Explore the vast range of titles available.

The emergence of tools based on large language models (LLMs), such as OpenAI’s ChatGPT and Google’s Gemini, has garnered immense public attention owing to their advanced natural language generation capabilities. These remarkably natural-sounding tools have the potential to be highly useful for various tasks. However, they also tend to produce false, erroneous or misleading content—commonly referred to as hallucinations. Moreover, LLMs can be misused to generate convincing, yet false, content and profiles on a large scale, posing a substantial societal challenge by potentially deceiving users and spreading inaccurate information. This makes fact-checking increasingly important. Despite their issues with factual accuracy, LLMs have shown proficiency in various subtasks that support fact-checking, which is essential to ensure factually accurate responses. In light of these concerns, we explore issues related to factuality in LLMs and their impact on fact-checking. We identify key challenges, imminent threats and possible solutions to these factuality issues. We also thoroughly examine these challenges, existing solutions and potential prospects for fact-checking. By analysing the factuality constraints within LLMs and their impact on fact-checking, we aim to contribute to a path towards maintaining accuracy at a time of confluence of generative artificial intelligence and misinformation. Large language models (LLMs) present challenges, including a tendency to produce false or misleading content and the potential to create misinformation or disinformation. Augenstein and colleagues explore issues related to factuality in LLMs and their impact on fact-checking.

Language is a uniquely human trait at the core of human interactions. The language people use often reflects their personality, intentions and state of mind. With the integration of the Internet and social media into everyday life, much of human communication is documented as written text. These online forms of communication (for example, blogs, reviews, social media posts and emails) provide a window into human behaviour and therefore present abundant research opportunities for behavioural science. In this Review, we describe how natural language processing (NLP) can be used to analyse text data in behavioural science. First, we review applications of text data in behavioural science. Second, we describe the NLP pipeline and explain the underlying modelling approaches (for example, dictionary-based approaches and large language models). We discuss the advantages and disadvantages of these methods for behavioural science, in particular with respect to the trade-off between interpretability and accuracy. Finally, we provide actionable recommendations for using NLP to ensure rigour and reproducibility.

How is knowledge about word meaning represented in the mental lexicon? Current computational models infer word meanings from lexical co-occurrence patterns. They learn to represent words as vectors in a multidimensional space, wherein words that are used in more similar linguistic contexts—that is, are more semantically related—are located closer together. However, whereas inter-word proximity captures only overall relatedness, human judgements are highly context dependent. For example, dolphins and alligators are similar in size but differ in dangerousness. Here, we use a domain-general method to extract context-dependent relationships from word embeddings: ‘semantic projection’ of word-vectors onto lines that represent features such as size (the line connecting the words ‘small’ and ‘big’) or danger (‘safe’ to ‘dangerous’), analogous to ‘mental scales’. This method recovers human judgements across various object categories and properties. Thus, the geometry of word embeddings explicitly represents a wealth of context-dependent world knowledge.Grand, Blank, et al. show that context-dependent knowledge about objects, such as the similarities between animals in terms of size versus danger versus habitat, can be recovered from word embeddings via a simple, interpretable geometrical operation.

Language diversity is under threat. While each language is subject to specific social, demographic and political pressures, there may also be common threatening processes. We use an analysis of 6,511 spoken languages with 51 predictor variables spanning aspects of population, documentation, legal recognition, education policy, socioeconomic indicators and environmental features to show that, counter to common perception, contact with other languages per se is not a driver of language loss. However, greater road density, which may encourage population movement, is associated with increased endangerment. Higher average years of schooling is also associated with greater endangerment, evidence that formal education can contribute to loss of language diversity. Without intervention, language loss could triple within 40 years, with at least one language lost per month. To avoid the loss of over 1,500 languages by the end of the century, urgent investment is needed in language documentation, bilingual education programmes and other community-based programmes. Using a global analysis of 6,511 spoken languages with 51 predictor variables spanning aspects of population, documentation, legal recognition, education policy, socioeconomic indicators and environmental features, the authors identify predictors of current and future language endangerment and loss.

Foundation models have demonstrated remarkable performance across various tasks, primarily due to their abilities to comprehend instructions and access extensive, high-quality data. These capabilities showcase the effectiveness of current foundation models and suggest a promising trajectory. Owing to multiple constraints, such as the extreme scarcity or inaccessibility of raw data used to train foundation models and the high cost of training large-scale foundation models from scratch, the use of pre-existing foundation models or application programming interfaces for downstream tasks has become a new research trend, which we call Learn from Model (LFM). LFM involves extracting and leveraging prior knowledge from foundation models through fine-tuning, editing and fusion methods and applying it to downstream tasks. We emphasize that maximizing the use of parametric knowledge in data-scarce scenarios is critical to LFM. Analysing the LFM paradigm can guide the selection of the most appropriate technology in a given scenario to minimize parameter storage and computational costs while improving the performance of foundation models on new tasks. This Review provides a comprehensive overview of current methods based on foundation models from the perspective of LFM. Large general-purpose models are becoming more prevalent and useful, but also harder to train and find suitable training data for. Zheng et al. discuss how models can be used to train other models.

Formulating segments or strings of text of varying length flanked by pauses of different durations is integral to the writing and translation processes. In recent years, these strings of texts produced within pauses, thereafter referred to as bursts of written language, have become the object of detailed analysis in both writing and translation, two sub-categories of text production. Although the linguistic units contained within these bursts have been the subject of extensive research, to date, their analysis has mostly built on traditional grammatical categories, and their linguistic properties remain only partially defined. This paper suggests a new analytical framework that can be deployed for the analysis of pause placement in bursts of writing in translation at the product-process interface. The study builds on a corpus of product and process data of English-to-French translations of a biotechnology text by L1 French translation students. It focuses on coding and analyzing keylogging data corresponding to similar translation renditions to explore patterns in pause placement. The new analytical framework contributes to a better understanding of how meaning unfolds in translation and how pause placement patterns can indicate translation difficulty.

If the structure of language vocabularies mirrors the structure of natural divisions that are universally perceived, then the meanings of words in different languages should closely align. By contrast, if shared word meanings are a product of shared culture, history and geography, they may differ between languages in substantial but predictable ways. Here, we analysed the semantic neighbourhoods of 1,010 meanings in 41 languages. The most-aligned words were from semantic domains with high internal structure (number, quantity and kinship). Words denoting natural kinds, common actions and artefacts aligned much less well. Languages that are more geographically proximate, more historically related and/or spoken by more-similar cultures had more aligned word meanings. These results provide evidence that the meanings of common words vary in ways that reflect the culture, history and geography of their users. A comparison of 41 languages reveals that words for common actions, artefacts and natural kinds are less translatable than expected. Translatability is related to the cultural similarity of language communities and to their historic relationships.

Transformer models such as GPT generate human-like language and are predictive of human brain responses to language. Here, using functional-MRI-measured brain responses to 1,000 diverse sentences, we first show that a GPT-based encoding model can predict the magnitude of the brain response associated with each sentence. We then use the model to identify new sentences that are predicted to drive or suppress responses in the human language network. We show that these model-selected novel sentences indeed strongly drive and suppress the activity of human language areas in new individuals. A systematic analysis of the model-selected sentences reveals that surprisal and well-formedness of linguistic input are key determinants of response strength in the language network. These results establish the ability of neural network models to not only mimic human language but also non-invasively control neural activity in higher-level cortical areas, such as the language network. Tuckute et al. use a machine learning approach to identify sentences that either maximally or minimally activate the human language processing network.

This study develops and evaluates a Large Language Model (LLM) system for hint recognition in Chinese and Russian diplomatic discourse by integrating a semantic–cognitive–pragmatic theoretical framework with Retrieval-Augmented Generation (RAG) and Chain-of-Thought (CoT) reasoning. Grounded in linguistic–pragmatic theory and employing discourse analysis, we systematically annotated real press-conference transcripts from the Chinese and Russian Ministries of Foreign Affairs, constructed a vectorized external knowledge base covering textual and logical hints, embedded CoT reasoning instructions into the prompts, and provided bilingual few-shot exemplars to guide LLM recognition. Experimental results demonstrate stable overall performance with consistently high recall across both corpora, with the Russian dataset achieving higher precision and F1 scores than the Chinese dataset. Error analysis reveals three major types of systematic bias in LLM hint recognition: semantic over–interpretation, hint-type misclassification, and literal-meaning misclassification. To address these problems, we propose several targeted improvements, including expanding negative samples with standardized “no-hint” diplomatic expressions, strengthening context anchoring to ensure pragmatic interpretations are grounded in discourse, introducing a repeated matching mechanism, calibrating sensitive trigger words, and introducing a pre-filtering and self-evaluation mechanism to better distinguish explicit statements from implicit meanings. This study provides a feasible pathway and practical guidance for improving the accuracy and stability of LLMs in multilingual, high-context implicit information recognition.