, Alphabet's shares had fallen roughly 7.7%, erasing approximately $100 billion in market capitalization.[^5][^6] Microsoft shares rose about 3% during the same window.[^5] Google said in a statement that the error "highlights the importance of a rigorous testing process," referring to the trusted tester program that had only just begun.[^5][^6] Internal messages later reported by Bloomberg captured employee anger, with one engineer writing "Bard is worse than useless: please do not launch" and another calling the system a "pathological liar."[^13] The error confirmed a narrative that had been building for months: Google, despite having published the foundational transformer research, was now shipping a product before it was ready, and the hallucination it had been so cautious about had appeared on opening night.
The episode also reset the press script. For the rest of 2023, virtually every story about AI quality, accuracy, or model error referenced the James Webb mistake as a touchpoint. It became shorthand for LLM overreach in general, cited routinely in coverage of Microsoft, Meta, and even unrelated Google products.
Bard opened to a public waitlist on March 21, 2023, in the United States and the United Kingdom only, after a closed period with roughly 10,000 trusted testers.[^2][^14] The web interface lived at bard.google.com and required a personal Google account; Workspace accounts and users in the European Union were excluded. The initial version still ran on a lightweight LaMDA, which limited reasoning, code generation, and multilingual coverage compared to ChatGPT.[^17]
The user experience was deliberately minimal: a prompt box, a response area, and a "View other drafts" control that exposed two alternate completions for each query. The Drafts feature was unique among contemporary chatbots and was Google's attempt to surface model uncertainty by showing the user multiple plausible outputs rather than a single confident answer.[^19] The interface also carried a prominent disclaimer that Bard "may display inaccurate or offensive information that doesn't represent Google's views," and Pichai personally pinned a thread of his own internal feedback emphasizing that the system was an experiment, not a search replacement.[^14]
The early access pool skewed toward existing Google users. Pixel Superfans, Pixel and Nest device owners, and Google One subscribers received priority invitations.[^16] The product offered text generation, basic summarization, brainstorming, and short-form creative writing. On March 31, 2023, Google quietly added math and logic capabilities, and in early April it added coding assistance across more than 20 programming languages, although reviewers noted that quality lagged GPT-4 in head-to-head comparisons.[^17] Kevin Roose of The New York Times called Bard during this period "cautious to a fault," writing that the model frequently refused requests that ChatGPT would attempt and that its outputs felt flatter, more reluctant, and more obviously hedged.[^28]
At Google I/O on May 10, 2023, Pichai announced that Bard would move to PaLM 2, a much larger model trained on a wider range of text including scientific papers and code in more than 20 programming languages.[^15][^16] The model came in four sizes (Gecko, Otter, Bison, and Unicorn), and the Bard version sat near the top of that range. The same announcement removed the waitlist and opened Bard to over 180 countries and territories, added Japanese and Korean as the first non-English languages, and previewed integrations with Adobe Firefly for image generation, Google Lens for image input, source citations for code blocks, a dark theme, and one-click export to Replit (Python), Gmail, and Google Docs.[^15][^16] Third-party extensions were previewed with partners including Kayak, OpenTable, ZipRecruiter, Instacart, Wolfram, and Khan Academy.[^15]
Reviewers noted improvements in coding, multilingual translation, and reasoning, although the gap to GPT-4 (released by OpenAI in March 2023) remained visible in head-to-head tests.[^16][^17] PaLM 2 closed the most embarrassing gaps in fluency and factual recall, but it did not produce a clear "Bard moment" that flipped public sentiment back toward Google.
The European Union launch was a separate problem. Google postponed the EU rollout in June 2023 after the Irish Data Protection Commission, Google's lead GDPR regulator, said the company had not provided a data protection impact assessment. Bard finally launched in the EU and Brazil on July 13, 2023, with expanded language support to 40 languages including Arabic, Mandarin, German, Hindi, Spanish, and Romanian.[^18]
On September 19, 2023, Google launched Bard Extensions in English, the most ambitious feature update under the Bard name.[^20] Extensions let Bard pull data from a connected user's Gmail, Google Drive, Google Docs, Google Maps, YouTube, and Google Flights and Hotels in a single conversation. The launch example was a trip-planning prompt that asked Bard to find dates from a user's Gmail thread, look up flights and hotels for those dates, generate Maps directions to the airport, and pull YouTube videos for the destination.[^20]
Google paired the launch with two other features. A "Google it" button (sometimes referred to in press coverage as the "double-check" feature) re-ran Bard's response through Google Search and highlighted phrases the search engine could or could not corroborate; phrases were tinted green when supporting search results were found and orange when results conflicted or were absent. A shared-conversation link let users publish a transcript to a public URL, enabling the kind of forwarding behavior that had been viral for ChatGPT but was not natively supported there at the time.[^20] Workspace data accessed through Extensions would not be used to train the model or shown to human reviewers.[^20]
Extensions was Google's clearest articulation of what made its chatbot different from ChatGPT: it sat on top of the user's own data graph as well as the public web. The pattern was later echoed by ChatGPT's own GPT Builder and connectors, by Microsoft Copilot's integration into Microsoft 365, and by Apple Intelligence's on-device personal context layer announced in mid-2024.
Google DeepMind announced the Gemini family of models on December 6, 2023, in three sizes (Nano, Pro, and Ultra) and described them as natively multimodal across text, code, images, audio, and video.[^21] On the same day, Bard was upgraded to a fine-tuned version of Gemini Pro for English-language users in more than 170 countries.[^21] Google reported that Gemini Pro outperformed GPT-3.5 on six of eight standard benchmarks, including MMLU and GSM8K, although the larger Gemini Ultra was still in safety testing and not yet shipping.[^21] By February 1, 2024, the Gemini Pro upgrade had rolled out globally, including the European Union, in more than 40 languages.[^22]
The Bard interface stayed the same; the change was under the hood. The same prompt window, the same Drafts feature, the same Extensions, but with a stronger foundation model behind them. The December update is also where Bard's identity began to dissolve. The product was now called Bard but was openly powered by Gemini, a separation that was confusing enough to make the rebrand almost inevitable.
The Gemini launch also reset Google's research org chart. Demis Hassabis, who had been folded into the merged Google DeepMind unit in April 2023, took on visible product responsibility alongside Pichai for the first time, and the Gemini models were positioned as Google's flagship across consumer, enterprise, and developer surfaces in a way that LaMDA and PaLM 2 had never quite been.[^30]
On February 8, 2024, exactly one year after the Paris livestream that contained the James Webb error, Google rebranded Bard to Gemini.[^3][^7] The same day, the company launched Gemini Advanced, a paid tier built on Gemini Ultra 1.0 and sold through Google One AI Premium for $19.99 per month with a two-month free trial.[^7][^23] The plan bundled Gemini Advanced with 2 TB of Google One storage and the existing premium Google One features, positioning it as a Workspace-plus subscription rather than a pure chatbot upgrade.[^7] The free tier, still backed by Gemini Pro, kept the same web interface at gemini.google.com and absorbed the Bard URL through a redirect. A standalone Gemini app for Android shipped on February 8 and replaced Google Assistant on supported devices; iOS users initially got access through the Google app, with a dedicated Gemini iOS app following later in 2024.[^7]
Sissie Hsiao, the vice president and general manager who had led Bard from launch, wrote in the announcement that Google wanted users to understand they were talking directly to the underlying model rather than a separately branded chat layer.[^7] Jack Krawczyk, the senior product director who had been the public face of Bard since 2023, continued in his product role under the Gemini brand. The rebrand also let Google retire two awkward sub-brands at once: Bard, which carried the baggage of the launch fiasco, and Duet AI, the workspace assistant that had never quite landed.[^7]
Gemini Advanced added features that had been held back from the free Bard tier: a longer context window (initially 32,768 tokens, soon expanded much further with the Gemini 1.5 launch), the ability to upload and analyze documents and images, code generation with a built-in Python sandbox, and priority access to new capabilities like deeper Workspace integration. Gemini Advanced subscribers were also the first to receive Imagen 2-powered human image generation, the feature that would produce the next major controversy within two weeks of launch.[^7][^23]
The rebrand honeymoon lasted about two weeks. In late February 2024, Gemini's newly added human image generation feature (powered by Imagen 2) produced widely shared outputs depicting historical figures in racially and gender-incongruent ways: Black founding fathers, Asian Wehrmacht soldiers in 1943 German uniforms, female popes, and racially diverse Vikings.[^24][^25] The model also refused several prompts asking specifically for white people while accepting analogous prompts for other groups.[^24][^26] Screenshots circulated on X and triggered political backlash plus broader criticism that the system was applying demographic diversification heuristics without context.
On February 22, 2024, Google paused human image generation in Gemini.[^26] Pichai sent an internal memo on February 27, 2024 calling the outputs "completely unacceptable" and saying Google had "got it wrong."[^27] Co-founder Sergey Brin told an audience at a March 2024 hackathon that the company had "definitely messed up on the image generation."[^28] Krawczyk deleted his X account after his older posts were dredged up.[^13] Image generation of people remained suspended for roughly six months and was reintroduced in August 2024 on the upgraded Imagen 3 model with narrower guardrails and a default that favored prompt fidelity over forced demographic variation.[^25]
The episode prompted a broader argument about how to align image generators across users with different priors, and it was a particularly visible example of the trade-offs that reinforcement learning from human feedback and demographic post-processing can introduce. It also became a case study for prompt engineering and red-teaming teams across the industry on how interventions intended to address representation in one direction can produce conspicuous failures in another.
During the Bard era, the chatbot competed mainly with OpenAI's ChatGPT and Microsoft's Bing Chat (later renamed Copilot). Anthropic's Claude became publicly available in mid-2023 as a smaller third entrant during this period.
| Feature (late 2023) | Bard | ChatGPT | Bing Chat |
|---|---|---|---|
| Consumer launch | Mar 21, 2023 (waitlist) | Nov 30, 2022 | Feb 7, 2023 |
| Underlying model | Gemini Pro | GPT-3.5 (free) / GPT-4 (paid) | GPT-4 + Prometheus |
| Real-time web data | Yes | No on free tier | Yes |
| Image generation | Imagen 2 (Feb 2024) | DALL-E 2/3 (paid) | DALL-E 3 |
| Image input | Google Lens (May 2023) | GPT-4V (Sept 2023, paid) | Yes |
| Inline citations | Limited; "Google it" button | None in 2023 | Yes |
| Paid tier | None until Feb 2024 ($19.99/mo Advanced) | $20/mo Plus | Free with Edge |
| Standalone mobile app | Feb 2024 | Nov 2023 iOS, Jul 2023 Android | Bing app |
| Peak reach | 230+ territories, 40+ languages | 160+ countries | 169 countries |
Reviewers in 2023 generally placed Bard third among the three. James Vincent of The Verge called Bard "faster than ChatGPT" but flatter and less useful for sustained tasks; Kevin Roose of The New York Times found ChatGPT "vastly superior" as a writing assistant during the LaMDA era.[^29] After the PaLM 2 and Gemini Pro upgrades, the gap closed: NewsGuard reported in 2024 that Gemini outperformed ChatGPT at debunking conspiracy theories.[^30] On coding, GPT-4 retained a clear lead through most of the Bard period, although Gemini's tighter integration with Google's stack made it more useful for users who lived inside Gmail and Drive.
Bard's reception was a function of expectations more than capabilities. Because it came from Google, which had defined modern web search and had published the foundational transformer paper, anything short of clearly beating ChatGPT registered as a disappointment.[^4] Early reviews focused on hallucinations, refusals, and a tone that several writers described as cautious or generic. The James Webb error became shorthand in business and tech press for AI overreach in general, cited routinely throughout 2023 in stories about Microsoft, Meta, and even other Google features.[^28][^29]
Internally, the launch produced a measurable shift. The "code red" reorganization that had begun in late 2022 accelerated through 2023, with Pichai folding Google Brain and DeepMind into a single Google DeepMind unit in April 2023 and elevating Demis Hassabis as its CEO.[^31] The Bard product team grew rapidly under Hsiao, who led the chatbot through both the rebrand and the image generation crisis before stepping down in April 2025 in favor of Josh Woodward, who had led NotebookLM.[^32] In October 2024, the Gemini app team itself was reorganized under DeepMind, completing the migration from a product-marketing surface to a research-led one.[^32]
External impact extended beyond Google. The James Webb error and the Imagen 2 controversy together produced two of the most-cited examples of AI failure in the early generative era, and they shaped how investors, regulators, and competitors evaluated subsequent AI launches. Microsoft, OpenAI, Meta, and Anthropic all drew lessons from how Google had handled both episodes, and several later launches (including OpenAI's GPT-4o reveal in May 2024) were deliberately staged in ways that avoided the demo-on-stage failure mode that had defined Bard's debut.
Bard exists today only as a redirect. Typing bard.google.com sends users to the Gemini app, and Google's blog posts refer to the 2023 chapter under the Gemini brand. In Hsiao's April 2025 farewell memo, she described her time leading the product as "chapter 1" and the new phase under Woodward as "chapter 2."[^32] The Gemini brand itself has expanded well beyond the chat interface: the family now includes a series of foundation models released over 2024 and 2025 (Gemini 1.5 Pro, 1.5 Flash, 2.0, and successors), a developer API, embedded surfaces inside Search, Workspace, Android, and Chrome, and the Gemini Live voice-chat mode that previewed at Google I/O in May 2024 and shipped as the default assistant on the Pixel 9 series in August 2024.
Bard's lasting contribution may be the precedent for tying a chatbot to a personal data graph through Extensions, a pattern ChatGPT later echoed with its own connectors and Microsoft pursued through Copilot for Microsoft 365. Conceptually, the Bard era established that a chatbot's value to a user could come less from raw model strength than from the data and surfaces it could reach on the user's behalf; that framing has since become a standard part of how every major AI assistant is positioned. The product also accelerated the broader industry race that drew Anthropic, Microsoft, Meta, and a long tail of open-weight model developers into a sustained competition over reasoning, multimodality, agentic behavior, and long-context understanding that defined the rest of the decade.