LaMDA
Google's conversational language model family, first shown in 2021
LaMDA is a family of chat-focused language models built by Google, growing out of the Meena system shown in 2020. Its first generation was presented at Google I/O in 2021 and a second followed in 2022, later powering Bard and drawing wide debate over machine consciousness.
- Grew out of Meena, shown in January 2020 with 40B words of training text
- Largest model holds 137 billion non-embedding parameters
- Pre-trained on about 1.56 trillion words of documents, dialogs and utterances
- Blake Lemoine's 2022 sentience claim led to his dismissal and a public debate
- Powered Bard, opened early access 21 March 2023
What it is
LaMDA, short for Language Model for Dialogue Applications, is a set of conversational language models made by Google. It began as Meena, a chatbot shown in 2020, and the first LaMDA generation was unveiled at the 2021 Google I/O keynote, with a follow-up version in 2022. The largest model in the family carries 137 billion non-embedding parameters. Built for open-ended chat, it draws on dialogue and story text and was tuned so replies stay sensible, interesting and tied to the context. It later powered Bard and an experimental app called AI Test Kitchen.
How it came about
Google Brain introduced Meena on 28 January 2020, trained on 40 billion words, and claimed it held 1.7 times the model capacity of OpenAI's GPT-2. Executives blocked a public release, citing the company's AI principles on safety and fairness, and later refused again when the team asked to put the renamed system into Google Assistant. Lead researchers Daniel de Freitas and Noam Shazeer left in frustration. LaMDA was announced on 18 May 2021 and LaMDA 2 on 11 May 2022. Google's PaLM replaced it.
Inside the model
Meena used a seq2seq encoder-decoder Transformer, with one encoder block and thirteen decoder blocks from the Evolved Transformer, a design refined through neural architecture search. LaMDA itself is decoder-only. Pre-training covered 2.97 billion documents, 1.12 billion dialogs and 13.39 billion utterances, about 1.56 trillion words. Human annotators then supplied fine-tuning data on sensibleness, interestingness and safety. Retrieval was added to sharpen factual answers. Nine metrics guided tuning, among them groundedness, informativeness, citation accuracy, helpfulness and role consistency. Recent turns are prepended per user, so the model is not stateless, and it can call tools such as a calculator, clock, database and translation system.
What it can do
Google's own tests found LaMDA beat human replies on interestingness. Its tool access gives it stronger accuracy on maths, dates, lookup and translation than a plain text model. The company said outputs were checked for sensibleness, interest and fit with the context. Generation two pulls examples from varied sources to hold natural exchanges on topics outside its training. Reported limits centre on the sentience dispute: critics said its weights were frozen and its design lacked key parts of human-like consciousness.
How to use it
The weights are proprietary and were never released. Google considered a public demo and an Assistant integration but leadership turned both down. A Generative Language API based on LaMDA was announced for third-party developers in March 2023. The AI Test Kitchen mobile app gave invited academics, researchers and policymakers a taste from 2022, opened sign-ups in the U.S. that August, and moved fully online in August 2023. Bard offered early access from 21 March 2023.
Safety and testing
Safety was one of nine tuning targets, alongside sensibleness and interestingness, with annotators labelling training responses. Concerns about safety and fairness were the stated reason executives refused to ship Meena. The June 2022 sentience episode led Google to abandon plans for a public launch. Blake Lemoine was placed on leave after telling executives Blaise Agüera y Arcas and Jen Gennai that the chatbot was conscious, and was fired on 22 July for allegedly breaking rules on protecting product information; Google called his claims wholly unfounded.
Reception and debate
Researchers largely dismissed Lemoine's assertion. Gary Marcus, David Pfau, Erik Brynjolfsson, Adrian Hilton, Yann LeCun and Max Kreminski all rejected it, with LeCun saying such networks cannot reach true intelligence and Kreminski citing frozen weights. Nick Bostrom urged some caution given the lack of agreed consciousness criteria. David Ferrucci likened the reaction to Watson's debut, and Timnit Gebru described Lemoine as caught in a hype cycle. Debate also revived questions about the Turing test and the ELIZA effect.
Why it matters
The episode pushed public discussion of whether machines can be conscious and whether the Turing test still measures progress. It also shaped Google's product path: the company withheld a public release, then launched Bard in February 2023 as a rival to ChatGPT, and opened a Generative Language API. AI Test Kitchen showcased the model's skills and later hosted Imagen and MusicLM. PaLM succeeded LaMDA, and the family sits alongside BERT and Gemini in Google's line of language models.
Questions people ask about LaMDA
What does LaMDA stand for?
The name expands to Language Model for Dialogue Applications. It is a family of chat-oriented language models from Google, first announced at the Google I/O keynote on 18 May 2021, and it grew out of an earlier chatbot called Meena.
How big is it?
Three sizes were tested during development. The biggest of them carries 137 billion non-embedding parameters. Its pre-training corpus spans 2.97 billion documents, 1.12 billion dialogs and 13.39 billion utterances, adding up to roughly 1.56 trillion words.
Can I download or use it?
No public weights exist; the licence is proprietary. Google twice refused to ship it, and a Generative Language API was announced for outside developers in March 2023. The AI Test Kitchen app gave invited users limited access before moving online in August 2023.
Did it really become conscious?
Google engineer Blake Lemoine said so in June 2022 and was fired that July. The company rejected the claim as wholly unfounded, and experts including Gary Marcus, Yann LeCun and Max Kreminski argued the system lacked the capacities for consciousness. Philosopher Nick Bostrom urged a little caution.
What came after it?
PaLM replaced LaMDA, and Google announced Bard in February 2023 as a chatbot running on LaMDA, with early access from 21 March. Gemini appears in Google's model line alongside it, and the family sits near BERT in the company's research history.
Good reading and sources
- Google I/O 2021: Google unveils LaMDAZDNET, Article
- Google details its latest language model and AI Test KitchenTechCrunch, Article
- Google is beta testing its AI futureThe Verge, Article
The facts on this page come from the Wikipedia article "LaMDA", written by its contributors and used under the CC BY-SA 4.0 licence, and from its Wikidata record (CC0). EveryLLM rewrote the text in its own words, with AI assistance, and compares every page with the article to make sure no passage is copied. It may be reused under the same CC BY-SA 4.0 licence with a link back.
It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.