Explore the vast range of titles available.

The need to make objectives explicit may expose policy trade-offs that had previously been implicit and obscured There is a growing demand to be able to “explain” machine learning (ML) systems' decisions and actions to human users, particularly when used in contexts where decisions have substantial implications for those affected and where there is a requirement for political accountability or legal compliance ( 1 ). Explainability is often discussed as a technical challenge in designing ML systems and decision procedures, to improve understanding of what is typically a “black box” phenomenon. But some of the most difficult challenges are nontechnical and raise questions about the broader accountability of organizations using ML in their decision-making. One reason for this is that many decisions by ML systems may exhibit bias, as systemic biases in society lead to biases in data used by the systems ( 2 ). But there is another reason, less widely appreciated. Because the quantities that ML systems seek to optimize have to be specified by their users, explainable ML will force policy-makers to be more explicit about their objectives, and thus about their values and political choices, exposing policy trade-offs that may have previously only been implicit and obscured. As the use of ML in policy spreads, there may have to be public debate that makes explicit the value judgments or weights to be used. Merely technical approaches to “explaining” ML will often only be effective if the systems are deployed by trustworthy and accountable organizations.

What do we want from machine intelligence? We envision machines that are not just tools for thought but partners in thought: reasonable, insightful, knowledgeable, reliable and trustworthy systems that think with us. Current artificial intelligence systems satisfy some of these criteria, some of the time. In this Perspective, we show how the science of collaborative cognition can be put to work to engineer systems that really can be called ‘thought partners’, systems built to meet our expectations and complement our limitations. We lay out several modes of collaborative thought in which humans and artificial intelligence thought partners can engage, and we propose desiderata for human-compatible thought partnerships. Drawing on motifs from computational cognitive science, we motivate an alternative scaling path for the design of thought partners and ecosystems around their use through a Bayesian lens, whereby the partners we construct actively build and reason over models of the human and world. In this Perspective, the authors advance a view for the science of collaborative cognition to engineer systems that can be considered thought partners, systems built to meet our expectations and complement our limitations.

The world is about to be swamped by an unprecedented wave of AI-generated content. We need reliable ways of identifying such content, to supplement the many existing social institutions that enable trust between people and organisations and ensure social resilience. In this paper, we begin by highlighting an important new development: providers of AI content generators have new obligations to support the creation of reliable detectors for the content they generate. These new obligations arise mainly from the EU’s newly finalised AI Act, but they are enhanced by the US President’s recent Executive Order on AI, and by several considerations of self-interest. These new steps towards reliable detection mechanisms are by no means a panacea—but we argue they will usher in a new adversarial landscape, in which reliable methods for identifying AI-generated content are commonly available. In this landscape, many new questions arise for policymakers. Firstly, if reliable AI-content detection mechanisms are available, who should be required to use them? And how should they be used? We argue that new duties arise for media and Web search companies arise for media companies, and for Web search companies, in the deployment of AI-content detectors. Secondly, what broader regulation of the tech ecosystem will maximise the likelihood of reliable AI-content detectors? We argue for a range of new duties, relating to provenance-authentication protocols, open-source AI generators, and support for research and enforcement. Along the way, we consider how the production of AI-generated content relates to ‘free expression’, and discuss the important case of content that is generated jointly by humans and AIs.

Artificial intelligence provides a promising solution for streamlining COVID-19 diagnoses; however, concerns surrounding security and trustworthiness impede the collection of large-scale representative medical data, posing a considerable challenge for training a well-generalized model in clinical practices. To address this, we launch the Unified CT-COVID AI Diagnostic Initiative (UCADI), where the artificial intelligence (AI) model can be distributedly trained and independently executed at each host institution under a federated learning framework without data sharing. Here we show that our federated learning framework model considerably outperformed all of the local models (with a test sensitivity/specificity of 0.973/0.951 in China and 0.730/0.942 in the United Kingdom), achieving comparable performance with a panel of professional radiologists. We further evaluated the model on the hold-out (collected from another two hospitals without the federated learning framework) and heterogeneous (acquired with contrast materials) data, provided visual explanations for decisions made by the model, and analysed the trade-offs between the model performance and the communication costs in the federated training process. Our study is based on 9,573 chest computed tomography scans from 3,336 patients collected from 23 hospitals located in China and the United Kingdom. Collectively, our work advanced the prospects of utilizing federated learning for privacy-preserving AI in digital health. The COVID-19 pandemic sparked the need for international collaboration in using clinical data for rapid development of diagnosis and treatment methods. But the sensitive nature of medical data requires special care and ideally potentially sensitive data would not leave the organization which collected it. Xiang Bai and colleagues present a privacy-preserving AI framework for CT-based COVID-19 diagnosis and demonstrate it on data from 23 hospitals in China and the United Kingdom.

Graphical models provide a flexible, powerful and compact way to model relationships between random variables, and have been applied with great success in many domains. Combining prior beliefs with observed evidence to form a prediction is called inference. Problems of great interest include finding a configuration with highest probability (MAP inference) or solving for the distribution over a subset of variables (marginal inference). Further, these methods are often critical subroutines for learning the relationships. However, inference is computationally intractable in general. Hence, much effort has focused on two themes: finding subdomains where exact inference is solvable efficiently, or identifying approximate methods that work well. We explore both these themes, restricting attention to undirected graphical models with discrete variables. First we address exact MAP inference by advancing the recent method of reducing the problem to finding a maximum weight stable set (MWSS) on a derived graph, which, if perfect, admits polynomial time inference. We derive new results for this approach, including a general decomposition theorem for models of any order and number of labels, extensions of results for binary pairwise models with submodular cost functions to higher order, and a characterization of which binary pairwise models can be efficiently solved with this method. This clarifies the power of the approach on this class of models, improves our toolbox and provides insight into the range of tractable models. Next we consider methods of approximate inference, with particular emphasis on the Bethe approximation, which is in widespread use and has proved remarkably effective, yet is still far from being completely understood. We derive new formulations and properties of the derivatives of the Bethe free energy, then use these to establish an algorithm to compute log of the optimum Bethe partition function to arbitrary ε-accuracy. Further, if the model is attractive, we demonstrate a fully polynomial-time approximation scheme (FPTAS), which is an important theoretical result, and demonstrate its practical applications. Next we explore ways to tease apart the two aspects of the Bethe approximation, i.e. the polytope relaxation and the entropy approximation. We derive analytic results, show how optimization may be explored over various polytopes in practice, even for large models, and remark on the observed performance compared to the true distribution and the tree-reweighted (TRW) approximation. This reveals important novel observations and helps guide inference in practice. Finally, we present results related to clamping a selection of variables in a model. We derive novel lower bounds on an array of approximate partition functions based only on the model’s topology. Further, we show that in an attractive binary pairwise model, clamping any variable and summing over the approximate sub-partition functions can only increase (hence improve) the Bethe approximation, then use this to provide a new, short proof that the Bethe partition function lower bounds the true value for this class of models. The bulk of this work focuses on the class of binary, pairwise models, but several results apply more generally.

[This corrects the article DOI: 10.1038/s42256-021-00421-z.].

We present a survey of tools used in the criminal justice system in the UK in three categories: data infrastructure, data analysis, and risk prediction. Many tools are currently in deployment, offering potential benefits, including improved efficiency and consistency. However, there are also important concerns. Transparent information about these tools, their purpose, how they are used, and by whom is difficult to obtain. Even when information is available, it is often insufficient to enable a satisfactory evaluation. More work is needed to establish governance mechanisms to ensure that tools are deployed in a transparent, safe and ethical way. We call for more engagement with stakeholders and greater documentation of the intended goal of a tool, how it will achieve this goal compared to other options, and how it will be monitored in deployment. We highlight additional points to consider when evaluating the trustworthiness of deployed tools and make concrete proposals for policy.

With the capability to write convincing and fluent natural language and generate code, Foundation Models present dual-use concerns broadly and within the cyber domain specifically. Generative AI has already begun to impact cyberspace through a broad illicit marketplace for assisting malware development and social engineering attacks through hundreds of malicious-AI-as-a-services tools. More alarming is that recent research has shown the potential for these advanced models to inform or independently execute offensive cyberspace operations. However, these previous investigations primarily focused on the threats posed by proprietary models due to the until recent lack of strong open-weight model and additionally leave the impacts of network defenses or potential countermeasures unexplored. Critically, understanding the aptitude of downloadable models to function as offensive cyber agents is vital given that they are far more difficult to govern and prevent their misuse. As such, this work evaluates several state-of-the-art FMs on their ability to compromise machines in an isolated network and investigates defensive mechanisms to defeat such AI-powered attacks. Using target machines from a commercial provider, the most recently released downloadable models are found to be on par with a leading proprietary model at conducting simple cyber attacks with common hacking tools against known vulnerabilities. To mitigate such LLM-powered threats, defensive prompt injection (DPI) payloads for disrupting the malicious cyber agent's workflow are demonstrated to be effective. From these results, the implications for AI safety and governance with respect to cybersecurity is analyzed.