Mountain View, California · 18 May 2021 — During the Google I/O 2021 keynote on 18 May 2021, Google CEO Sundar Pichai unveiled LaMDA (Language Model for Dialogue Applications), a conversational large language model designed to engage in open-ended dialogue on virtually any topic. Unlike task-specific chatbots that handle narrow, predefined interactions — such as booking a flight or checking the weather — LaMDA was built to sustain free-flowing conversations that could shift naturally between subjects, follow logical threads, and maintain contextual coherence over multiple turns.
The announcement positioned Google in the emerging competition over conversational AI, a field that was rapidly gaining attention following the release of OpenAI's GPT-3 in June 2020. While GPT-3 was a general-purpose language model capable of dialogue among many other tasks, LaMDA was specifically trained and optimised for conversation.
From Meena to LaMDA
LaMDA's lineage traced back to Meena, a 2.6-billion-parameter chatbot unveiled by Google on 28 January 2020. Meena was developed by the Google Brain research team, which hoped to release it publicly in a limited capacity. However, corporate executives declined, citing concerns that Meena violated Google's AI principles around safety and fairness.
As Meena's training data and computing power were scaled up, it was renamed LaMDA. The Google Brain team again sought to deploy it to Google Assistant and open a public demo, but leadership once more denied the request. LaMDA's two lead researchers, Daniel de Freitas and Noam Shazeer, eventually left Google in frustration — though their departure was not publicly known at the time of the I/O announcement.
Architecture and Training
LaMDA is a decoder-only transformer language model, built on the seq2seq architecture and transformer-based neural networks that Google Research introduced in 2017. The model was pre-trained on a massive corpus of text that included both documents and dialogue data, then fine-tuned using manually annotated responses evaluated on multiple quality dimensions.
Google identified three core quality metrics for LaMDA responses:
- Sensibleness: Does the response make sense in context?
- Specificity: Is the response specific to the conversation, not a generic reply?
- Interestingness: Is the response engaging and likely to keep the conversation going?
In addition to these three, LaMDA was tuned on six further metrics: safety, groundedness, informativeness, citation accuracy, helpfulness, and role consistency.
Dual-Process Architecture
LaMDA incorporated several design features that distinguished it from standard language models. The system had access to external tools — including a database, a real-time clock and calendar, a mathematical calculator, and a natural language translation system — making it one of the first "dual-process" chatbots. This tool-augmented approach, later known as retrieval-augmented generation, improved factual accuracy by grounding responses in external data rather than relying solely on the model's internal knowledge.
Additionally, LaMDA was not stateless. Its sensibleness metric was fine-tuned by "pre-conditioning" each dialogue turn with many of the most recent dialogue interactions on a per-user basis, allowing the model to maintain conversational context across extended exchanges.
The Google I/O Demonstration
At the I/O keynote, Pichai demonstrated LaMDA by having it converse as if it were the dwarf planet Pluto and later as if it were a paper airplane. In the Pluto demonstration, LaMDA responded to questions about its characteristics, its relationship with other celestial bodies, and the New Horizons flyby with contextually appropriate, conversational responses that maintained the persona throughout multiple turns.
The demonstrations highlighted LaMDA's ability to stay in character, provide factual information within a conversational framework, and handle topic transitions naturally. However, Google noted that LaMDA was still in the research and development phase and was not yet integrated into any Google product.
Safety and Responsible AI
Google emphasised safety as a central design consideration. The company stated that LaMDA's responses had been ensured to be "sensible, interesting, and specific to the context" while also meeting safety standards. The model was trained with safety as one of its nine tuning metrics, and Google invested in adversarial testing to identify and mitigate harmful outputs.
This emphasis on safety reflected broader industry concerns about the risks of large language models, including the generation of misinformation, biased or toxic content, and the potential for misuse in social engineering. Google's caution in deploying LaMDA — declining to release it publicly at the time of announcement — contrasted with OpenAI's more aggressive API rollout for GPT-3.
Industry Context and Competitive Landscape
LaMDA was announced at a time of intensifying competition in large language models. OpenAI's GPT-3, released in June 2020, had demonstrated that massive scale could yield emergent capabilities across diverse language tasks. Microsoft had exclusively licensed GPT-3 in September 2020. Other players, including DeepMind and Meta (then Facebook), were investing heavily in language model research.
Google's decision to focus specifically on dialogue — rather than general-purpose text generation — reflected a strategic bet that conversational interfaces would become a primary mode of human-computer interaction. The company's search engine, which processed billions of queries daily, was a natural candidate for conversational enhancement, and LaMDA was widely interpreted as a step toward making search more interactive.
Legacy
LaMDA would not be publicly released in 2021, but its announcement established Google's position in the conversational AI race. The second generation, LaMDA 2, was unveiled at Google I/O 2022 alongside the AI Test Kitchen app, which allowed limited public interaction with the model.
In February 2023, following the viral success of OpenAI's ChatGPT, Google announced Bard — a conversational AI chatbot powered by LaMDA — in what was widely seen as a rapid response to competitive pressure. Bard's initial demo, however, produced a factual error about the James Webb Space Telescope, contributing to a $100 billion drop in Alphabet's market value and underscoring the ongoing challenges of factual accuracy in large language models.
The LaMDA project's trajectory — from internal research to cautious public preview to competitive product launch — illustrated the complex interplay between AI capability, safety considerations, and commercial pressure that would define the large language model era.
Sources
- Google Blog, "LaMDA: our breakthrough conversation technology," 18 May 2021
- Wikipedia, "LaMDA,"
- ZDNet, "Google I/O 2021: Google unveils LaMDA," 18 May 2021
- Wired, "Google Hopes AI Can Turn Search Into a Conversation," 2021
- 9to5Google, "Google 'LaMDA' is the next breakthrough in AI understanding," 18 May 2021
FIRAT Editorial Board
Institutional Research Desk · Foresight Institute of Research and Translation
The collective editorial and research translation board of FIRAT, synthesising peer-reviewed evidence, policy briefs, and division milestones across our seven foundational research pillars.
