---
title: "LaMDA (Google)"
source: "https://systems-analysis.info/eng/LaMDA_(Google)"
wiki: "systems-analysis.info/eng"
article: "LaMDA_(Google)"
language: "en"
categories:
  - "Category:English"
  - "Category:Google"
  - "Category:Large language models"
  - "Category:LLM families"
  - "Category:Machine learning"
  - "Category:Technology"
revision_id: 207
wiki_created_at: 2026-09-06T22:18:54Z
wiki_modified_at: 2026-09-06T22:18:54Z
downloaded_at: 2026-09-07T22:21:52Z
---

# LaMDA (Google)

**LaMDA** (**L**anguage **M**odel for **D**ialogue **A**pplications) is a family of [large language models](https://systems-analysis.info/eng/Large_language_model "Large language model") based on the [Transformer architecture](https://systems-analysis.info/eng/Transformer_architecture "Transformer architecture"), developed by Google and specialized in conducting meaningful, open-ended dialogues<sup>[\[1\]](https://systems-analysis.info/eng/LaMDA_(Google)#cite_note-thoppilan2022-1)</sup>. Unlike many general-purpose models of its time, LaMDA was specifically trained to maintain coherent, multi-turn conversations on virtually any topic, freely switching between contexts<sup>[\[2\]](https://systems-analysis.info/eng/LaMDA_(Google)#cite_note-google-blog-lamda-2)</sup>.

The model was first publicly unveiled at the **Google I/O conference in May 2021**<sup>[\[3\]](https://systems-analysis.info/eng/LaMDA_(Google)#cite_note-verge-io-2021-3)</sup>. LaMDA was positioned as a fundamental step toward more natural human-computer interaction, such as through conversational interfaces in search and voice assistants<sup>[\[4\]](https://systems-analysis.info/eng/LaMDA_(Google)#cite_note-google-helpful-moments-4)</sup>.

## Architecture and Training

### Fundamental Architecture: Decoder-Only

LaMDA is a **[decoder-only](https://systems-analysis.info/eng/Decoder-only_models_(architecture) "Decoder-only models (architecture)")** language model built on the Transformer architecture. This architecture is standard for text generation tasks. The model operates **autoregressively**—it predicts the next word ([token](https://systems-analysis.info/eng/Token_(LLM) "Token (LLM)")) in a sequence based on all preceding words. This allows it to generate coherent and logical text, continuing a given conversation, but it limits its ability to see the "right-side" context, unlike [BERT](https://systems-analysis.info/eng/BERT_(language_model) "BERT (language model)")<sup>[\[1\]](https://systems-analysis.info/eng/LaMDA_(Google)#cite_note-thoppilan2022-1)</sup>. The largest model in the family uses 64 Transformer layers with a model dimension of 8192 and 128 attention heads, together with T5-style relative attention and a gated-GeLU activation<sup>[\[1\]](https://systems-analysis.info/eng/LaMDA_(Google)#cite_note-thoppilan2022-1)</sup>.

### Scale and Training Data

The LaMDA family includes three models with **2 billion**, **8 billion**, and **137 billion** parameters (the largest counted as non-embedding parameters, roughly 50 times more than Meena). For [pre-training](https://systems-analysis.info/eng/Pre-training_of_large_language_models "Pre-training of large language models"), a massive corpus of **1.56 trillion words** was used, drawn from public dialogue data and web documents (2.97 billion documents, 1.12 billion dialogues, and 13.39 billion utterances, over 90% of it in English). This volume was nearly 40 times larger than the data used to train LaMDA's predecessor, the **Meena** model. The largest model was pre-trained on 1024 TPU-v3 chips over about 57.7 days<sup>[\[1\]](https://systems-analysis.info/eng/LaMDA_(Google)#cite_note-thoppilan2022-1)</sup>.

### Fine-Tuning Process and Metrics

Google researchers concluded that scaling alone was insufficient to ensure safety and factual accuracy in responses. Therefore, a multi-stage [fine-tuning](https://systems-analysis.info/eng/Fine-tuning_(deep_learning) "Fine-tuning (deep learning)") process was developed, during which the model was specifically optimized on three key metrics evaluated by human raters<sup>[\[1\]](https://systems-analysis.info/eng/LaMDA_(Google)#cite_note-thoppilan2022-1)</sup>:

- **Quality**: Assessed through three components:
  - *Sensibleness'*: Logicality and contextual relevance.
  - *Specificity'*: Concreteness and informativeness of the responses.
  - *Interestingness'*: Insightfulness and wit.

<!-- -->

- **Safety**: Preventing the generation of harmful, biased, or toxic statements. A special classifier-filter was fine-tuned for this purpose.
- **Groundedness**: Aimed at combating "[hallucinations](https://systems-analysis.info/eng/LLM_hallucinations "LLM hallucinations")" (fabricated facts). LaMDA was fine-tuned to consult a suite of **external tools** (a search engine, calculator, translator) when necessary to verify and clarify factual information<sup>[\[1\]](https://systems-analysis.info/eng/LaMDA_(Google)#cite_note-thoppilan2022-1)</sup>. This innovation was one of the first systemic solutions to the problem of factuality in large language models.

## Development and Implementation History

### Public Announcements and LaMDA 2

At **Google I/O 2021**, CEO Sundar Pichai demonstrated LaMDA's capabilities by showcasing dialogues where the model conversed from the perspective of the planet Pluto and a paper airplane<sup>[\[5\]](https://systems-analysis.info/eng/LaMDA_(Google)#cite_note-vincent-pluto-5)</sup>.

A year later, at **Google I/O 2022**, **LaMDA 2** was introduced, described as an "even more advanced conversationalist." Concurrently, Google launched the **AI Test Kitchen** application—a "lab" for public testing where users could experience LaMDA in several demo scenarios<sup>[\[6\]](https://systems-analysis.info/eng/LaMDA_(Google)#cite_note-google-io-2022-keynote-6)</sup>. This allowed for the collection of large-scale feedback to further improve the model.

### Integration into Google Bard

In February 2023, amid the rapid rise in popularity of ChatGPT, Google announced the launch of its own experimental chatbot, **Bard**<sup>[\[7\]](https://systems-analysis.info/eng/LaMDA_(Google)#cite_note-google-bard-announcement-7)</sup>. Initially, Bard ran on a lightweight version of LaMDA to reduce computational requirements. LaMDA served as a key "transitional" technology, enabling Google to quickly bring a competitive product to market while more powerful models like [PaLM](https://systems-analysis.info/eng/PaLM_(Pathways_Language_Model) "PaLM (Pathways Language Model)") were being prepared for deployment. Bard was soon moved off LaMDA — to **PaLM 2** in May 2023 and then to the **[Gemini](https://systems-analysis.info/eng/Gemini_(Google) "Gemini (Google)")** models from December 2023 — and in February 2024 the product itself was renamed Gemini, effectively ending LaMDA's role as a front-line technology<sup>[\[8\]](https://systems-analysis.info/eng/LaMDA_(Google)#cite_note-bard-evolution-8)</sup>.

## The Blake Lemoine Incident

In June 2022, LaMDA became the center of a widespread public debate after **Blake Lemoine**, an engineer from Google's Responsible AI organization, publicly claimed that the model had, in his opinion, achieved sentience. He published excerpts of his conversations with LaMDA in which the model discussed self-awareness, feelings, and expressed a fear of being turned off<sup>[\[9\]](https://systems-analysis.info/eng/LaMDA_(Google)#cite_note-guardian-leave-9)</sup>.

### Official Position and Scientific Community's Reaction

Google strongly refuted Lemoine's claims, stating that after a review, no evidence of sentience was found, and there was "extensive evidence to the contrary"<sup>[\[9\]](https://systems-analysis.info/eng/LaMDA_(Google)#cite_note-guardian-leave-9)</sup>. In July 2022, Lemoine was fired for violating the company's confidentiality policy<sup>[\[10\]](https://systems-analysis.info/eng/LaMDA_(Google)#cite_note-guardian-fires-10)</sup>.

The vast majority of scientists and AI experts also rejected the idea of LaMDA's sentience. Linguist **Emily M. Bender** and other researchers emphasized that such models are "[stochastic parrots](https://systems-analysis.info/eng/Stochastic_parrot "Stochastic parrot")"—complex algorithms that statistically generate coherent text by mimicking human speech but without true understanding or consciousness<sup>[\[11\]](https://systems-analysis.info/eng/LaMDA_(Google)#cite_note-wapo-lemoine-11)</sup>. The incident clearly demonstrated how easily humans are prone to anthropomorphism, attributing human qualities to machines, and it spurred a global discussion about the nature of AI.

## Contribution and Legacy

Despite its relatively short lifecycle as a flagship technology, LaMDA left a significant mark on the history of conversational AI.

- **Technological Contribution**: LaMDA demonstrated the feasibility of creating open-ended, context-aware dialogue systems and pioneered a systemic approach to ensuring **safety** (value-based filtering) and **factual groundedness** (consulting external tools).
- **Role in the Google Ecosystem**: LaMDA became a critically important transitional technology that allowed Google to urgently enter the "chatbot wars" with its Bard product and served as a testing ground for methods that formed the basis of more powerful models like PaLM and [Gemini](https://systems-analysis.info/eng/Gemini_(Google) "Gemini (Google)").
- **Public Impact**: The Blake Lemoine incident elevated the discussion about the nature of AI, consciousness, and the risks of anthropomorphism to a new global level.

## External links

- <a href="https://blog.google/technology/ai/lamda/" class="external text" rel="nofollow">Official LaMDA announcement on the Google AI Blog</a>
- <a href="https://research.google/pubs/lamda-language-models-for-dialog-applications/" class="external text" rel="nofollow">LaMDA scientific publication page on Google Research</a>
- <a href="https://en.wikipedia.org/wiki/LaMDA" class="external text" rel="nofollow">LaMDA — Wikipedia</a>

## Literature

- Vaswani, A. et al. (2017). *Attention Is All You Need*. <a href="https://arxiv.org/abs/1706.03762" class="external text" rel="nofollow">arXiv:1706.03762</a>.
- So, D. R. et al. (2019). *The Evolved Transformer*. <a href="https://arxiv.org/abs/1901.11117" class="external text" rel="nofollow">arXiv:1901.11117</a>.
- Zhang, Y. et al. (2020). *DialoGPT: Large-Scale Generative Pre-training for Conversational Response Generation*. <a href="https://arxiv.org/abs/1911.00536" class="external text" rel="nofollow">arXiv:1911.00536</a>.
- Adiwardana, D. et al. (2020). *Towards a Human-like Open-Domain Chatbot*. <a href="https://arxiv.org/abs/2001.09977" class="external text" rel="nofollow">arXiv:2001.09977</a>.
- Roller, S. et al. (2021). *Recipes for Building an Open-Domain Chatbot*. <a href="https://arxiv.org/abs/2004.13637" class="external text" rel="nofollow">arXiv:2004.13637</a>.
- Lin, S. et al. (2021). *TruthfulQA: Measuring How Models Mimic Human Falsehoods*. <a href="https://arxiv.org/abs/2109.07958" class="external text" rel="nofollow">arXiv:2109.07958</a>.
- Thoppilan, R. et al. (2022). *LaMDA: Language Models for Dialog Applications*. <a href="https://arxiv.org/abs/2201.08239" class="external text" rel="nofollow">arXiv:2201.08239</a>.
- Bai, Y. et al. (2022). *Constitutional AI: Harmlessness from AI Feedback*. <a href="https://arxiv.org/abs/2212.08073" class="external text" rel="nofollow">arXiv:2212.08073</a>.

## References

1.  <span id="cite_note-thoppilan2022-1">↑ <sup>[1.0](https://systems-analysis.info/eng/LaMDA_(Google)#cite_ref-thoppilan2022_1-0)</sup> <sup>[1.1](https://systems-analysis.info/eng/LaMDA_(Google)#cite_ref-thoppilan2022_1-1)</sup> <sup>[1.2](https://systems-analysis.info/eng/LaMDA_(Google)#cite_ref-thoppilan2022_1-2)</sup> <sup>[1.3](https://systems-analysis.info/eng/LaMDA_(Google)#cite_ref-thoppilan2022_1-3)</sup> <sup>[1.4](https://systems-analysis.info/eng/LaMDA_(Google)#cite_ref-thoppilan2022_1-4)</sup> <sup>[1.5](https://systems-analysis.info/eng/LaMDA_(Google)#cite_ref-thoppilan2022_1-5)</sup> Thoppilan, Romal; De Freitas, Daniel; Hall, Jamie; et al. "LaMDA: Language Models for Dialog Applications". *arXiv*. <a href="https://arxiv.org/abs/2201.08239" class="external autonumber" rel="nofollow">[1]</a></span>
2.  <span id="cite_note-google-blog-lamda-2">[↑](https://systems-analysis.info/eng/LaMDA_(Google)#cite_ref-google-blog-lamda_2-0) Collins, Eli; Ghahramani, Zoubin. "LaMDA: our breakthrough conversation technology". *Google AI Blog*. <a href="https://blog.google/technology/ai/lamda/" class="external autonumber" rel="nofollow">[2]</a></span>
3.  <span id="cite_note-verge-io-2021-3">[↑](https://systems-analysis.info/eng/LaMDA_(Google)#cite_ref-verge-io-2021_3-0) Peters, Jay. "Google I/O 2021: the 14 biggest announcements". *The Verge*. <a href="https://www.theverge.com/2021/5/18/22435419/google-io-2021-event-recap-biggest-announcements-pixel-android-12-wear-os-workspace" class="external autonumber" rel="nofollow">[3]</a></span>
4.  <span id="cite_note-google-helpful-moments-4">[↑](https://systems-analysis.info/eng/LaMDA_(Google)#cite_ref-google-helpful-moments_4-0) "Google I/O 2021: Being helpful in moments that matter". *Official Google Blog*. <a href="https://blog.google/technology/developers/io21-helpful-google/" class="external autonumber" rel="nofollow">[4]</a></span>
5.  <span id="cite_note-vincent-pluto-5">[↑](https://systems-analysis.info/eng/LaMDA_(Google)#cite_ref-vincent-pluto_5-0) Vincent, James. "Google showed off its next-generation AI by talking to Pluto and a paper airplane". *The Verge*. <a href="https://www.theverge.com/2021/5/18/22442328/google-io-2021-ai-language-model-lamda-pluto" class="external autonumber" rel="nofollow">[5]</a></span>
6.  <span id="cite_note-google-io-2022-keynote-6">[↑](https://systems-analysis.info/eng/LaMDA_(Google)#cite_ref-google-io-2022-keynote_6-0) "Google I/O 2022: Advancing knowledge and computing (Keynote)". *Official Google Blog*. <a href="https://blog.google/technology/developers/io-2022-keynote/" class="external autonumber" rel="nofollow">[6]</a></span>
7.  <span id="cite_note-google-bard-announcement-7">[↑](https://systems-analysis.info/eng/LaMDA_(Google)#cite_ref-google-bard-announcement_7-0) Pichai, Sundar. "An important next step on our AI journey". *Official Google Blog*. <a href="https://blog.google/technology/ai/bard-google-ai-search-updates/" class="external autonumber" rel="nofollow">[7]</a></span>
8.  <span id="cite_note-bard-evolution-8">[↑](https://systems-analysis.info/eng/LaMDA_(Google)#cite_ref-bard-evolution_8-0) "What is the Google Gemini AI Model (Formerly Bard)?". *TechTarget*. <a href="https://www.techtarget.com/searchenterpriseai/definition/Google-Gemini" class="external autonumber" rel="nofollow">[8]</a></span>
9.  <span id="cite_note-guardian-leave-9">↑ <sup>[9.0](https://systems-analysis.info/eng/LaMDA_(Google)#cite_ref-guardian-leave_9-0)</sup> <sup>[9.1](https://systems-analysis.info/eng/LaMDA_(Google)#cite_ref-guardian-leave_9-1)</sup> Luscombe, Richard. "Google engineer put on leave after saying AI chatbot has become sentient". *The Guardian*. <a href="https://www.theguardian.com/technology/2022/jun/12/google-engineer-ai-bot-sentient-blake-lemoine" class="external autonumber" rel="nofollow">[9]</a></span>
10. <span id="cite_note-guardian-fires-10">[↑](https://systems-analysis.info/eng/LaMDA_(Google)#cite_ref-guardian-fires_10-0) "Google fires software engineer who claims AI chatbot is sentient". *The Guardian*. <a href="https://www.theguardian.com/technology/2022/jul/23/google-fires-software-engineer-who-claims-ai-chatbot-is-sentient" class="external autonumber" rel="nofollow">[10]</a></span>
11. <span id="cite_note-wapo-lemoine-11">[↑](https://systems-analysis.info/eng/LaMDA_(Google)#cite_ref-wapo-lemoine_11-0) Tiku, Nitasha. "The Google engineer who thinks the company's AI has come to life". *The Washington Post*. <a href="https://www.washingtonpost.com/technology/2022/06/11/google-ai-lamda-blake-lemoine/" class="external autonumber" rel="nofollow">[11]</a></span>