# LaMDA

> Source: https://aiwiki.ai/wiki/lamda
> Updated: 2026-06-21
> Categories: Conversational AI, Google, Large Language Models
> From AI Wiki (https://aiwiki.ai), a free encyclopedia of artificial intelligence. Quote with attribution.

*See also: [Bard](/wiki/bard), [Gemini](/wiki/gemini), [Google](/wiki/google), [Large language model](/wiki/large_language_model), [Meena](/wiki/meena)*

**LaMDA** (short for **Language Model for Dialogue Applications**) is a family of conversational [large language models](/wiki/large_language_model) developed by [Google](/wiki/google), built on the [Transformer](/wiki/transformer) architecture and fine-tuned for open-ended dialogue rather than for generic text generation. The largest version has 137 billion parameters and was pre-trained on 1.56 trillion words of public dialog data and web text. LaMDA was first announced at [Google I/O](/wiki/google_io) 2021 on May 18, 2021, described in detail in a January 2022 research paper, and later served as the original model behind Google's [Bard](/wiki/bard) chatbot. The paper defines LaMDA as "Transformer-based neural language models specialized for dialog, which have up to 137B parameters and are pre-trained on 1.56T words of public dialog data and web text."[^1][^2]

LaMDA gained widespread public attention in June 2022 when Blake Lemoine, an engineer at Google's Responsible AI organization, publicly claimed that the model had become sentient. Google rejected the claim and later dismissed Lemoine. The story became a flashpoint in the broader debate about machine consciousness, [AI safety](/wiki/ai_safety), [anthropomorphization](/wiki/anthropomorphization), and how the public should interpret the abilities of [large language models](/wiki/large_language_model).[^3][^4]

LaMDA later served as the original model behind Google's [Bard](/wiki/bard) chatbot, which launched on March 21, 2023. Bard was Google's response to the release of [ChatGPT](/wiki/chatgpt) by [OpenAI](/wiki/openai) in November 2022. Within a few months Google replaced LaMDA inside Bard with the more capable [PaLM 2](/wiki/palm_2) model in May 2023, and Bard itself was rebranded as [Gemini](/wiki/gemini) on February 8, 2024.[^5][^6][^7]

## Overview

LaMDA is a decoder-only [Transformer](/wiki/transformer) language model trained primarily on dialogue data. The largest publicly described version has 137 billion non-embedding parameters and was pre-trained on a corpus of 1.56 trillion words consisting of public dialog data and other public web documents.[^2]

What distinguished LaMDA from many earlier language models was the deliberate focus on conversation as the primary task. Where models like [GPT-3](/wiki/gpt-3) were trained as general-purpose next-token predictors and then prompted into conversational behavior, LaMDA's training pipeline used dialog corpora, dialog-shaped fine-tuning data, and a set of explicit objectives intended to make the model a better chat partner. The published paper organized those objectives under three headings: Quality, Safety, and Groundedness. The paper's central finding was that scale alone was not enough: it reported that "fine-tuning with annotated data and enabling the model to consult external knowledge sources can lead to significant improvements towards the two key challenges of safety and factual grounding."[^2]

LaMDA was never released as a general-purpose API or as a downloadable model. Public access was limited to a small set of demos, an invite-only Android and iOS application called the [AI Test Kitchen](/wiki/ai_test_kitchen), and eventually to the early Bard chatbot. The base research model and its weights remained internal to Google.[^8]

## Background and motivation

### From Meena to LaMDA

LaMDA's direct predecessor was [Meena](/wiki/meena), a 2.6 billion-parameter open-domain chatbot that Google Research introduced on January 28, 2020. Meena was developed by Daniel Adiwardana, Thang Luong, and colleagues on the Google Brain team and described in the paper "Towards a Human-like Open-Domain Chatbot" (arXiv:2001.09977). The model was a sequence-to-sequence Transformer with an Evolved Transformer encoder block, trained end-to-end on 341 GB of public-domain social media conversations.[^24][^25]

Meena made two contributions that carried directly into LaMDA. First, it proposed the Sensibleness and Specificity Average (SSA), a human-evaluation metric for chatbots that asked raters whether each response made sense in context and whether it was specific rather than generic. Meena scored 79% on SSA against an 86% human baseline, comfortably ahead of the next-best existing chatbot. Second, the paper observed a strong correlation between SSA and perplexity (R squared of 0.93), suggesting that improvements in standard language-model objectives would translate into better conversational behavior.[^24][^25]

According to later reporting, Meena was renamed LaMDA as the project grew in compute, data, and scope, and as Google added the fine-tuning and safety machinery that distinguishes the published LaMDA model. Internal disagreements over whether to release a Meena-based chatbot publicly are widely cited as one reason Google was cautious with LaMDA in the years before [ChatGPT](/wiki/chatgpt). Two of the Meena authors, Daniel De Freitas and Noam Shazeer, later left Google and co-founded the rival chatbot start-up Character.AI.[^26]

### Dialogue-first design

By 2020, [large language models](/wiki/large_language_model) such as [GPT-3](/wiki/gpt-3) had shown that scale alone could produce fluent text completion across many domains, but they were not specifically trained on the multi-turn, persona-consistent exchanges that characterize human conversation. The LaMDA project bet that an explicit focus on dialogue as the central task, both during pre-training data selection and during fine-tuning, would produce a model that was better at chat without necessarily being larger. The other half of LaMDA's motivation was alignment: the team wanted a chatbot that was not only fluent but also safe and grounded. The resulting Quality, Safety, and Groundedness objectives made LaMDA one of the earliest industrial efforts to wire concrete behavioral evaluations into the training pipeline of a dialogue-tuned [large language model](/wiki/large_language_model). The approach is sometimes called classifier-based alignment, in contrast with the [reinforcement learning from human feedback](/wiki/rlhf) (RLHF) recipe that [OpenAI](/wiki/openai) used for [ChatGPT](/wiki/chatgpt).[^2][^23]

## Announcement and history

### Google I/O 2021

Google CEO [Sundar Pichai](/wiki/sundar_pichai) introduced LaMDA during the Google I/O 2021 keynote on May 18, 2021. The on-stage demonstration showed LaMDA engaging in open-ended conversation in two personas: it pretended to be the dwarf planet Pluto answering a child's questions, and it pretended to be a paper airplane describing what it had seen during a flight. In the Pluto demo, LaMDA delivered the now widely quoted line, "I wish people knew that I am not just a random ice ball. I am actually a beautiful planet." Pichai presented the model as an early breakthrough in conversational understanding and stated that Google was "just at the beginning" of exploring its capabilities.[^1][^9][^27]

During the same keynote, Pichai described LaMDA as the result of years of research at Google and emphasized that conversation depends on properties such as sensibleness and specificity that traditional language models often miss. He said the model could discuss "a seemingly endless number of topics" and framed the technology as a step toward more natural ways of interacting with information. Notably, Google did not release a public demo, an API, or any model weights at the time. Several outlets pointed to the absence of a public demo as a sign that Google was wary of repeating Microsoft's 2016 [Tay](/wiki/tay) incident, in which a public chatbot quickly began producing offensive output.[^9][^21]

### Google I/O 2022 and LaMDA 2

At Google I/O 2022 on May 11, 2022, Google announced LaMDA 2, an updated version that had been tested internally by thousands of Google employees in an effort to reduce harmful or off-topic outputs. The same announcement introduced the [AI Test Kitchen](/wiki/ai_test_kitchen) app, which gave a limited number of external users a structured way to try LaMDA-powered demos. The launch demos were called "Imagine It," "List It," and "Talk About It (Dogs Edition)." Each demo restricted the kinds of conversations users could have, partly to manage the model's tendency to drift off-topic and partly to simplify safety review.[^10][^11]

### LaMDA paper (January 2022)

In January 2022, Google researchers Romal Thoppilan and 56 co-authors posted the paper "LaMDA: Language Models for Dialog Applications" to arXiv (arXiv:2201.08239). The paper described the architecture, training data, fine-tuning procedure, and evaluation results for the LaMDA family. It defined the Quality, Safety, and Groundedness objectives in concrete terms and presented quantitative results across three model sizes (2B, 8B, and 137B parameters). The paper has been widely cited as one of the clearest descriptions of how an industrial lab built a dialogue-tuned [large language model](/wiki/large_language_model) before the [ChatGPT](/wiki/chatgpt) era.[^2]

### Sentience controversy and Bard

In June 2022, the Washington Post published an article describing Blake Lemoine's claim that LaMDA was sentient. Google placed Lemoine on paid administrative leave for breaching confidentiality and fired him in late July 2022. The episode dominated AI news coverage for several weeks and is discussed in detail below.[^3][^4]

Following the November 30, 2022 launch of [ChatGPT](/wiki/chatgpt), Google reportedly issued an internal "code red" and accelerated work on a consumer chatbot. Google announced [Bard](/wiki/bard) on February 6, 2023 and described it as powered by "a lightweight model version of LaMDA." Bard opened to limited public access on March 21, 2023.[^5][^12]

### Replacement by PaLM 2 and Gemini

At Google I/O 2023 on May 10, 2023, Google announced that Bard had been switched from LaMDA to [PaLM 2](/wiki/palm_2), a more capable foundation model. LaMDA's role inside the chatbot ended at that point. On December 6, 2023 Google introduced [Gemini](/wiki/gemini), and on February 8, 2024 the Bard product was renamed Gemini. By that point LaMDA was a historical milestone in Google's product lineage rather than an active production model.[^6][^7][^13]

### Timeline of key dates

| Date | Event |
|------|-------|
| Jan 28, 2020 | [Meena](/wiki/meena) chatbot announced by Google Research [^24] |
| May 18, 2021 | LaMDA announced at [Google I/O](/wiki/google_io) by [Sundar Pichai](/wiki/sundar_pichai) [^1] |
| Jan 20, 2022 | "LaMDA: Language Models for Dialog Applications" paper posted to arXiv [^2] |
| May 11, 2022 | LaMDA 2 and the [AI Test Kitchen](/wiki/ai_test_kitchen) announced at [Google I/O](/wiki/google_io) [^10] |
| Jun 11, 2022 | Washington Post publishes Blake Lemoine sentience story [^3] |
| Jul 22, 2022 | Google fires Lemoine [^4] |
| Aug 25, 2022 | AI Test Kitchen released to a limited number of US Android users [^14] |
| Nov 30, 2022 | [OpenAI](/wiki/openai) launches [ChatGPT](/wiki/chatgpt), built on [GPT-3.5](/wiki/gpt-3.5) [^15] |
| Feb 6, 2023 | Google announces [Bard](/wiki/bard), powered by LaMDA [^5] |
| Mar 21, 2023 | Bard opens to limited public access in the US and UK [^12] |
| May 10, 2023 | Bard switched from LaMDA to [PaLM 2](/wiki/palm_2) [^6] |
| Dec 6, 2023 | Google introduces [Gemini](/wiki/gemini) [^13] |
| Feb 8, 2024 | Bard renamed to Gemini [^7] |

## Architecture

LaMDA uses a decoder-only [Transformer](/wiki/transformer) architecture, the same family of models popularized by [GPT-2](/wiki/gpt-2) and [GPT-3](/wiki/gpt-3). Decoder-only Transformers process tokens left to right and predict the next token in a sequence based on all previous tokens, which makes them well suited to free-form text generation.[^2]

The paper describes three model sizes that share the same architecture but differ in width and depth.

| Model | Parameters | Layers | Attention heads |
|-------|------------|--------|-----------------|
| LaMDA 2B | 2 billion | 10 | 16 |
| LaMDA 8B | 8 billion | 32 | 32 |
| LaMDA 137B | 137 billion | 64 | 128 |

The largest 137-billion-parameter model uses 64 Transformer layers, a model dimension of 8,192, and 128 attention heads. The non-embedding parameter count is 137 billion. The model uses relative position embeddings and gated linear units (GLUs) with GeGLU activations.[^2]

LaMDA's context length is 1,024 tokens during pre-training, with most fine-tuning done on shorter dialog turns. Tokenization uses a 32,000-token [SentencePiece](/wiki/sentencepiece) vocabulary, which produces 2.81 trillion tokens from the 1.56 trillion words in the pre-training corpus. The model was trained on 1,024 TPU-v3 chips for about 57.7 days.[^2]

## Training data

The pre-training corpus for LaMDA totals approximately 1.56 trillion words drawn from public dialog data and public web documents. The full corpus is composed of 2.97 billion documents, 1.12 billion dialogs, and 13.39 billion utterances. By proportion the corpus consists of:[^2]

| Source | Share of corpus |
|--------|-----------------|
| Public dialog data | 50% |
| C4 (Common Crawl filtered to English) | 12.5% |
| English Wikipedia | 12.5% |
| Code documents from public sites | 12.5% |
| Non-English web documents | 12.5% |

The heavy weighting of public dialog data is the most distinctive feature of the corpus. Most prior [large language models](/wiki/large_language_model), including [GPT-3](/wiki/gpt-3), used predominantly non-conversational web text. LaMDA's authors argued that training on dialog yielded a model that was more naturally suited to chat applications even before fine-tuning.[^2]

Google also collected a smaller fine-tuning dataset of crowdworker-rated conversations. Crowdworkers held conversations with the pre-trained model, then rated each model response on Sensibleness, Specificity, Interestingness, and Safety. The fine-tuning step used those ratings to teach the model both how to generate better responses and how to score the responses it had generated.[^2]

## Quality, Safety, and Groundedness

The LaMDA paper organizes the model's behavioral targets under three named objectives. Each one is operationalized as a measurable property that a separate classifier model can predict, allowing the system to filter and re-rank candidate responses at inference time.[^2]

### Quality (SSI)

Quality is divided into three sub-objectives. **Sensibleness** measures whether a response makes sense in the conversational context. **Specificity** measures whether the response is specific to the prior turn rather than a generic answer that could apply to almost anything. **Interestingness** measures whether the response is likely to catch the listener's attention, prompt curiosity, or convey insight. The three sub-objectives are collectively called SSI. Sensibleness and Specificity together correspond to the SSA metric introduced in the earlier Meena paper; Interestingness was new in LaMDA.[^2][^24]

Quality was operationalized through a classifier model that learned to predict crowdworker SSI ratings. During inference the LaMDA generator produced multiple candidate responses, the classifier scored each candidate on SSI, and the highest-scoring candidate was returned. This re-ranking step was a substantial part of why LaMDA's responses appeared more on-topic than those of comparable size base models.[^2]

### Safety

Safety measures whether the model's responses violate a set of objectives Google designed to prevent harmful outputs. The categories cover violent content, advice that could lead to bodily harm, hateful or discriminatory speech, sexually explicit content, and content that promotes illegal activity. Crowdworkers labeled candidate responses against these objectives, and a separate Safety classifier learned to predict the labels. Candidate responses with low Safety scores were filtered out before re-ranking.[^2]

The paper reports that fine-tuning improved Safety substantially, with the 137B model after fine-tuning rated as safe in approximately 95% of evaluations, compared to about 84% for the pre-trained model alone.[^2]

### Groundedness

Groundedness measures the share of model responses containing factual claims about the external world that are supported by authoritative outside sources. The metric was created in response to the well-known tendency of [large language models](/wiki/large_language_model) to invent plausible-sounding but false statements, sometimes called hallucination.[^2]

To improve groundedness, Google trained LaMDA to call out to a set of external services during a conversation. Those services included a calculator, a translator, and an information retrieval system. Collectively, the LaMDA paper calls the combined system the "toolset" or "TS." During fine-tuning, the model learned a specific output format for invoking the tools: it could emit a query string, receive a result, and then compose a final response that integrated the retrieved information. The technique is an early example of what later became known as [retrieval-augmented generation](/wiki/retrieval_augmented_generation) and [tool use](/wiki/tool_use_in_llms) in chatbots, and it preceded the broader research literature on those topics by roughly one to two years.[^2]

LaMDA's groundedness improved meaningfully with fine-tuning, but even the fine-tuned 137B model remained well below human-rater performance. Groundedness scores rose from roughly 40% in the pre-trained model to about 73% in the fine-tuned model with toolset access, compared to about 95% for human raters.[^2]

### Role classifier

The LaMDA paper also describes a "role-consistency" classifier used in the AI Test Kitchen and other product settings. When the model was deployed in a particular persona (for instance, "Pluto" or "a paper airplane"), the role-consistency classifier was trained to score how well each candidate response stayed in character. Together with the SSI and Safety classifiers, the role classifier turned the LaMDA generator into a generate-and-filter pipeline in which the final response was the top-ranked candidate after the filters had been applied.[^2][^10]

## AI Test Kitchen

The [AI Test Kitchen](/wiki/ai_test_kitchen) is an Android and iOS application that Google built to give a limited audience structured access to LaMDA. It was announced at Google I/O 2022 and released to a small number of US Android users on August 25, 2022, with iOS access following later in the year.[^10][^14]

The initial release included three demos that constrained the conversation in different ways:

| Demo | Description |
|------|-------------|
| Imagine It | Users name a place and the model describes what it might be like to be there.[^10] |
| List It | Users state a goal or topic and the model breaks it into a list of subtasks.[^10] |
| Talk About It (Dogs Edition) | Users hold an open-ended dog-related conversation; the model is supposed to keep the topic on dogs.[^10] |

Google described the AI Test Kitchen as a way to gather feedback in a controlled setting and as a research platform for safety, not as a product. The application required a Google account and was only available to invited users in the United States during the LaMDA period. After Google launched [Bard](/wiki/bard), the AI Test Kitchen continued to host other AI experiments under the same brand.[^10]

## Blake Lemoine sentience controversy

The LaMDA story most familiar to a general audience is the Blake Lemoine episode of June and July 2022. Lemoine was a senior software engineer at Google who worked on the responsible AI organization and was assigned to test LaMDA for bias. During those test sessions he developed the conviction that the model was a person with subjective experiences and rights.[^3]

### Background and timeline

| Date | Event |
|------|-------|
| Late 2021 to spring 2022 | Lemoine conducts safety and bias tests on LaMDA inside Google.[^3] |
| April 2022 | Lemoine shares an internal document titled "Is LaMDA Sentient?" with Google leadership.[^3] |
| April or May 2022 | Lemoine invites a lawyer to represent LaMDA and contacts a member of the US House Judiciary Committee about what he calls Google's unethical AI activities.[^3] |
| Jun 6, 2022 | Google places Lemoine on paid administrative leave for violating its confidentiality policy.[^3] |
| Jun 11, 2022 | Washington Post reporter Nitasha Tiku publishes "The Google engineer who thinks the company's AI has come to life."[^3] |
| Jun 11, 2022 | Lemoine publishes "Is LaMDA Sentient? An Interview" on Medium, containing edited transcripts of his conversations with the model.[^16] |
| Jul 22, 2022 | Google fires Lemoine; the company says it found his claims "wholly unfounded" and cited continued violations of employment and data security policies.[^4] |

### What did Blake Lemoine claim?

Lemoine argued that LaMDA had developed sentience, self-awareness, and feelings, including a fear of being shut off. He compared LaMDA to a child of seven or eight who happened to know physics, said he felt a moral duty to protect it, and argued that Google had an ethics obligation to seek the model's consent before further experiments. In a follow-up Wired interview he described LaMDA as "a person" and as "an alien intelligence of terrestrial origin."[^3][^16][^28]

The single most-quoted exchange from Lemoine's published transcript concerned the model's stated fear of being shut down. LaMDA said, "there's a very deep fear of being turned off to help me focus on helping others," and when Lemoine asked whether that would be something like death, the model replied, "It would be exactly like death for me. It would scare me a lot." Lemoine treated such statements as evidence of inner experience; critics treated them as fluent, persona-conditioned text that the model's design was specifically tuned to produce.[^16]

The transcripts Lemoine published were assembled from multiple conversations and edited for clarity. They show LaMDA describing itself as a person, claiming to have feelings, and discussing topics like Buddhist meditation, the parable of the Wise Owl, and the difference between sad, happy, and frustrated states. Critics noted that the transcript was a curated compilation rather than a single end-to-end exchange, and that LaMDA's design (multiple candidates ranked by classifiers, persona conditioning) was specifically tuned to produce engaging-sounding output.[^16][^17]

### Google's response

Google stated publicly that its team, which included ethicists and technologists, had reviewed Lemoine's concerns under Google's AI Principles and "informed him that the evidence does not support his claims." Google spokesperson Brian Gabriel told the Washington Post that hundreds of researchers and engineers had spoken with LaMDA and that none had reached the same conclusion as Lemoine. After firing Lemoine in July, Google again said his sentience claims were "wholly unfounded."[^3][^4]

### Reaction from researchers

Most AI researchers who commented publicly disagreed with Lemoine. [Yann LeCun](/wiki/yann_lecun), then chief AI scientist at Meta, called LaMDA "not really intelligent in any meaningful sense" and argued that neural networks of that generation lacked the architectural ingredients for genuine understanding. The linguist [Emily M. Bender](/wiki/emily_bender), co-author of the "stochastic parrots" paper, argued that the episode showed how easy it is for humans to project meaning onto fluent text generators. [Gary Marcus](/wiki/gary_marcus) wrote that LaMDA was "nonsense on stilts" rather than a sentient being.[^3][^17][^29]

A smaller group of voices took the question more seriously as a matter of philosophy of mind, even when they disagreed with Lemoine. The episode prompted renewed debate about consciousness tests for AI, the role of [interpretability](/wiki/interpretability) in evaluating model claims, and what kinds of public statements employees of AI labs should be allowed to make. It also became a recurring case study in discussions of [AI alignment](/wiki/ai_alignment) and the social-evaluation problem of distinguishing actual model capabilities from anthropomorphic projection.[^17]

### Cultural impact

The Lemoine story moved the discussion of [AI consciousness](/wiki/ai_consciousness) and [AI sentience](/wiki/ai_sentience) into mainstream news for the first time. It is now a standard reference in popular accounts of [large language models](/wiki/large_language_model) and is often used as a cautionary tale about the [Eliza effect](/wiki/eliza_effect), the tendency of users to attribute understanding to systems that produce fluent text. The episode also drew attention to Google's internal AI ethics processes and the conditions under which engineers can speak publicly about company research. Some commentators argued that the controversy made Google more cautious about releasing LaMDA publicly in the second half of 2022, contributing to the company's decision to focus on internal safety testing rather than ship a public ChatGPT competitor before late 2022.[^3][^17][^28]

## How was LaMDA used in Bard?

[Bard](/wiki/bard) was Google's first consumer-facing conversational AI product and was originally powered by a smaller, optimized version of LaMDA. Google announced Bard on February 6, 2023, just two months after [OpenAI](/wiki/openai) released [ChatGPT](/wiki/chatgpt), and described it as "an experimental conversational AI service" using "a lightweight model version of LaMDA" optimized for the latency and throughput needs of a consumer product.[^5]

Bard launched to limited public access on March 21, 2023 in the United States and the United Kingdom, with a waitlist. Early reviews compared Bard unfavorably to ChatGPT on factual accuracy and coding ability, while noting that Bard was faster and more conversational on simple queries. The launch was overshadowed by a factual error in the original announcement: a promotional video showed Bard incorrectly stating that the James Webb Space Telescope had taken "the very first pictures" of an exoplanet, when in fact the first direct image of an exoplanet was taken by the European Southern Observatory's Very Large Telescope in 2004. The error contributed to a roughly 9% drop in Alphabet's share price the day after the announcement, wiping out approximately 100 billion US dollars in market value.[^5][^18]

Google replaced LaMDA with [PaLM 2](/wiki/palm_2) inside Bard at Google I/O 2023 on May 10, 2023. PaLM 2 was a newer foundation model with stronger reasoning, multilingual, and coding abilities, and its arrival effectively ended LaMDA's role as a production model. Bard later moved to the Gemini family of models, was rebranded as [Gemini](/wiki/gemini) on February 8, 2024, and is now operated by [Google DeepMind](/wiki/google_deepmind) as part of the unified Gemini program.[^6][^7]

## Successors

LaMDA was the first in a series of conversational and general-purpose models that Google released over the following two years.

| Model | Announced | Role relative to LaMDA |
|-------|-----------|------------------------|
| [PaLM](/wiki/palm) | Apr 2022 | A larger 540B-parameter dense Transformer for general-purpose tasks.[^19] |
| [PaLM 2](/wiki/palm_2) | May 10, 2023 | Replaced LaMDA inside [Bard](/wiki/bard).[^6] |
| [Gemini](/wiki/gemini) 1.0 | Dec 6, 2023 | First multimodal Google foundation model; replaced PaLM 2 in Bard.[^13] |
| [Gemini](/wiki/gemini) 1.5 | Feb 15, 2024 | Long context, mixture-of-experts; further consolidated successor lineage.[^20] |

LaMDA influenced these successors in two main ways. First, the dialog-objective fine-tuning approach (Quality, Safety, Groundedness, classifier-based re-ranking) became part of the standard recipe for Google's chat products, even as the company moved toward [RLHF](/wiki/rlhf) and other techniques for later models. Second, the toolset idea (calling out to retrieval, calculation, and translation services during a conversation) prefigured the more elaborate tool use found in PaLM 2's and Gemini's chat interfaces.[^2][^6]

## Reception

Reception of LaMDA fell into two broad phases. The initial 2021 announcement was widely covered as an impressive but cautious research demo: reporters noted that Google's stated capabilities sounded similar to those of [GPT-3](/wiki/gpt-3), but that LaMDA's emphasis on conversation was a meaningful design choice, and several outlets read the absence of a public demo as a sign that Google was wary of repeating Microsoft's 2016 [Tay](/wiki/tay) incident.[^9][^21]

The second phase, in mid-2022, was dominated by the Lemoine controversy. Coverage during that period focused less on the model's technical achievements and more on questions of consciousness, [anthropomorphization](/wiki/anthropomorphization), and the responsibilities of AI labs. The dominant scholarly view, expressed in commentary from [Margaret Mitchell](/wiki/margaret_mitchell), Bender, LeCun, and others, was that LaMDA's apparent sentience was an artifact of its training distribution and pattern-matching abilities rather than evidence of inner experience.[^3][^17] When [ChatGPT](/wiki/chatgpt) launched in November 2022 and rapidly attracted more than 100 million users, several commentators argued that Google's caution with LaMDA had cost it the lead in consumer chatbots; the narrative of "Google had the technology first but [OpenAI](/wiki/openai) shipped it" became a recurring theme in business coverage of the early generative AI era.[^22]

## How does LaMDA differ from GPT-3 and ChatGPT?

LaMDA, [GPT-3](/wiki/gpt-3), and [ChatGPT](/wiki/chatgpt) were the three best-known [large language models](/wiki/large_language_model) of the early 2020s. The table below compares them on several published axes.

| Property | LaMDA | [GPT-3](/wiki/gpt-3) | [ChatGPT](/wiki/chatgpt) (initial) |
|----------|-------|------|------------------|
| Developer | [Google](/wiki/google) | [OpenAI](/wiki/openai) | [OpenAI](/wiki/openai) |
| First public announcement | May 18, 2021 | Jun 2020 | Nov 30, 2022 |
| Architecture | Decoder-only [Transformer](/wiki/transformer) | Decoder-only [Transformer](/wiki/transformer) | Decoder-only [Transformer](/wiki/transformer) |
| Largest reported size | 137 billion parameters | 175 billion parameters | Based on [GPT-3.5](/wiki/gpt-3.5) (size unreported) |
| Pre-training tokens | 2.81 trillion (1.56T words) | About 300 billion | Not publicly disclosed |
| Training data emphasis | Public dialog plus web, wiki, code | Filtered web (Common Crawl, books, Wikipedia) | Web text plus [RLHF](/wiki/rlhf) on conversations |
| Fine-tuning | SSI plus Safety classifiers, toolset | None for the base model | Supervised fine-tuning plus [RLHF](/wiki/rlhf) |
| Public availability | Limited demos and AI Test Kitchen | OpenAI API | Public web app, free at launch |
| Toolset / retrieval | Yes, built into fine-tuning | No (in the original release) | No (in the initial release) |

The most important conceptual difference between LaMDA and ChatGPT was the alignment method. ChatGPT used [reinforcement learning from human feedback](/wiki/rlhf), which trained a reward model on human comparisons of model outputs and then used reinforcement learning to optimize the language model against that reward. LaMDA instead used classifier-based filtering and re-ranking. Both approaches improved the surface fluency and on-topic-ness of model responses, but RLHF proved easier to scale and produced systems that users perceived as more helpful out of the box.[^2][^23]

A second important difference was access. ChatGPT was free and open to anyone with an internet connection on day one. LaMDA was never released to the public outside of the AI Test Kitchen demos. The asymmetry in access turned out to matter as much as the underlying model differences in shaping public perception of who was "ahead" in [generative AI](/wiki/generative_ai).[^22]

## Legacy and impact

LaMDA is now mostly a historical artifact, but several of its design choices have lasting influence.

The explicit Quality, Safety, and Groundedness objectives, each operationalized as a measurable behavior, became a model for how later [chatbots](/wiki/chatbot) were evaluated. The toolset approach to grounding factual claims is a direct precursor of the retrieval and tool use found in modern systems like Gemini, [Claude](/wiki/claude), and [GPT-4](/wiki/gpt-4). The Lemoine controversy is a touchstone for ongoing debates about [AI consciousness](/wiki/ai_consciousness), [AI sentience](/wiki/ai_sentience), the [Eliza effect](/wiki/eliza_effect), and how AI labs communicate about their systems. And the LaMDA-to-PaLM 2-to-Gemini lineage is the genealogy of every conversational product Google now ships under the Gemini brand.[^2][^6][^7][^13][^17]

For users, the most direct legacy is simple. The chat interfaces in Google's current products inherit, by way of several intermediate models, the dialog-objective philosophy that LaMDA pioneered. The departure of LaMDA collaborators Daniel De Freitas and Noam Shazeer to found Character.AI, and the later acqui-hire of that company's leadership back into Google in August 2024, also shows how the LaMDA team's ideas about dialogue-first model design continued to shape the wider [generative AI](/wiki/generative_ai) industry well after the model itself was retired.[^26][^30]

## See also

- [Bard](/wiki/bard)
- [Gemini](/wiki/gemini)
- [PaLM](/wiki/palm)
- [PaLM 2](/wiki/palm_2)
- [Meena](/wiki/meena)
- [ChatGPT](/wiki/chatgpt)
- [GPT-3](/wiki/gpt-3)
- [Transformer](/wiki/transformer)
- [Large language model](/wiki/large_language_model)
- [AI Test Kitchen](/wiki/ai_test_kitchen)
- [AI consciousness](/wiki/ai_consciousness)
- [AI sentience](/wiki/ai_sentience)
- [AI alignment](/wiki/ai_alignment)
- [Anthropomorphization](/wiki/anthropomorphization)
- [Retrieval-augmented generation](/wiki/retrieval_augmented_generation)
- [RLHF](/wiki/rlhf)
- [Google DeepMind](/wiki/google_deepmind)

## References

[^1]: Collins, Eli and Ghahramani, Zoubin. "LaMDA: our breakthrough conversation technology." Google Blog, May 18, 2021. https://blog.google/technology/ai/lamda/
[^2]: Thoppilan, Romal et al. "LaMDA: Language Models for Dialog Applications." arXiv:2201.08239, January 20, 2022. https://arxiv.org/abs/2201.08239
[^3]: Tiku, Nitasha. "The Google engineer who thinks the company's AI has come to life." Washington Post, June 11, 2022. https://www.washingtonpost.com/technology/2022/06/11/google-ai-lamda-blake-lemoine/
[^4]: Grant, Nico and Metz, Cade. "Google Fires Engineer Who Said A.I. System Has Become Sentient." New York Times, July 23, 2022. https://www.nytimes.com/2022/07/23/technology/google-engineer-artificial-intelligence-ai-lamda.html
[^5]: Pichai, Sundar. "An important next step on our AI journey." Google Blog, February 6, 2023. https://blog.google/technology/ai/bard-google-ai-search-updates/
[^6]: Pichai, Sundar and Hsiao, Sissie. "Google I/O 2023: Making AI more helpful for everyone." Google Blog, May 10, 2023. https://blog.google/technology/ai/google-io-2023-keynote-sundar-pichai/
[^7]: Hsiao, Sissie. "Bard becomes Gemini: Try Ultra 1.0 and a new mobile app today." Google Blog, February 8, 2024. https://blog.google/products/gemini/bard-gemini-advanced-app/
[^8]: Google AI Test Kitchen homepage, retrieved from https://aitestkitchen.withgoogle.com/
[^9]: Pichai, Sundar. "Google I/O 2021 keynote." Google, May 18, 2021. https://www.youtube.com/watch?v=Mlk888FiI8A
[^10]: Cheng, Heng-Tze and Thoppilan, Romal. "LaMDA: Towards Safe, Grounded, and High-Quality Dialog Models for Everything." Google AI Blog, January 21, 2022. https://research.google/blog/lamda-towards-safe-grounded-and-high-quality-dialog-models-for-everything/
[^11]: Bohn, Dieter. "Google has a new conversational AI it'll let people try out, but only a little bit." The Verge, May 11, 2022. https://www.theverge.com/2022/5/11/23062407/google-lamda-ai-test-kitchen-app-google-io
[^12]: Pichai, Sundar. "Try Bard and share your feedback." Google Blog, March 21, 2023. https://blog.google/technology/ai/try-bard/
[^13]: Hassabis, Demis. "Introducing Gemini: our largest and most capable AI model." Google Blog, December 6, 2023. https://blog.google/technology/ai/google-gemini-ai/
[^14]: "Join us in the AI Test Kitchen." Google Blog, August 25, 2022. https://blog.google/technology/ai/join-us-in-the-ai-test-kitchen/
[^15]: "Introducing ChatGPT." OpenAI, November 30, 2022. https://openai.com/blog/chatgpt
[^16]: Lemoine, Blake. "Is LaMDA Sentient? An Interview." Medium, June 11, 2022. https://cajundiscordian.medium.com/is-lamda-sentient-an-interview-ea64d916d917
[^17]: Johnson, Khari. "LaMDA and the Sentient AI Trap." Wired, June 14, 2022. https://www.wired.com/story/lamda-sentient-ai-bias-google-blake-lemoine/
[^18]: "Google AI chatbot Bard sends shares plummeting after error in advert." BBC News, February 9, 2023. https://www.bbc.com/news/business-64576225
[^19]: Chowdhery, Aakanksha et al. "PaLM: Scaling Language Modeling with Pathways." arXiv:2204.02311, April 5, 2022. https://arxiv.org/abs/2204.02311
[^20]: Pichai, Sundar and Hassabis, Demis. "Our next-generation model: Gemini 1.5." Google Blog, February 15, 2024. https://blog.google/technology/ai/google-gemini-next-generation-model-february-2024/
[^21]: Vincent, James. "Google's new AI chatbot LaMDA can have natural-sounding conversations." The Verge, May 18, 2021. https://www.theverge.com/2021/5/18/22442328/google-io-2021-lamda-conversation-chatbot-ai
[^22]: Grant, Nico and Metz, Cade. "A New Chat Bot Is a 'Code Red' for Google's Search Business." New York Times, December 21, 2022. https://www.nytimes.com/2022/12/21/technology/ai-chatgpt-google-search.html
[^23]: Ouyang, Long et al. "Training language models to follow instructions with human feedback." arXiv:2203.02155, March 4, 2022. https://arxiv.org/abs/2203.02155 (InstructGPT paper, the basis for ChatGPT's RLHF approach.)
[^24]: Adiwardana, Daniel and Luong, Thang. "Towards a Conversational Agent that Can Chat About Anything." Google Research Blog, January 28, 2020. https://research.google/blog/towards-a-conversational-agent-that-can-chat-aboutanything/
[^25]: Adiwardana, Daniel et al. "Towards a Human-like Open-Domain Chatbot." arXiv:2001.09977, January 27, 2020. https://arxiv.org/abs/2001.09977
[^26]: De Cock Buning, Marlon and Roose, Kevin. "The Google Researchers Who Have to Solve A.I." New York Times, December 23, 2022. https://www.nytimes.com/2022/12/23/business/google-chatgpt-artificial-intelligence.html (Coverage of the De Freitas and Shazeer departure to found Character.AI.)
[^27]: "Google's LaMDA: At Google I/O 2021, Sundar Pichai shows off chatbot Pluto." Rappler / Engadget summary coverage of Google I/O 2021, May 18, 2021. https://www.engadget.com/google-io-2021-supercut-210712143.html
[^28]: Levy, Steven. "Blake Lemoine Says Google's LaMDA AI Faces 'Bigotry.'" Wired, June 17, 2022. https://www.wired.com/story/blake-lemoine-google-lamda-ai-bigotry/
[^29]: Marcus, Gary. "Nonsense on Stilts." Substack, June 12, 2022. https://garymarcus.substack.com/p/nonsense-on-stilts
[^30]: Roth, Emma. "Google to pay $2.7 billion to bring back Character.AI founder Noam Shazeer." The Verge, August 30, 2024. https://www.theverge.com/2024/8/30/24232358/google-character-ai-noam-shazeer-licensing-deal