Google Research
Last reviewed
Sources
41 citations
Review status
Source-backed
Revision
v2 ยท 3,067 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Sources
41 citations
Review status
Source-backed
Revision
v2 ยท 3,067 words
Add missing citations, update stale details, or suggest a clearer explanation.
Google Research is Google's organization for foundational and applied computer science research, with work spanning quantum computing, algorithms and theory, machine learning, health, climate, and crisis response. It is distinct from Google DeepMind, the unit created in April 2023 when Google merged its Google Brain team with DeepMind to concentrate frontier AI model development under Demis Hassabis [1]. Google Research kept the company's broader scientific portfolio after that split, and since April 2024 it has operated under a three-part mandate: computing systems including quantum, foundational machine learning and algorithms, and applied science and society [2]. Yossi Matias, the founding managing director of Google's research and development center in Israel, has served as vice president and head of Google Research since April 2024, reporting to James Manyika, Google's senior vice president and president for Research, Labs, Technology & Society [3][4]. The organization describes its mission as "Research, to reality": driving breakthroughs that benefit society, businesses, and Google products [5].
Google Research is the institutional descendant of the research division Google has run since the early 2000s, and it retains that group's defining habit: a "hybrid" model in which researchers work close to engineering and product teams rather than in a separate corporate laboratory [6]. It publishes hundreds of papers a year, runs faculty and student programs, and releases open-source software and datasets [5]. Jeff Dean, a Google Brain co-founder and the division's former leader, serves as Google's chief scientist, a role that spans both Google Research and Google DeepMind [1].
Since the April 2024 reorganization, the group's work has been organized around three pillars [2][5]:
| Pillar | Representative research areas |
|---|---|
| Applied AI and the sciences | Earth AI, health AI, science AI, sustainability and crisis resilience |
| Foundational machine learning and algorithms | Algorithms and theory, information retrieval, machine intelligence, machine perception, natural language processing |
| People, systems, and quantum | Human-computer interaction and visualization, networking, quantum AI, responsible AI, anti-abuse, software engineering and software systems |
Responsible AI remains a listed research area, although the teams working most directly on model safety moved to Google DeepMind in April 2024 [2][5].
Since April 2024 the division of labor has been explicit: Google DeepMind builds and scales Google's AI models, including the Gemini family, while Google Research pursues foundational computer science and applied scientific research [2]. Announcing the consolidation that created Google DeepMind in 2023, chief executive Sundar Pichai wrote that "Combining all this talent into one focused team, backed by the computational resources of Google, will significantly accelerate our progress in AI" [1]. A year later he gave Google Research what he called a mandate to "continue investing in foundational and applied computer science research in three key areas that tie directly to Google's mission: computing systems (including quantum), foundational ML and algorithms, and applied science and society" [2].
The boundary is porous in practice. Jeff Dean's chief scientist role spans both organizations [1]; health projects such as AMIE, the Med-PaLM lineage, and MedGemma are staffed jointly [26][27][29]; and the two groups co-author papers and share infrastructure such as TPUs. In weather and climate the programs are complementary, with Google DeepMind's GraphCast and GenCast forecasting models alongside Google Research's NeuralGCM and flood forecasting systems [35]. The historical lineages also interleave: the Transformer, TensorFlow, and word2vec came from teams now inside Google DeepMind, while quantum computing, which predates the merger, stayed with Google Research [1][2][11].
The arrangement leaves Google with two research organizations of different character. DeepMind carries the frontier model and artificial general intelligence agenda alongside its own scientific programs such as AlphaFold, while Google Research anchors the company's longer-horizon scientific bets, from quantum hardware to flood forecasting, and feeds results into products across Search, Cloud, Android, and beyond [1][2][5].
Google built a research function early in its history. Peter Norvig, who joined in 2001, directed search quality from 2002 to 2005 and then became the company's director of research [7]. Alfred Spector, vice president of research and special initiatives from 2007 to 2015, codified the group's philosophy with Norvig and Slav Petrov in a July 2012 Communications of the ACM article, "Google's Hybrid Approach to Research," which argued that embedding researchers in product teams let Google run experiments at a scale no isolated laboratory could match [6][8]. The division's early output shaped the wider industry: systems papers such as MapReduce (2004) and Bigtable (2006), written by Google engineers and published through its research arm, inspired open-source counterparts including Apache Hadoop.
Google Brain began in 2011 as a part-time collaboration between Google engineers Jeff Dean and Greg Corrado and Stanford professor Andrew Ng, incubated inside the Google X moonshot lab [9]. In June 2012 the team drew global attention when a network running on 16,000 processors taught itself to recognize cats from 10 million unlabeled YouTube frames [9][10]. The project's value was immediate; X head Astro Teller later said Brain alone "paid for the entire cost" of Google X [9]. Brain soon moved into Google's research organization, and in March 2013 Google hired Geoffrey Hinton and acquired his startup DNNresearch [9].
Brain's output defined the deep learning era at Google: it open-sourced TensorFlow in November 2015, rebuilt Google Translate around neural machine translation in 2016, and produced generative projects such as Magenta and the text-to-image model Imagen [9]. Its most consequential paper, "Attention Is All You Need" (June 2017), introduced the Transformer architecture that underpins modern large language models; the paper's eight authors held affiliations split between Google Brain and Google Research, making the Transformer a joint product of both lineages [11]. All eight authors later left Google, several to found AI startups including Character.AI, Cohere, and Sakana AI [12].
From 2016, Google's search and AI efforts reported jointly to John Giannandrea. When he left for Apple in April 2018, the portfolio was split: Ben Gomes took search, and Jeff Dean, until then Brain's leader, took charge of AI [13]. A month later Google renamed its research division Google AI, retiring Google Research as the umbrella brand [14]. The new name never fully displaced the old one; the division kept publishing under the Google Research banner, and by the early 2020s Google Research was again the organization's name, with Dean as its senior vice president. In its final years as a unified group it produced the conversational model LaMDA (2021), which later powered the first version of the Bard chatbot, and the 540-billion-parameter PaLM (2022) [15].
On April 20, 2023, Sundar Pichai announced that the Brain team would leave Google Research and merge with DeepMind to form Google DeepMind, led by DeepMind co-founder Demis Hassabis; Dean became Google's chief scientist, reporting to Pichai and serving both organizations [1]. The rest of Google Research moved under James Manyika, whose role expanded to president for Research, Labs, Technology & Society, with a continuing remit covering "algorithms and theory, privacy and security, quantum computing, health, climate and sustainability and responsible AI" [1][4].
A second reorganization followed on April 18, 2024. Pichai consolidated all AI model building, including teams that had remained in Google Research, inside Google DeepMind, and moved Research's responsible AI teams there as well, writing that the company was "moving Responsible AI teams in Research to Google DeepMind, to be closer to where the models are built and scaled"; other responsibility teams moved to Google's central trust and safety organization [2]. Google Research received what Pichai called "a clear and distinct mandate" for foundational and applied computer science, and Yossi Matias, who had led Google's Israel center since 2006 along with global AI efforts in health, climate, and crisis response, was named its head, relocating from Tel Aviv to Mountain View [2][3]. The drift of product-facing AI toward DeepMind continued that October, when the Gemini app team also moved from Google's knowledge and information group into Google DeepMind [16].
Google Quantum AI traces its origins to 2012, when Hartmut Neven founded Google's quantum computing effort, which went on to run the Quantum Artificial Intelligence Lab jointly with NASA and the Universities Space Research Association; today quantum falls within Google Research's computing systems remit [2][17][18]. In October 2019 the 53-qubit Sycamore processor performed a sampling computation in 200 seconds that Google estimated would take the fastest classical supercomputer 10,000 years, the first claimed demonstration of quantum supremacy [19]; IBM disputed the estimate, and improved classical algorithms later closed much of the gap [18]. On December 9, 2024, the team announced Willow, a 105-qubit processor that achieved the first "below threshold" quantum error correction: logical error rates roughly halved each time the encoded qubit array grew, from 3 by 3 to 5 by 5 to 7 by 7 [20][21]. Willow also completed a random circuit sampling benchmark in about five minutes that Google estimated would take a leading supercomputer 10 septillion (10^25) years [20]. On October 22, 2025, the group reported the first verifiable quantum advantage with its Quantum Echoes algorithm, published in Nature, which measured out-of-time-order correlators about 13,000 times faster than the best known classical algorithm on a top supercomputer and demonstrated a molecular-structure application with the University of California, Berkeley [22].
Google Research's health portfolio, led since 2016 by Brain co-founder Greg Corrado as head of health AI, grew out of mid-2010s medical imaging work, including a 2016 study in JAMA showing that deep learning could detect diabetic retinopathy in retinal photographs [23][24]. With DeepMind it developed Med-PaLM, announced in December 2022 as the first AI system to exceed the pass mark on US Medical Licensing Examination style questions in the MedQA benchmark, with the study published in Nature in July 2023, and Med-PaLM 2, which reached 86.5 percent on the same benchmark in 2023 [25][26]. The jointly built diagnostic research system AMIE (Articulate Medical Intelligence Explorer) outperformed board-certified primary care physicians on 30 of 32 specialist-rated axes in simulated text consultations, in a randomized, double-blind study published in Nature in April 2025 [27]; a multimodal version built on Gemini 2.0 Flash added the ability to request and interpret images such as skin photographs and ECGs [28]. In May 2025 Google Research and Google DeepMind released MedGemma, open medical models built on Gemma 3 within the Health AI Developer Foundations collection [29]. Related human-centered work includes accessibility projects such as Project Euphonia, which adapts speech recognition to impaired speech.
Google Research runs an AI flood forecasting program whose public Flood Hub expanded in November 2024 to river basins in more than 100 countries, covering areas home to 700 million people, up from 80 countries and about 460 million [30]. The underlying global river model, described in Nature in 2024, provides forecasts up to seven days ahead with multi-day reliability comparable to the best same-day nowcasts, including in ungauged basins, and has been opened to outside researchers and partners [31]. In wildfire work, Google Research helped create the FireSat program with satellite maker Muon Space and the nonprofit Earth Fire Alliance, supported by $13 million from Google.org; the first prototype satellite launched on a SpaceX rideshare on March 14, 2025, toward a planned constellation of more than 50 satellites able to detect fires as small as 5 by 5 meters within about 20 minutes [32][33]. With American Airlines and Breakthrough Energy, Google Research showed in 2023 that pilots using its AI contrail forecasts could cut contrail formation by 54 percent across 70 test flights [34], and with the European Centre for Medium-Range Weather Forecasts it built NeuralGCM, a hybrid physics and machine learning atmospheric model published in Nature in 2024 [35]. Other efforts include wildfire boundary tracking in Google's products, the Project Green Light traffic signal optimization initiative, and the Earth AI family of geospatial models [5].
Foundational work continues in algorithms, theory, privacy, and systems. Google researchers introduced federated learning in 2016 to 2017, letting models such as Gboard's keyboard predictions train across devices without centralizing user data [36]. The group also developed speculative decoding, a now industry-standard technique that accelerates large language model inference by drafting tokens with a small model and verifying them with a larger one [37]. Differential privacy tooling, anti-abuse research, and networking and software systems work round out the portfolio [5]. Newer architecture research includes the Titans long-term memory models and the Nested Learning paradigm for continual learning, presented with its proof-of-concept Hope architecture at NeurIPS 2025 [38]. The organization also runs longer-horizon moonshots: Project Suncatcher, announced in November 2025, is exploring constellations of solar-powered satellites carrying Tensor Processing Units, with two prototype satellites planned with Planet for launch by early 2027 [39].
The table below lists selected outputs of Google's research organizations, including the Brain era; Google counts the pre-2023 Brain lineage among the accomplishments that moved to Google DeepMind [1].
| Year | Contribution | Origin | Significance |
|---|---|---|---|
| 2013 | word2vec | Google Brain | Efficient word embeddings that helped launch modern representation learning [40] |
| 2015 | TensorFlow | Google Brain | Open-source machine learning framework [9] |
| 2016 | Federated learning | Google Research | Privacy-preserving on-device training, first deployed in Gboard [36] |
| 2017 | Transformer | Google Brain and Google Research | Architecture underpinning modern large language models [11] |
| 2018 | BERT | Google AI | Bidirectional pretraining that reset natural language understanding benchmarks [41] |
| 2019 | Sycamore experiment | Google Quantum AI | First claimed demonstration of quantum supremacy [19] |
| 2022 | Med-PaLM | Google Research and DeepMind | First AI past the pass mark on USMLE-style MedQA questions [26] |
| 2022 | Speculative decoding | Google Research | Widely adopted technique for faster LLM inference [37] |
| 2024 | Global flood forecasting model | Google Research | Nature-published model extending reliable river flood warnings to seven days [31] |
| 2024 | Willow | Google Quantum AI | First below-threshold quantum error correction [20] |
| 2025 | AMIE | Google Research and Google DeepMind | Conversational diagnostic AI evaluated against physicians in Nature [27] |
| 2025 | Quantum Echoes | Google Quantum AI | First verifiable quantum advantage claim, about 13,000x faster than classical [22] |