Sparrow
DeepMind's rule-following English chatbot that cites web evidence for its answers
Sparrow is a research chatbot built by DeepMind, an Alphabet subsidiary, and presented in a September 2022 paper on aligning dialogue agents with targeted human judgements. It fine-tunes the 70-billion-parameter Chinchilla model with human feedback, searches Google to cite evidence, and follows safety rules, though it was never released publicly.
- Fine-tuned from DeepMind's Chinchilla model with 70 billion parameters
- Uses Google Search to find and cite evidence for factual claims
- Trained with two reward models: a preference model and a rule model
- Broke its safety rules 8% of the time under adversarial probing
- Introduced in September 2022 but never released publicly
What it is
Sparrow is a chatbot created by DeepMind, the artificial intelligence research lab owned by Alphabet Inc. It answers questions in English while trying to cut down on unsafe, biased or wrong replies, a known weakness of large language models. The system is fine-tuned from Chinchilla, DeepMind's pre-trained language model with 70 billion parameters. Rather than relying on the base model alone, Sparrow is shaped by human judgements so that its answers are more helpful, correct and harmless. It was presented as a research project in September 2022 and was not made available to the public.
How it came about
DeepMind introduced Sparrow in a September 2022 paper titled "Improving alignment of dialogue agents via targeted human judgements". Paid study participants talked with the system and their preferences were gathered to train a model of answer quality. In January 2023, DeepMind CEO Demis Hassabis said the company was thinking about opening Sparrow as a private beta at some point during 2023.
Inside the model
Sparrow is a deep neural network built on the transformer architecture and fine-tuned from Chinchilla, DeepMind's 70-billion-parameter pre-trained language model. Training uses reinforcement learning from human feedback, with some supervised fine-tuning as well. Two reward models capture human judgements: a preference model that predicts which answer a participant would favour, and a rule model that flags rule violations. To reduce hallucination, the bot can query Google Search and cite evidence for factual claims. Behaviour is bounded by rules such as avoiding threats, hateful or insulting comments, harmful advice, and any claim to be a person.
What it can do
Sparrow answers questions and follow-up questions, backing factual claims with evidence found through Google Search. DeepMind's work on it focused on an English-speaking agent, and the lab said more effort was needed to reach similar results in other languages and cultural settings. When study participants deliberately tried to provoke rule-breaking, the system violated the rules 8% of the time, which the researchers reported as three times less often than the prompted Chinchilla baseline.
How to use it
Sparrow was never released publicly. It appeared only through the September 2022 research paper and related DeepMind writing. In January 2023 Demis Hassabis said a private beta was being considered for that year.
Safety and testing
Safety is built into the training process. A set of rules constrains the bot, covering threats, hateful or insulting language, possibly harmful advice and pretending to be human. Study participants were asked to converse with the system and attempt to trick it into breaking those rules, and their judgements trained a rule model used in further training. Under adversarial probing the bot still broke rules 8% of the time.
Why it matters
Sparrow is presented as a step toward dialogue agents that are more helpful, correct and harmless than plain pre-trained language models. Its combination of human preference training, an explicit rule model and search-backed citations offered a concrete approach to the problem of language models producing incorrect, biased or harmful output.
Questions people ask about Sparrow
What is Sparrow?
Sparrow is a research chatbot from DeepMind, the Alphabet Inc. lab. It answers questions in English and tries to stay helpful, correct and harmless. It is fine-tuned from Chinchilla, DeepMind's 70-billion-parameter language model, and was shown in a September 2022 paper rather than released as a product.
How does Sparrow avoid making things up?
It can search the internet through Google Search and cite evidence for the factual claims it makes. That is meant to reduce hallucination, where a language model states something incorrect with confidence.
How was it trained to behave safely?
Reinforcement learning from human feedback was used, along with some supervised fine-tuning. Two reward models guide it: one predicts which answer a human would prefer, the other predicts whether a rule has been broken. Paid participants also tried to trick the bot so their judgements could train the rule model.
Can I use Sparrow?
No public release happened. The work appeared in a September 2022 paper, and in January 2023 DeepMind CEO Demis Hassabis said a private beta was being considered for some time in 2023.
What languages does it support?
DeepMind's research on Sparrow centred on an English-speaking agent. The company said further work was required to obtain comparable results in other languages and cultural contexts.
Good reading and sources
- The secret to Sparrow, DeepMindThe Register, Article
- Building safer dialogue agents — Google DeepMindGoogle DeepMind, Article
The facts on this page come from the Wikipedia article "Sparrow (chatbot)", 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.