search research articles

An official website of the United States government

The .gov means it’s official. Federal government websites often end in .gov or .mil. Before sharing sensitive information, make sure you’re on a federal government site.

The site is secure. The https:// ensures that you are connecting to the official website and that any information you provide is encrypted and transmitted securely.

• Publications
• Account settings

Trending Articles

• Global burden and strength of evidence for 88 risk factors in 204 countries and 811 subnational locations, 1990-2021: a systematic analysis for the Global Burden of Disease Study 2021. GBD 2021 Risk Factors Collaborators. Lancet. 2024. PMID: 38762324
• Burden of disease scenarios for 204 countries and territories, 2022-2050: a forecasting analysis for the Global Burden of Disease Study 2021. GBD 2021 Forecasting Collaborators. Lancet. 2024. PMID: 38762325
• A modern way to teach and practice manual therapy. Kerry R, et al. Chiropr Man Therap. 2024. PMID: 38773515 Review.
• An age-progressive platelet differentiation path from hematopoietic stem cells causes exacerbated thrombosis. Poscablo DM, et al. Cell. 2024. PMID: 38749423
• ARID1A suppresses R-loop-mediated STING-type I interferon pathway activation of anti-tumor immunity. Maxwell MB, et al. Cell. 2024. PMID: 38754421

Latest Literature

• Cochrane Database Syst Rev (3)
• J Immunol (2)
• Nature (12)
• PLoS One (83)
• Pediatrics (2)
• Proc Natl Acad Sci U S A (12)

NCBI Literature Resources

MeSH PMC Bookshelf Disclaimer

The PubMed wordmark and PubMed logo are registered trademarks of the U.S. Department of Health and Human Services (HHS). Unauthorized use of these marks is strictly prohibited.

Explore millions of high-quality primary sources and images from around the world, including artworks, maps, photographs, and more.

Explore migration issues through a variety of media types

• Part of The Streets are Talking: Public Forms of Creative Expression from Around the World
• Part of The Journal of Economic Perspectives, Vol. 34, No. 1 (Winter 2020)
• Part of Cato Institute (Aug. 3, 2021)
• Part of University of California Press
• Part of Open: Smithsonian National Museum of African American History & Culture
• Part of Indiana Journal of Global Legal Studies, Vol. 19, No. 1 (Winter 2012)
• Part of R Street Institute (Nov. 1, 2020)
• Part of Leuven University Press
• Part of UN Secretary-General Papers: Ban Ki-moon (2007-2016)
• Part of Perspectives on Terrorism, Vol. 12, No. 4 (August 2018)
• Part of Leveraging Lives: Serbia and Illegal Tunisian Migration to Europe, Carnegie Endowment for International Peace (Mar. 1, 2023)
• Part of UCL Press

Harness the power of visual materials—explore more than 3 million images now on JSTOR.

Enhance your scholarly research with underground newspapers, magazines, and journals.

Explore collections in the arts, sciences, and literature from the world’s leading museums, archives, and scholars.

200 million monthly downloads

24 million monthly readers

3 million authors submit annually

SpringerLink - Home for all research

Discover open access.

• Publish with us

• Track your research

Featured articles and journals

Browse by subject, about springerlink.

Calls for papers

Log in for personalised recommendations.

Ageing in Place

The world’s population is ageing, and with ageing comes an increased risk of disability, multimorbidity and dementia, and an increased need for support. Older people are motivated to stay in their own homes as they age as an alternative to intramural care that is cost-beneficial and often provides...

Urban Lifelines and Supply Chains: Enhancing Resilience and Sustainability with Foundational Technologies

An urban community is an intricate network connected by a myriad of functional systems and subsystems, including civil structures, underground spaces, multi-modal transportation, telecom, energy grids, retail facilities, food systems, and healthcare facilities, all operating with unique spatial and...

Photons to Fuels: Recent Progress of Photocatalytic For CO2 Reduction and H2 Production

To address the detrimental effects of climate change, innovative and timely approaches in renewable energy and environmental chemistry are essential. Photocatalytic conversion of carbon dioxide has garnered significant attention as a viable method for carbon capture and utilization, offering the...

Trending research

Social connection and mortality in UK Biobank: a prospective cohort analysis

From clicks to consequences: a multi-method review of online grocery shopping

Highly efficient flexible structured metasurface by roll-to-roll printing for diurnal radiative cooling

Persistent cognitive slowing in post-covid patients: longitudinal study over 6 months.

Blame it on my youth: the origins of attitudes towards immigration

Creating the ICU of the future: patient-centred design to optimise recovery

Featured journals

SN Applied Sciences is now Discover Applied Sciences! We are excited to announce that SN Applied Sciences moved into our fully OA Discover journal...

Discover Sustainability is an open access journal publishing research across all fields relevant to sustainability. Average number of article...

The Journal of Epidemiology and Global Healthis an international peer reviewed journal which aims to impact global epidemiology and international...

Featured books

As part of Springer Nature, SpringerLink delivers fast access to the depth and breadth of our online collection of journals, eBooks, reference works and protocols across a vast range of subject disciplines.

SpringerLink is the reading platform of choice for hundreds of thousands of researchers worldwide. Find out how to publish your research with Springer Nature .

• Find a journal

A free, AI-powered research tool for scientific literature

• Brenda Milner
• Metamorphic Rock
• Thermal Expansion

New & Improved API for Developers

Introducing semantic reader in beta.

Stay Connected With Semantic Scholar Sign Up What Is Semantic Scholar? Semantic Scholar is a free, AI-powered research tool for scientific literature, based at the Allen Institute for AI.

The top list of academic search engines

1. Google Scholar

4. science.gov, 5. semantic scholar, 6. baidu scholar, get the most out of academic search engines, frequently asked questions about academic search engines, related articles.

Academic search engines have become the number one resource to turn to in order to find research papers and other scholarly sources. While classic academic databases like Web of Science and Scopus are locked behind paywalls, Google Scholar and others can be accessed free of charge. In order to help you get your research done fast, we have compiled the top list of free academic search engines.

Google Scholar is the clear number one when it comes to academic search engines. It's the power of Google searches applied to research papers and patents. It not only lets you find research papers for all academic disciplines for free but also often provides links to full-text PDF files.

• Coverage: approx. 200 million articles
• Abstracts: only a snippet of the abstract is available
• Related articles: ✔
• References: ✔
• Cited by: ✔
• Links to full text: ✔
• Export formats: APA, MLA, Chicago, Harvard, Vancouver, RIS, BibTeX

BASE is hosted at Bielefeld University in Germany. That is also where its name stems from (Bielefeld Academic Search Engine).

• Coverage: approx. 136 million articles (contains duplicates)
• Abstracts: ✔
• Related articles: ✘
• References: ✘
• Cited by: ✘
• Export formats: RIS, BibTeX

CORE is an academic search engine dedicated to open-access research papers. For each search result, a link to the full-text PDF or full-text web page is provided.

• Coverage: approx. 136 million articles
• Links to full text: ✔ (all articles in CORE are open access)
• Export formats: BibTeX

Science.gov is a fantastic resource as it bundles and offers free access to search results from more than 15 U.S. federal agencies. There is no need anymore to query all those resources separately!

• Coverage: approx. 200 million articles and reports
• Links to full text: ✔ (available for some databases)
• Export formats: APA, MLA, RIS, BibTeX (available for some databases)

Semantic Scholar is the new kid on the block. Its mission is to provide more relevant and impactful search results using AI-powered algorithms that find hidden connections and links between research topics.

• Coverage: approx. 40 million articles
• Export formats: APA, MLA, Chicago, BibTeX

Although Baidu Scholar's interface is in Chinese, its index contains research papers in English as well as Chinese.

• Coverage: no detailed statistics available, approx. 100 million articles
• Abstracts: only snippets of the abstract are available
• Export formats: APA, MLA, RIS, BibTeX

RefSeek searches more than one billion documents from academic and organizational websites. Its clean interface makes it especially easy to use for students and new researchers.

• Coverage: no detailed statistics available, approx. 1 billion documents
• Abstracts: only snippets of the article are available
• Export formats: not available

Consider using a reference manager like Paperpile to save, organize, and cite your references. Paperpile integrates with Google Scholar and many popular databases, so you can save references and PDFs directly to your library using the Paperpile buttons:

Google Scholar is an academic search engine, and it is the clear number one when it comes to academic search engines. It's the power of Google searches applied to research papers and patents. It not only let's you find research papers for all academic disciplines for free, but also often provides links to full text PDF file.

Semantic Scholar is a free, AI-powered research tool for scientific literature developed at the Allen Institute for AI. Sematic Scholar was publicly released in 2015 and uses advances in natural language processing to provide summaries for scholarly papers.

BASE , as its name suggest is an academic search engine. It is hosted at Bielefeld University in Germany and that's where it name stems from (Bielefeld Academic Search Engine).

CORE is an academic search engine dedicated to open access research papers. For each search result a link to the full text PDF or full text web page is provided.

Science.gov is a fantastic resource as it bundles and offers free access to search results from more than 15 U.S. federal agencies. There is no need any more to query all those resources separately!

An official website of the United States government

The .gov means it’s official. Federal government websites often end in .gov or .mil. Before sharing sensitive information, make sure you’re on a federal government site.

The site is secure. The https:// ensures that you are connecting to the official website and that any information you provide is encrypted and transmitted securely.

• Publications
• Account settings

Preview improvements coming to the PMC website in October 2024. Learn More or Try it out now .

• Advanced Search
• Journal List
• J Med Libr Assoc
• v.106(4); 2018 Oct

A systematic approach to searching: an efficient and complete method to develop literature searches

Associated data.

Creating search strategies for systematic reviews, finding the best balance between sensitivity and specificity, and translating search strategies between databases is challenging. Several methods describe standards for systematic search strategies, but a consistent approach for creating an exhaustive search strategy has not yet been fully described in enough detail to be fully replicable. The authors have established a method that describes step by step the process of developing a systematic search strategy as needed in the systematic review. This method describes how single-line search strategies can be prepared in a text document by typing search syntax (such as field codes, parentheses, and Boolean operators) before copying and pasting search terms (keywords and free-text synonyms) that are found in the thesaurus. To help ensure term completeness, we developed a novel optimization technique that is mainly based on comparing the results retrieved by thesaurus terms with those retrieved by the free-text search words to identify potentially relevant candidate search terms. Macros in Microsoft Word have been developed to convert syntaxes between databases and interfaces almost automatically. This method helps information specialists in developing librarian-mediated searches for systematic reviews as well as medical and health care practitioners who are searching for evidence to answer clinical questions. The described method can be used to create complex and comprehensive search strategies for different databases and interfaces, such as those that are needed when searching for relevant references for systematic reviews, and will assist both information specialists and practitioners when they are searching the biomedical literature.

INTRODUCTION

Librarians and information specialists are often involved in the process of preparing and completing systematic reviews (SRs), where one of their main tasks is to identify relevant references to include in the review [ 1 ]. Although several recommendations for the process of searching have been published [ 2 – 6 ], none describe the development of a systematic search strategy from start to finish.

Traditional methods of SR search strategy development and execution are highly time consuming, reportedly requiring up to 100 hours or more [ 7 , 8 ]. The authors wanted to develop systematic and exhaustive search strategies more efficiently, while preserving the high sensitivity that SR search strategies necessitate. In this article, we describe the method developed at Erasmus University Medical Center (MC) and demonstrate its use through an example search. The efficiency of the search method and outcome of 73 searches that have resulted in published reviews are described in a separate article [ 9 ].

As we aimed to describe the creation of systematic searches in full detail, the method starts at a basic level with the analysis of the research question and the creation of search terms. Readers who are new to SR searching are advised to follow all steps described. More experienced searchers can consider the basic steps to be existing knowledge that will already be part of their normal workflow, although step 4 probably differs from general practice. Experienced searchers will gain the most from reading about the novelties in the method as described in steps 10–13 and comparing the examples given in the supplementary appendix to their own practice.

CREATING A SYSTEMATIC SEARCH STRATEGY

Our methodology for planning and creating a multi-database search strategy consists of the following steps:

• Determine a clear and focused question
• Describe the articles that can answer the question
• Decide which key concepts address the different elements of the question
• Decide which elements should be used for the best results
• Choose an appropriate database and interface to start with
• Document the search process in a text document
• Identify appropriate index terms in the thesaurus of the first database
• Identify synonyms in the thesaurus
• Add variations in search terms
• Use database-appropriate syntax, with parentheses, Boolean operators, and field codes
• Optimize the search
• Evaluate the initial results
• Check for errors
• Translate to other databases
• Test and reiterate

Each step in the process is reflected by an example search described in the supplementary appendix .

1. Determine a clear and focused question

A systematic search can best be applied to a well-defined and precise research or clinical question. Questions that are too broad or too vague cannot be answered easily in a systematic way and will generally result in an overwhelming number of search results. On the other hand, a question that is too specific will result into too few or even zero search results. Various papers describe this process in more detail [ 10 – 12 ].

2. Describe the articles that can answer the question

Although not all clinical or research questions can be answered in the literature, the next step is to presume that the answer can indeed be found in published studies. A good starting point for a search is hypothesizing what the research that can answer the question would look like. These hypothetical (when possible, combined with known) articles can be used as guidance for constructing the search strategy.

3. Decide which key concepts address the different elements of the question

Key concepts are the topics or components that the desired articles should address, such as diseases or conditions, actions, substances, settings, domains (e.g., therapy, diagnosis, etiology), or study types. Key concepts from the research question can be grouped to create elements in the search strategy.

Elements in a search strategy do not necessarily follow the patient, intervention, comparison, outcome (PICO) structure or any other related structure. Using the PICO or another similar framework as guidance can be helpful to consider, especially in the inclusion and exclusion review stage of the SR, but this is not necessary for good search strategy development [ 13 – 15 ]. Sometimes concepts from different parts of the PICO structure can be grouped together into one search element, such as when the desired outcome is frequently described in a certain study type.

4. Decide which elements should be used for the best results

Not all elements of a research question should necessarily be used in the search strategy. Some elements are less important than others or may unnecessarily complicate or restrict a search strategy. Adding an element to a search strategy increases the chance of missing relevant references. Therefore, the number of elements in a search strategy should remain as low as possible to optimize recall.

Using the schema in Figure 1 , elements can be ordered by their specificity and importance to determine the best search approach. Whether an element is more specific or more general can be measured objectively by the number of hits retrieved in a database when searching for a key term representing that element. Depending on the research question, certain elements are more important than others. If articles (hypothetically or known) exist that can answer the question but lack a certain element in their titles, abstracts, or keywords, that element is unimportant to the question. An element can also be unimportant because of expected bias or an overlap with another element.

Schema for determining the optimal order of elements

Bias in elements

The choice of elements in a search strategy can introduce bias through use of overly specific terminology or terms often associated with positive outcomes. For the question “does prolonged breastfeeding improve intelligence outcomes in children?,” searching specifically for the element of duration will introduce bias, as articles that find a positive effect of prolonged breastfeeding will be much more likely to mention time factors in their titles or abstracts.

Overlapping elements

Elements in a question sometimes overlap in their meaning. Sometimes certain therapies are interventions for one specific disease. The Lichtenstein technique, for example, is a repair method for inguinal hernias. There is no need to include an element of “inguinal hernias” to a search for the effectiveness of the Lichtenstein therapy. Likewise, sometimes certain diseases are only found in certain populations. Adding such an overlapping element could lead to missing relevant references.

The elements to use in a search strategy can be found in the plot of elements in Figure 1 , by following the top row from left to right. For this method, we recommend starting with the most important and specific elements. Then, continue with more general and important elements until the number of results is acceptable for screening. Determining how many results are acceptable for screening is often a matter of negotiation with the SR team.

5. Choose an appropriate database and interface to start with

Important factors for choosing databases to use are the coverage and the presence of a thesaurus. For medically oriented searches, the coverage and recall of Embase, which includes the MEDLINE database, are superior to those of MEDLINE [ 16 ]. Each of these two databases has its own thesaurus with its own unique definitions and structure. Because of the complexity of the Embase thesaurus, Emtree, which contains much more specific thesaurus terms than the MEDLINE Medical Subject Headings (MeSH) thesaurus, translation from Emtree to MeSH is easier than the other way around. Therefore, we recommend starting in Embase.

MEDLINE and Embase are available through many different vendors and interfaces. The choice of an interface and primary database is often determined by the searcher’s accessibility. For our method, an interface that allows searching with proximity operators is desirable, and full functionality of the thesaurus, including explosion of narrower terms, is crucial. We recommend developing a personal workflow that always starts with one specific database and interface.

6. Document the search process in a text document

We advise designing and creating the complete search strategies in a log document, instead of directly in the database itself, to register the steps taken and to make searches accountable and reproducible. The developed search strategies can be copied and pasted into the desired databases from the log document. This way, the searcher is in control of the whole process. Any change to the search strategy should be done in the log document, assuring that the search strategy in the log is always the most recent.

7. Identify appropriate index terms in the thesaurus of the first database

Searches should start by identifying appropriate thesaurus terms for the desired elements. The thesaurus of the database is searched for matching index terms for each key concept. We advise restricting the initial terms to the most important and most relevant terms. Later in the process, more general terms can be added in the optimization process, in which the effect on the number of hits, and thus the desirability of adding these terms, can be evaluated more easily.

Several factors can complicate the identification of thesaurus terms. Sometimes, one thesaurus term is found that exactly describes a specific element. In contrast, especially in more general elements, multiple thesaurus terms can be found to describe one element. If no relevant thesaurus terms have been found for an element, free-text terms can be used, and possible thesaurus terms found in the resulting references can be added later (step 11).

Sometimes, no distinct thesaurus term is available for a specific key concept that describes the concept in enough detail. In Emtree, one thesaurus term often combines two or more elements. The easiest solution for combining these terms for a sensitive search is to use such a thesaurus term in all elements where it is relevant. Examples are given in the supplementary appendix .

8. Identify synonyms in the thesaurus

Most thesauri offer a list of synonyms on their term details page (named Synonyms in Emtree and Entry Terms in MeSH). To create a sensitive search strategy for SRs, these terms need to be searched as free-text keywords in the title and abstract fields, in addition to searching their associated thesaurus terms.

The Emtree thesaurus contains more synonyms (300,000) than MeSH does (220,000) [ 17 ]. The difference in number of terms is even higher considering that many synonyms in MeSH are permuted terms (i.e., inversions of phrases using commas).

Thesaurus terms are ordered in a tree structure. When searching for a more general thesaurus term, the more specific (narrower) terms in the branches below that term will also be searched (this is frequently referred to as “exploding” a thesaurus term). However, to perform a sensitive search, all relevant variations of the narrower terms must be searched as free-text keywords in the title or abstract, in addition to relying on the exploded thesaurus term. Thus, all articles that describe a certain narrower topic in their titles and abstracts will already be retrieved before MeSH terms are added.

9. Add variations in search terms (e.g., truncation, spelling differences, abbreviations, opposites)

Truncation allows a searcher to search for words beginning with the same word stem. A search for therap* will, thus, retrieve therapy, therapies, therapeutic, and all other words starting with “therap.” Do not truncate a word stem that is too short. Also, limitations of interfaces should be taken into account, especially in PubMed, where the number of search term variations that can be found by truncation is limited to 600.

Databases contain references to articles using both standard British and American English spellings. Both need to be searched as free-text terms in the title and abstract. Alternatively, many interfaces offer a certain code to replace zero or one characters, allowing a search for “pediatric” or “paediatric” as “p?ediatric.” Table 1 provides a detailed description of the syntax for different interfaces.

Field codes in five most used interfaces for biomedical literature searching

Searching for abbreviations can identify extra, relevant references and retrieve more irrelevant ones. The search can be more focused by combining the abbreviation with an important word that is relevant to its meaning or by using the Boolean “NOT” to exclude frequently observed, clearly irrelevant results. We advise that searchers do not exclude all possible irrelevant meanings, as it is very time consuming to identify all the variations, it will result in unnecessarily complicated search strategies, and it may lead to erroneously narrowing the search and, thereby, reduce recall.

Searching partial abbreviations can be useful for retrieving relevant references. For example, it is very likely that an article would mention osteoarthritis (OA) early in the abstract, replacing all further occurrences of osteoarthritis with OA . Therefore, it may not contain the phrase “hip osteoarthritis” but only “hip oa.”

It is also important to search for the opposites of search terms to avoid bias. When searching for “disease recurrence,” articles about “disease free” may be relevant as well. When the desired outcome is survival , articles about mortality may be relevant.

10. Use database-appropriate syntax, with parentheses, Boolean operators, and field codes

Different interfaces require different syntaxes, the special set of rules and symbols unique to each database that define how a correctly constructed search operates. Common syntax components include the use of parentheses and Boolean operators such as “AND,” “OR,” and “NOT,” which are available in all major interfaces. An overview of different syntaxes for four major interfaces for bibliographic medical databases (PubMed, Ovid, EBSCOhost, Embase.com, and ProQuest) is shown in Table 1 .

Creating the appropriate syntax for each database, in combination with the selected terms as described in steps 7–9, can be challenging. Following the method outlined below simplifies the process:

• Create single-line queries in a text document (not combining multiple record sets), which allows immediate checking of the relevance of retrieved references and efficient optimization.
• Type the syntax (Boolean operators, parentheses, and field codes) before adding terms, which reduces the chance that errors are made in the syntax, especially in the number of parentheses.
• Use predefined proximity structures including parentheses, such as (() ADJ3 ()) in Ovid, that can be reused in the query when necessary.
• Use thesaurus terms separately from free-text terms of each element. Start an element with all thesaurus terms (using “OR”) and follow with the free-text terms. This allows the unique optimization methods as described in step 11.
• When adding terms to an existing search strategy, pay close attention to the position of the cursor. Make sure to place it appropriately either in the thesaurus terms section, in the title/abstract section, or as an addition (broadening) to an existing proximity search.

The supplementary appendix explains the method of building a query in more detail, step by step for different interfaces: PubMed, Ovid, EBSCOhost, Embase.com, and ProQuest. This method results in a basic search strategy designed to retrieve some relevant references upon which a more thorough search strategy can be built with optimization such as described in step 11.

11. Optimize the search

The most important question when performing a systematic search is whether all (or most) potentially relevant articles have been retrieved by the search strategy. This is also the most difficult question to answer, since it is unknown which and how many articles are relevant. It is, therefore, wise first to broaden the initial search strategy, making the search more sensitive, and then check if new relevant articles are found by comparing the set results (i.e., search for Strategy #2 NOT Strategy #1 to see the unique results).

A search strategy should be tested for completeness. Therefore, it is necessary to identify extra, possibly relevant search terms and add them to the test search in an OR relationship with the already used search terms. A good place to start, and a well-known strategy, is scanning the top retrieved articles when sorted by relevance, looking for additional relevant synonyms that could be added to the search strategy.

We have developed a unique optimization method that has not been described before in the literature. This method often adds valuable extra terms to our search strategy and, therefore, extra, relevant references to our search results. Extra synonyms can be found in articles that have been assigned a certain set of thesaurus terms but that lack synonyms in the title and/or abstract that are already present in the current search strategy. Searching for thesaurus terms NOT free-text terms will help identify missed free-text terms in the title or abstract. Searching for free-text terms NOT thesaurus terms will help identify missed thesaurus terms. If this is done repeatedly for each element, leaving the rest of the query unchanged, this method will help add numerous relevant terms to the query. These steps are explained in detail for five different search platforms in the supplementary appendix .

12. Evaluate the initial results

The results should now contain relevant references. If the interface allows relevance ranking, use that in the evaluation. If you know some relevant references that should be included in the research, search for those references specifically; for example, combine a specific (first) author name with a page number and the publication year. Check whether those references are retrieved by the search. If the known relevant references are not retrieved by the search, adapt the search so that they are. If it is unclear which element should be adapted to retrieve a certain article, combine that article with each element separately.

Different outcomes are desired for different types of research questions. For instance, in the case of clinical question answering, the researcher will not be satisfied with many references that contain a lot of irrelevant references. A clinical search should be rather specific and is allowed to miss a relevant reference. In the case of an SR, the researchers do not want to miss any relevant reference and are willing to handle many irrelevant references to do so. The search for references to include in an SR should be very sensitive: no included reference should be missed. A search that is too specific or too sensitive for the intended goal can be adapted to become more sensitive or specific. Steps to increase sensitivity or specificity of a search strategy can be found in the supplementary appendix .

13. Check for errors

Errors might not be easily detected. Sometimes clues can be found in the number of results, either when the number of results is much higher or lower than expected or when many retrieved references are not relevant. However, the number expected is often unknown, and very sensitive search strategies will always retrieve many irrelevant articles. Each query should, therefore, be checked for errors.

One of the most frequently occurring errors is missing the Boolean operator “OR.” When no “OR” is added between two search terms, many interfaces automatically add an “AND,” which unintentionally reduces the number of results and likely misses relevant references. One good strategy to identify missing “OR”s is to go to the web page containing the full search strategy, as translated by the database, and using Ctrl-F search for “AND.” Check whether the occurrences of the “AND” operator are deliberate.

Ideally, search strategies should be checked by other information specialists [ 18 ]. The Peer Review of Electronic Search Strategies (PRESS) checklist offers good guidance for this process [ 4 ]. Apart from the syntax (especially Boolean operators and field codes) of the search strategy, it is wise to have the search terms checked by the clinician or researcher familiar with the topic. At Erasmus MC, researchers and clinicians are involved during the complete process of structuring and optimizing the search strategy. Each word is added after the combined decision of the searcher and the researcher, with the possibility of directly comparing results with and without the new term.

14. Translate to other databases

To retrieve as many relevant references as possible, one has to search multiple databases. Translation of complex and exhaustive queries between different databases can be very time consuming and cumbersome. The single-line search strategy approach detailed above allows quick translations using the find and replace method in Microsoft Word (<Ctrl-H>).

At Erasmus MC, macros based on the find-and-replace method in Microsoft Word have been developed for easy and fast translation between the most used databases for biomedical and health sciences questions. The schema that is followed for the translation between databases is shown in Figure 2 . Most databases simply follow the structure set by the Embase.com search strategy. The translation from Emtree terms to MeSH terms for MEDLINE in Ovid often identifies new terms that need to be added to the Embase.com search strategy before the translation to other databases.

Schematic representation of translation between databases used at Erasmus University Medical Center

Dotted lines represent databases that are used in less than 80% of the searches.

Using five different macros, a thoroughly optimized query in Embase.com can be relatively quickly translated into eight major databases. Basic search strategies will be created to use in many, mostly smaller, databases, because such niche databases often do not have extensive thesauri or advanced syntax options. Also, there is not much need to use extensive syntax because the number of hits and, therefore, the amount of noise in these databases is generally low. In MEDLINE (Ovid), PsycINFO (Ovid), and CINAHL (EBSCOhost), the thesaurus terms must be adapted manually, as each database has its own custom thesaurus. These macros and instructions for their installation, use, and adaptation are available at bit.ly/databasemacros.

15. Test and reiterate

Ideally, exhaustive search strategies should retrieve all references that are covered in a specific database. For SR search strategies, checking searches for their recall is advised. This can be done after included references have been determined by the authors of the systematic review. If additional papers have been identified through other non-database methods (i.e., checking references in included studies), results that were not identified by the database searches should be examined. If these results were available in the databases but not located by the search strategy, the search strategy should be adapted to try to retrieve these results, as they may contain terms that were omitted in the original search strategies. This may enable the identification of additional relevant results.

A methodology for creating exhaustive search strategies has been created that describes all steps of the search process, starting with a question and resulting in thorough search strategies in multiple databases. Many of the steps described are not new, but together, they form a strong method creating high-quality, robust searches in a relatively short time frame.

Our methodology is intended to create thoroughness for literature searches. The optimization method, as described in step 11, will identify missed synonyms or thesaurus terms, unlike any other method that largely depends on predetermined keywords and synonyms. Using this method results in a much quicker search process, compared to traditional methods, especially because of the easier translation between databases and interfaces (step 13). The method is not a guarantee for speed, since speed depends on many factors, including experience. However, by following the steps and using the tools as described above, searchers can gain confidence first and increase speed through practice.

What is new?

This method encourages searchers to start their search development process using empty syntax first and later adding the thesaurus terms and free-text synonyms. We feel this helps the searcher to focus on the search terms, instead of on the structure of the search query. The optimization method in which new terms are found in the already retrieved articles is used in some other institutes as well but has to our knowledge not been described in the literature. The macros to translate search strategies between interfaces are unique in this method.

What is different compared to common practice?

Traditionally, librarians and information specialists have focused on creating complex, multi-line (also called line-by-line) search strategies, consisting of multiple record sets, and this method is frequently advised in the literature and handbooks [ 2 , 19 – 21 ]. Our method, instead, uses single-line searches, which is critical to its success. Single-line search strategies can be easily adapted by adding or dropping a term without having to recode numbers of record sets, which would be necessary in multi-line searches. They can easily be saved in a text document and repeated by copying and pasting for search updates. Single-line search strategies also allow easy translation to other syntaxes using find-and-replace technology to update field codes and other syntax elements or using macros (step 13).

When constructing a search strategy, the searcher might experience that certain parentheses in the syntax are unnecessary, such as parentheses around all search terms in the title/abstract portion, if there is only one such term, there are double parentheses in the proximity statement, or one of the word groups exists for only one word. One might be tempted to omit those parentheses for ease of reading and management. However, during the optimization process, the searcher is likely to find extra synonyms that might consist of one word. To add those terms to the first query (with reduced parentheses) requires adding extra parentheses (meticulously placing and counting them), whereas, in the latter search, it only requires proper placement of those terms.

Many search methods highly depend on the PICO framework. Research states that often PICO or PICOS is not suitable for every question [ 22 , 23 ]. There are other acronyms than PICO—such as sample, phenomenon of interest, design, evaluation, research type (SPIDER) [ 24 ]—but each is just a variant. In our method, the most important and specific elements of a question are being analyzed for building the best search strategy.

Though it is generally recommended that searchers search both MEDLINE and Embase, most use MEDLINE as the starting point. It is considered the gold standard for biomedical searching, partially due to historical reasons, since it was the first of its kind, and more so now that it is freely available via the PubMed interface. Our method can be used with any database as a starting point, but we use Embase instead of MEDLINE or another database for a number of reasons. First, Embase provides both unique content and the complete content of MEDLINE. Therefore, searching Embase will be, by definition, more complete than searching MEDLINE only. Second, the number of terms in Emtree (the Embase thesaurus) is three times as high as that of MeSH (the MEDLINE thesaurus). It is easier to find MeSH terms after all relevant Emtree terms have been identified than to start with MeSH and translate to Emtree.

At Erasmus MC, the researchers sit next to the information specialist during most of the search strategy design process. This way, the researchers can deliver immediate feedback on the relevance of proposed search terms and retrieved references. The search team then combines knowledge about databases with knowledge about the research topic, which is an important condition to create the highest quality searches.

Limitations of the method

One disadvantage of single-line searches compared to multi-line search strategies is that errors are harder to recognize. However, with the methods for optimization as described (step 11), errors are recognized easily because missed synonyms and spelling errors will be identified during the process. Also problematic is that more parentheses are needed, making it more difficult for the searcher and others to assess the logic of the search strategy. However, as parentheses and field codes are typed before the search terms are added (step 10), errors in parentheses can be prevented.

Our methodology works best if used in an interface that allows proximity searching. It is recommended that searchers with access to an interface with proximity searching capabilities select one of those as the initial database to develop and optimize the search strategy. Because the PubMed interface does not allow proximity searches, phrases or Boolean “AND” combinations are required. Phrase searching complicates the process and is more specific, with the higher risk of missing relevant articles, and using Boolean “AND” combinations increases sensitivity but at an often high loss of specificity. Due to some searchers’ lack of access to expensive databases or interfaces, the freely available PubMed interface may be necessary to use, though it should never be the sole database used for an SR [ 2 , 16 , 25 ]. A limitation of our method is that it works best with subscription-based and licensed resources.

Another limitation is the customization of the macros to a specific institution’s resources. The macros for the translation between different database interfaces only work between the interfaces as described. To mitigate this, we recommend using the find-and-replace functionality of text editors like Microsoft Word to ease the translation of syntaxes between other databases. Depending on one’s institutional resources, custom macros can be developed using similar methods.

Results of the method

Whether this method results in exhaustive searches where no important article is missed is difficult to determine, because the number of relevant articles is unknown for any topic. A comparison of several parameters of 73 published reviews that were based on a search developed with this method to 258 reviews that acknowledged information specialists from other Dutch academic hospitals shows that the performance of the searches following our method is comparable to those performed in other institutes but that the time needed to develop the search strategies was much shorter than the time reported for the other reviews [ 9 ].

CONCLUSIONS

With the described method, searchers can gain confidence in their search strategies by finding many relevant words and creating exhaustive search strategies quickly. The approach can be used when performing SR searches or for other purposes such as answering clinical questions, with different expectations of the search’s precision and recall. This method, with practice, provides a stepwise approach that facilitates the search strategy development process from question clarification to final iteration and beyond.

SUPPLEMENTAL FILE

Acknowledgments.

We highly appreciate the work that was done by our former colleague Louis Volkers, who in his twenty years as an information specialist in Erasmus MC laid the basis for our method. We thank Professor Oscar Franco for reviewing earlier drafts of this article.

• SUGGESTED TOPICS
• The Magazine
• Newsletters
• Managing Yourself
• Managing Teams
• Work-life Balance
• The Big Idea
• Data & Visuals
• Reading Lists
• Case Selections
• HBR Learning
• Topic Feeds
• Account Settings
• Email Preferences

Research: What Companies Don’t Know About How Workers Use AI

• Jeremie Brecheisen

Three Gallup studies shed light on when and why AI is being used at work — and how employees and customers really feel about it.

Leaders who are exploring how AI might fit into their business operations must not only navigate a vast and ever-changing landscape of tools, but they must also facilitate a significant cultural shift within their organizations. But research shows that leaders do not fully understand their employees’ use of, and readiness for, AI. In addition, a significant number of Americans do not trust business’ use of AI. This article offers three recommendations for leaders to find the right balance of control and trust around AI, including measuring how their employees currently use AI, cultivating trust by empowering managers, and adopting a purpose-led AI strategy that is driven by the company’s purpose instead of a rules-heavy strategy that is driven by fear.

If you’re a leader who wants to shift your workforce toward using AI, you need to do more than manage the implementation of new technologies. You need to initiate a profound cultural shift. At the heart of this cultural shift is trust. Whether the use case for AI is brief and experimental or sweeping and significant, a level of trust must exist between leaders and employees for the initiative to have any hope of success.

• Jeremie Brecheisen is a partner and managing director of The Gallup CHRO Roundtable.

Partner Center

Suggestions or feedback?

MIT News | Massachusetts Institute of Technology

• Machine learning
• Social justice
• Black holes
• Classes and programs

Departments

• Aeronautics and Astronautics
• Brain and Cognitive Sciences
• Architecture
• Political Science
• Mechanical Engineering

Centers, Labs, & Programs

• Abdul Latif Jameel Poverty Action Lab (J-PAL)
• Picower Institute for Learning and Memory
• Lincoln Laboratory
• School of Architecture + Planning
• School of Engineering
• School of Humanities, Arts, and Social Sciences
• Sloan School of Management
• School of Science
• MIT Schwarzman College of Computing

Using ideas from game theory to improve the reliability of language models

Press contact :.

Previous image Next image

Imagine you and a friend are playing a game where your goal is to communicate secret messages to each other using only cryptic sentences. Your friend's job is to guess the secret message behind your sentences. Sometimes, you give clues directly, and other times, your friend has to guess the message by asking yes-or-no questions about the clues you've given. The challenge is that both of you want to make sure you're understanding each other correctly and agreeing on the secret message.

MIT Computer Science and Artificial Intelligence Laboratory (CSAIL) researchers have created a similar "game" to help improve how AI understands and generates text. It is known as a “consensus game” and it involves two parts of an AI system — one part tries to generate sentences (like giving clues), and the other part tries to understand and evaluate those sentences (like guessing the secret message).

The researchers discovered that by treating this interaction as a game, where both parts of the AI work together under specific rules to agree on the right message, they could significantly improve the AI's ability to give correct and coherent answers to questions. They tested this new game-like approach on a variety of tasks, such as reading comprehension, solving math problems, and carrying on conversations, and found that it helped the AI perform better across the board.

Traditionally, large language models answer one of two ways: generating answers directly from the model (generative querying) or using the model to score a set of predefined answers (discriminative querying), which can lead to differing and sometimes incompatible results. With the generative approach, "Who is the president of the United States?" might yield a straightforward answer like "Joe Biden." However, a discriminative query could incorrectly dispute this fact when evaluating the same answer, such as "Barack Obama."

So, how do we reconcile mutually incompatible scoring procedures to achieve coherent, efficient predictions? 

"Imagine a new way to help language models understand and generate text, like a game. We've developed a training-free, game-theoretic method that treats the whole process as a complex game of clues and signals, where a generator tries to send the right message to a discriminator using natural language. Instead of chess pieces, they're using words and sentences," says Athul Jacob, an MIT PhD student in electrical engineering and computer science and CSAIL affiliate. "Our way to navigate this game is finding the 'approximate equilibria,' leading to a new decoding algorithm called 'equilibrium ranking.' It's a pretty exciting demonstration of how bringing game-theoretic strategies into the mix can tackle some big challenges in making language models more reliable and consistent."

When tested across many tasks, like reading comprehension, commonsense reasoning, math problem-solving, and dialogue, the team's algorithm consistently improved how well these models performed. Using the ER algorithm with the LLaMA-7B model even outshone the results from much larger models. "Given that they are already competitive, that people have been working on it for a while, but the level of improvements we saw being able to outperform a model that's 10 times the size was a pleasant surprise," says Jacob. 

"Diplomacy," a strategic board game set in pre-World War I Europe, where players negotiate alliances, betray friends, and conquer territories without the use of dice — relying purely on skill, strategy, and interpersonal manipulation — recently had a second coming. In November 2022, computer scientists, including Jacob, developed “Cicero,” an AI agent that achieves human-level capabilities in the mixed-motive seven-player game, which requires the same aforementioned skills, but with natural language. The math behind this partially inspired the Consensus Game. 

While the history of AI agents long predates when OpenAI's software entered the chat in November 2022, it's well documented that they can still cosplay as your well-meaning, yet pathological friend. 

The consensus game system reaches equilibrium as an agreement, ensuring accuracy and fidelity to the model's original insights. To achieve this, the method iteratively adjusts the interactions between the generative and discriminative components until they reach a consensus on an answer that accurately reflects reality and aligns with their initial beliefs. This approach effectively bridges the gap between the two querying methods. 

In practice, implementing the consensus game approach to language model querying, especially for question-answering tasks, does involve significant computational challenges. For example, when using datasets like MMLU, which have thousands of questions and multiple-choice answers, the model must apply the mechanism to each query. Then, it must reach a consensus between the generative and discriminative components for every question and its possible answers. 

The system did struggle with a grade school right of passage: math word problems. It couldn't generate wrong answers, which is a critical component of understanding the process of coming up with the right one. 

“The last few years have seen really impressive progress in both strategic decision-making and language generation from AI systems, but we’re just starting to figure out how to put the two together. Equilibrium ranking is a first step in this direction, but I think there’s a lot we’ll be able to do to scale this up to more complex problems,” says Jacob.   

An avenue of future work involves enhancing the base model by integrating the outputs of the current method. This is particularly promising since it can yield more factual and consistent answers across various tasks, including factuality and open-ended generation. The potential for such a method to significantly improve the base model's performance is high, which could result in more reliable and factual outputs from ChatGPT and similar language models that people use daily. 

"Even though modern language models, such as ChatGPT and Gemini, have led to solving various tasks through chat interfaces, the statistical decoding process that generates a response from such models has remained unchanged for decades," says Google Research Scientist Ahmad Beirami, who was not involved in the work. "The proposal by the MIT researchers is an innovative game-theoretic framework for decoding from language models through solving the equilibrium of a consensus game. The significant performance gains reported in the research paper are promising, opening the door to a potential paradigm shift in language model decoding that may fuel a flurry of new applications."

Jacob wrote the paper with MIT-IBM Watson Lab researcher Yikang Shen and MIT Department of Electrical Engineering and Computer Science assistant professors Gabriele Farina and Jacob Andreas, who is also a CSAIL member. They presented their work at the International Conference on Learning Representations (ICLR) earlier this month, where it was highlighted as a "spotlight paper." The research also received a “best paper award” at the NeurIPS R0-FoMo Workshop in December 2023.

Share this news article on:

Press mentions, quanta magazine.

MIT researchers have developed a new procedure that uses game theory to improve the accuracy and consistency of large language models (LLMs), reports Steve Nadis for Quanta Magazine . “The new work, which uses games to improve AI, stands in contrast to past approaches, which measured an AI program’s success via its mastery of games,” explains Nadis. 

Previous item Next item

Related Links

• Article: "Game Theory Can Make AI More Correct and Efficient"
• Jacob Andreas
• Athul Paul Jacob
• Language & Intelligence @ MIT
• Computer Science and Artificial Intelligence Laboratory (CSAIL)
• Department of Electrical Engineering and Computer Science
• MIT-IBM Watson AI Lab

Related Topics

• Computer science and technology
• Artificial intelligence
• Human-computer interaction
• Natural language processing
• Game theory
• Electrical Engineering & Computer Science (eecs)

Related Articles

Reasoning and reliability in AI

Explained: Generative AI

Synthetic imagery sets new bar in AI training efficiency

Simulating discrimination in virtual reality

More mit news.

Turning up the heat on next-generation semiconductors

Read full story →

Sarah Millholland receives 2024 Vera Rubin Early Career Award

A community collaboration for progress

MIT scholars will take commercial break with entrepreneurial scholarship

MIT scientists learn how to control muscles with light

Adhesive coatings can prevent scarring around medical implants

• More news on MIT News homepage →

Massachusetts Institute of Technology 77 Massachusetts Avenue, Cambridge, MA, USA

• Map (opens in new window)
• Events (opens in new window)
• People (opens in new window)
• Careers (opens in new window)
• Accessibility
• Social Media Hub
• MIT on Facebook
• MIT on YouTube
• MIT on Instagram

To revisit this article, visit My Profile, then View saved stories .

• Backchannel
• Newsletters
• WIRED Insider
• WIRED Consulting

Will Knight

OpenAI’s Long-Term AI Risk Team Has Disbanded

In July last year, OpenAI announced the formation of a new research team that would prepare for the advent of supersmart artificial intelligence capable of outwitting and overpowering its creators. Ilya Sutskever, OpenAI’s chief scientist and one of the company’s cofounders, was named as the colead of this new team. OpenAI said the team would receive 20 percent of its computing power.

Now OpenAI’s “superalignment team” is no more, the company confirms. That comes after the departures of several researchers involved, Tuesday’s news that Sutskever was leaving the company, and the resignation of the team’s other colead. The group’s work will be absorbed into OpenAI’s other research efforts.

Sutskever’s departure made headlines because although he’d helped CEO Sam Altman start OpenAI in 2015 and set the direction of the research that led to ChatGPT, he was also one of the four board members who fired Altman in November . Altman was restored as CEO five chaotic days later after a mass revolt by OpenAI staff and the brokering of a deal in which Sutskever and two other company directors left the board . Hours after Sutskever’s departure was announced on Tuesday, Jan Leike, the former DeepMind researcher who was the superalignment team’s other colead, posted on X that he had resigned.

Neither Sutskever nor Leike responded to requests for comment. Sutskever did not offer an explanation for his decision to leave but offered support for OpenAI’s current path in a post on X . “The company’s trajectory has been nothing short of miraculous, and I’m confident that OpenAI will build AGI that is both safe and beneficial” under its current leadership, he wrote.

Leike posted a thread on X on Friday explaining that his decision came from a disagreement over the company’s priorities and how much resources his team was being allocated.

“I have been disagreeing with OpenAI leadership about the company's core priorities for quite some time, until we finally reached a breaking point,” Leike wrote. “Over the past few months my team has been sailing against the wind. Sometimes we were struggling for compute and it was getting harder and harder to get this crucial research done.”

The dissolution of OpenAI’s superalignment team adds to recent evidence of a shakeout inside the company in the wake of last November’s governance crisis. Two researchers on the team, Leopold Aschenbrenner and Pavel Izmailov, were dismissed for leaking company secrets, The Information reported last month. Another member of the team, William Saunders, left OpenAI in February, according to an internet forum post in his name.

Two more OpenAI researchers working on AI policy and governance also appear to have left the company recently. Cullen O'Keefe left his role as research lead on policy frontiers in April, according to LinkedIn. Daniel Kokotajlo, an OpenAI researcher who has coauthored several papers on the dangers of more capable AI models, “quit OpenAI due to losing confidence that it would behave responsibly around the time of AGI,” according to a posting on an internet forum in his name. None of the researchers who have apparently left responded to requests for comment.

By Carlton Reid

By Emily Mullin

By Steven Levy

By Andy Greenberg

OpenAI declined to comment on the departures of Sutskever or other members of the superalignment team, or the future of its work on long-term AI risks. Research on the risks associated with more powerful models will now be led by John Schulman, who coleads the team responsible for fine-tuning AI models after training.

The superalignment team was not the only team pondering the question of how to keep AI under control, although it was publicly positioned as the main one working on the most far-off version of that problem. The blog post announcing the superalignment team last summer stated: “Currently, we don't have a solution for steering or controlling a potentially superintelligent AI, and preventing it from going rogue.” OpenAI’s charter binds it to safely developing so-called artificial general intelligence, or technology that rivals or exceeds humans, safely and for the benefit of humanity. Sutskever and other leaders there have often spoken about the need to proceed cautiously. But OpenAI has also been early to develop and publicly release experimental AI projects to the public.

OpenAI was once unusual among prominent AI labs for the eagerness with which research leaders like Sutskever talked of creating superhuman AI and of the potential for such technology to turn on humanity. That kind of doomy AI talk became much more widespread last year, after ChatGPT turned OpenAI into the most prominent and closely-watched technology company on the planet. As researchers and policymakers wrestled with the implications of ChatGPT and the prospect of vastly more capable AI, it became less controversial to worry about AI harming humans or humanity as a whole .

The existential angst has since cooled—and AI has yet to make another massive leap—but the need for AI regulation remains a hot topic. And this week OpenAI showcased a new version of ChatGPT that could once again change people’s relationship with the technology in powerful and perhaps problematic new ways.

The departures of Sutskever and Leike come shortly after OpenAI’s latest big reveal—a new “multimodal” AI model called GPT-4o that allows ChatGPT to see the world and converse in a more natural and humanlike way. A livestreamed demonstration showed the new version of ChatGPT mimicking human emotions and even attempting to flirt with users. OpenAI has said it will make the new interface available to paid users within a couple of weeks.

There is no indication that the recent departures have anything to do with OpenAI’s efforts to develop more humanlike AI or to ship products. But the latest advances do raise ethical questions around privacy, emotional manipulation, and cybersecurity risks. OpenAI maintains another research group called the Preparedness team which focuses on these issues.

Update 5/17/24 12:23 pm ET: This story has been updated to include comments from posts on X by Jan Leike.

You Might Also Like …

In your inbox: Will Knight's Fast Forward explores advances in AI

Indian voters are being bombarded with millions of deepfakes

They bought tablets in prison —and found a broken promise

The one thing that’s holding back the heat pump

It's always sunny: Here are the best sunglasses for every adventure

Reece Rogers

Kate Knibbs

Steven Levy

We've detected unusual activity from your computer network

To continue, please click the box below to let us know you're not a robot.

Why did this happen?

Please make sure your browser supports JavaScript and cookies and that you are not blocking them from loading. For more information you can review our Terms of Service and Cookie Policy .

For inquiries related to this message please contact our support team and provide the reference ID below.

• My View My View
• Following Following
• Saved Saved

China should hike tariffs on large cars to 25%, says research body

• Medium Text

• Company Bayerische Motoren Werke AG Follow
• Company Carparts.Com Inc Follow
• Company China Automotive Technology and Research Center Co., Ltd. Follow

Sign up here.

Reporting by Philip Blenkinsop, additional reporting by Sarah Wu and Zoey Zhang in Beijing; Editing by Elaine Hardcastle

Our Standards: The Thomson Reuters Trust Principles. New Tab , opens new tab

Business Chevron

Nvidia forecasts quarterly revenue above estimates, unveils stock split

Nvidia forecast quarterly revenue above estimates on Wednesday and announced a stock split, lifting its shares to record-high territory and impressing investors who have tripled the chipmaker's market value in the past year on AI optimism.

Thank you for visiting nature.com. You are using a browser version with limited support for CSS. To obtain the best experience, we recommend you use a more up to date browser (or turn off compatibility mode in Internet Explorer). In the meantime, to ensure continued support, we are displaying the site without styles and JavaScript.

• View all journals
• Explore content
• About the journal
• Publish with us
• Sign up for alerts
• Published: 14 May 2024

2023 summer warmth unparalleled over the past 2,000 years

• Jan Esper   ORCID: orcid.org/0000-0003-3919-014X 1 , 2 ,
• Max Torbenson   ORCID: orcid.org/0000-0003-2720-2238 1 &
• Ulf Büntgen 2 , 3 , 4  

Nature ( 2024 ) Cite this article

5844 Accesses

3408 Altmetric

Metrics details

We are providing an unedited version of this manuscript to give early access to its findings. Before final publication, the manuscript will undergo further editing. Please note there may be errors present which affect the content, and all legal disclaimers apply.

• Climate change
• Palaeoclimate

Including an exceptionally warm Northern Hemisphere (NH) summer 1 ,2 , 2023 has been reported as the hottest year on record 3-5 . Contextualizing recent anthropogenic warming against past natural variability is nontrivial, however, because the sparse 19 th century meteorological records tend to be too warm 6 . Here, we combine observed and reconstructed June-August (JJA) surface air temperatures to show that 2023 was the warmest NH extra-tropical summer over the past 2000 years exceeding the 95% confidence range of natural climate variability by more than half a degree Celsius. Comparison of the 2023 JJA warming against the coldest reconstructed summer in 536 CE reveals a maximum range of pre-Anthropocene-to-2023 temperatures of 3.93°C. Although 2023 is consistent with a greenhouse gases-induced warming trend 7 that is amplified by an unfolding El Niño event 8 , this extreme emphasizes the urgency to implement international agreements for carbon emission reduction.

This is a preview of subscription content, access via your institution

Access options

Access Nature and 54 other Nature Portfolio journals

Get Nature+, our best-value online-access subscription

24,99 € / 30 days

cancel any time

Subscribe to this journal

Receive 51 print issues and online access

185,98 € per year

only 3,65 € per issue

Rent or buy this article

Prices vary by article type

Prices may be subject to local taxes which are calculated during checkout

Similar content being viewed by others

The economic commitment of climate change

Near-synchronous Northern Hemisphere and Patagonian Ice Sheet variation over the last glacial cycle

Urban development pattern’s influence on extreme rainfall occurrences

Author information, authors and affiliations.

Department of Geography, Johannes Gutenberg University, Mainz, Germany

Jan Esper & Max Torbenson

Global Change Research Institute of the Czech Academy of Sciences, Brno, Czech Republic

Jan Esper & Ulf Büntgen

Department of Geography, University of Cambridge, Cambridge, United Kingdom

Ulf Büntgen

Department of Geography, Masaryk University, Brno, Czech Republic

You can also search for this author in PubMed   Google Scholar

Corresponding author

Correspondence to Jan Esper .

Rights and permissions

Reprints and permissions

About this article

Cite this article.

Esper, J., Torbenson, M. & Büntgen, U. 2023 summer warmth unparalleled over the past 2,000 years. Nature (2024). https://doi.org/10.1038/s41586-024-07512-y

Download citation

Received : 16 January 2024

Accepted : 02 May 2024

Published : 14 May 2024

DOI : https://doi.org/10.1038/s41586-024-07512-y

Share this article

Anyone you share the following link with will be able to read this content:

Sorry, a shareable link is not currently available for this article.

Provided by the Springer Nature SharedIt content-sharing initiative

By submitting a comment you agree to abide by our Terms and Community Guidelines . If you find something abusive or that does not comply with our terms or guidelines please flag it as inappropriate.

Quick links

• Explore articles by subject
• Guide to authors
• Editorial policies

Sign up for the Nature Briefing newsletter — what matters in science, free to your inbox daily.