Google announced a dialogue model built for open-ended chat
LaMDA's announcement and later paper name sensibleness and safety metrics a companion product still has to earn separately.
- Historical event
- May 18, 2021
- First source published
- May 18, 2021
- Site publication
- September 18, 2026
What happened
On 18 May 2021, Google announced LaMDA (Language Model for Dialogue Applications), a family of Transformer-based models trained specifically on dialogue rather than general text. The announcement frames LaMDA as a step toward software that can engage in a free-flowing way about a seemingly endless number of topics, and names three qualities it optimises for: sensibleness, specificity and interestingness. A fuller paper followed in January 2022, describing a 137-billion-parameter model and the fine-tuning behind it.
What the documents show
The 2021 announcement is candid about a limit a finished product's marketing rarely admits: it states that factuality is something language models often struggle with, and that responses need to be compelling and correct, not just one or the other. The 2022 paper adds a method the blog post only gestures at: a separate LaMDA classifier, fine-tuned on a small set of crowdworker-labelled examples, filters candidate replies for safety before one is shown. The paper is explicit that scaling the model up alone improved general quality but 'shows less improvements on safety and factual grounding' — the annotated fine-tuning step, not size, did that work.
The mechanism
Two distinct mechanisms are named across the two documents and are worth keeping separate. Groundedness, per the paper, means checking a candidate response against an external source, such as an information-retrieval system, before offering it, so the model does not rely only on what it memorised in training. Safety filtering is a different step: a classifier trained on human-labelled examples of undesirable replies screens outputs after generation. Neither mechanism is memory of a specific user, and neither is a claim that the model 'knows' what is true; both are constraints applied around a model that otherwise predicts plausible text.
What it leaves open
LaMDA was announced as research, not as a product whose users form ongoing relationships with it; the documents describe conversation quality and safety filtering in the abstract, not what happens when the same person returns to the same chat across months. Whether groundedness or safety filtering, as described here, would hold up inside a companion app with a persistent persona is a question the two documents do not address, and treating them as a specification for such a product would be an editorial extrapolation this record does not make.
- Is a safety filter applied before or after a response is generated, and what does it check against?
- Does groundedness mean the system consulted an outside source, or only that it sounds confident?
- Was a safety claim tested by the same team that built the model, using its own criteria?
Read side by side, the 2021 announcement and the 2022 paper show a research lab naming its own unsolved problems before a product shipped around the same underlying technique. That sequence — a public claim, followed by a technical paper that qualifies it — is a useful pattern to look for whenever a company describes a dialogue system's safety in the present tense.
Sources & reading trail
Names sensibleness, specificity and interestingness and discloses factuality as an unresolved problem.
Source published: 18 May 2021 · Retrieved: 16 September 2026
Describes a safety classifier fine-tuned on annotated data and groundedness via external retrieval.
Source published: 20 January 2022 · Retrieved: 16 September 2026
Company documents, filings, studies and official records establish the record; the reading and the questions are Lovebot Journal editorial analysis. This retrospective draft does not imply the site published on the event date.
Continue reading
- Two 2020 papers proposed a way to grade how human a chatbot sounds
- Character.AI opened its beta built entirely on user-made personas
- A written constitution trained a model, not just labels
- Browse the complete the archive
Sources & reading trail
- LaMDA: our breakthrough conversation technology
Source published: May 18, 2021 · Retrieved: September 16, 2026 - LaMDA: Language Models for Dialog Applications
Source published: January 20, 2022 · Retrieved: September 16, 2026
The documents above establish the record. The reading and the questions are this publication’s editorial analysis, written after the fact.
Published September 18, 2026, not on the date of the event described.