I am honoured to present the International AI Safety Report. It is the work of 96 international AI experts who collaborated in an unprecedented effort to establish an internationally shared scientific understanding of risks from advanced AI and methods for managing them. We embarked on this journey just over a year ago, shortly after the countries present at the Bletchley Park AI Safety Summit agreed to support the creation of this report. Since then, we published an Interim Report in May 2024, which was presented at the AI Seoul Summit. We are now pleased to publish the present, full report ahead of the AI Action Summit in Paris in February 2025. Since the Bletchley Summit, the capabilities of general-purpose AI, the type of AI this report focuses on, have increased further. For example, new models have shown markedly better performance at tests of Professor Yoshua Bengio programming and scientific reasoning.

I am honoured to be chairing the delivery of the inaugural International Scientific Report on Advanced AI Safety. I am proud to publish this interim report which is the culmination of huge efforts by many experts over the six months since the work was commissioned at the Bletchley Park AI Safety Summit in November 2023. We know that advanced AI is developing very rapidly, and that there is considerable uncertainty over how these advanced AI systems might affect how we live and work in the future. AI has tremendous potential to change our lives for the better, but it also poses risks of harm. That is why having this thorough analysis of the available scientific literature and expert opinion is essential. The more we know, the better equipped we are to shape our collective destiny.

Exogenous state variables and rewards can slow reinforcement learning by injecting uncontrolled variation into the reward signal. This paper formalizes exogenous state variables and rewards and shows that if the reward function decomposes additively into endogenous and exogenous components, the MDP can be decomposed into an exogenous Markov Reward Process (based on the exogenous reward) and an endogenous Markov Decision Process (optimizing the endogenous reward). Any optimal policy for the endogenous MDP is also an optimal policy for the original MDP, but because the endogenous reward typically has reduced variance, the endogenous MDP is easier to solve. We study settings where the decomposition of the state space into exogenous and endogenous state spaces is not given but must be discovered. The paper introduces and proves correctness of algorithms for discovering the exogenous and endogenous subspaces of the state space when they are mixed through linear combination. These algorithms can be applied during reinforcement learning to discover the exogenous space, remove the exogenous reward, and focus reinforcement learning on the endogenous MDP. Experiments on a variety of challenging synthetic MDPs show that these methods, applied online, discover large exogenous state spaces and produce substantial speedups in reinforcement learning.

Trustworthy classifiers are essential to the adoption of machine learning predictions in many real-world settings. The predicted probability of possible outcomes can inform high-stakes decision making, particularly when assessing the expected value of alternative decisions or the risk of bad outcomes. These decisions require well-calibrated probabilities, not just the correct prediction of the most likely class. Black-box classifier calibration methods can improve the reliability of a classifier's output without requiring retraining. However, these methods are unable to detect subpopulations where calibration could also improve prediction accuracy. Such subpopulations are said to exhibit "hidden heterogeneity" (HH), because the original classifier did not detect them. This paper proposes a quantitative measure for HH. It also introduces two similarity-weighted calibration methods that can address HH by adapting locally to each test item: SWC weights the calibration set by similarity to the test item, and SWC-HH explicitly incorporates hidden heterogeneity to filter the calibration set. Experiments show that the improvements in calibration achieved by similarity-based calibration methods correlate with the amount of HH present and, given sufficient calibration data, generally exceed calibration achieved by global methods. HH can therefore serve as a useful diagnostic tool for identifying when local calibration methods would be beneficial.

As an autonomous system performs a task, it should maintain a calibrated estimate of the probability that it will achieve the user's goal. If that probability falls below some desired level, it should alert the user so that appropriate interventions can be made. This paper considers settings where the user's goal is specified as a target interval for a real-valued performance summary, such as the cumulative reward, measured at a fixed horizon $H$. At each time $t \in \{0, \ldots, H-1\}$, our method produces a calibrated estimate of the probability that the final cumulative reward will fall within a user-specified target interval $[y^-,y^+].$ Using this estimate, the autonomous system can raise an alarm if the probability drops below a specified threshold. We compute the probability estimates by inverting conformal prediction. Our starting point is the Conformalized Quantile Regression (CQR) method of Romano et al., which applies split-conformal prediction to the results of quantile regression. CQR is not invertible, but by using the conditional cumulative distribution function (CDF) as the non-conformity measure, we show how to obtain an invertible modification that we call \textbf{P}robability-space \textbf{C}onformalized \textbf{Q}uantile \textbf{R}egression (PCQR). Like CQR, PCQR produces well-calibrated conditional prediction intervals with finite-sample marginal guarantees. By inverting PCQR, we obtain marginal guarantees for the probability that the cumulative reward of an autonomous system will fall within an arbitrary user-specified target intervals. Experiments on two domains confirm that these probabilities are well-calibrated.

In many object recognition applications, the set of possible categories is an open set, and the deployed recognition system will encounter novel objects belonging to categories unseen during training. Detecting such "novel category" objects is usually formulated as an anomaly detection problem. Anomaly detection algorithms for feature-vector data identify anomalies as outliers, but outlier detection has not worked well in deep learning. Instead, methods based on the computed logits of visual object classifiers give state-of-the-art performance. This paper proposes the Familiarity Hypothesis that these methods succeed because they are detecting the absence of familiar learned features rather than the presence of novelty. This distinction is important, because familiarity-based detection will fail in many situations where novelty is present. For example when an image contains both a novel object and a familiar one, the familiarity score will be high, so the novel object will not be noticed. The paper reviews evidence from the literature and presents additional evidence from our own experiments that provide strong support for this hypothesis. The paper concludes with a discussion of whether familiarity-based detection is an inevitable consequence of representation learning.

Before delegating a task to an autonomous system, a human operator may want a guarantee about the behavior of the system. This paper extends previous work on conformal prediction for functional data and conformalized quantile regression to provide conformal prediction intervals over the future behavior of an autonomous system executing a fixed control policy on a Markov Decision Process (MDP). The prediction intervals are constructed by applying conformal corrections to prediction intervals computed by quantile regression. The resulting intervals guarantee that with probability $1-\delta$ the observed trajectory will lie inside the prediction interval, where the probability is computed with respect to the starting state distribution and the stochasticity of the MDP. The method is illustrated on MDPs for invasive species management and StarCraft2 battles.

Deep learning approaches to anomaly detection have recently improved the state of the art in detection performance on complex datasets such as large collections of images or text. These results have sparked a renewed interest in the anomaly detection problem and led to the introduction of a great variety of new methods. With the emergence of numerous such methods, including approaches based on generative models, one-class classification, and reconstruction, there is a growing need to bring methods of this field into a systematic and unified perspective. In this review we aim to identify the common underlying principles as well as the assumptions that are often made implicitly by various methods. In particular, we draw connections between classic 'shallow' and novel deep approaches and show how this relation might cross-fertilize or extend both directions. We further provide an empirical assessment of major existing methods that is enriched by the use of recent explainability techniques, and present specific worked-through examples together with practical advice. Finally, we outline critical open challenges and identify specific paths for future research in anomaly detection.

It is important to detect and handle anomalous inputs when deploying machine learning systems. The use of larger and more complex inputs in deep learning magnifies the difficulty of distinguishing between anomalous and in-distribution examples. At the same time, diverse image and text data commonly used by deep learning systems are available in enormous quantities. We propose leveraging these data to improve deep anomaly detection by training anomaly detectors against an auxiliary dataset of outliers, an approach we call Outlier Exposure (OE). This approach enables anomaly detectors to generalize and detect unseen anomalies. In extensive experiments in vision and natural language processing settings, we find that Outlier Exposure significantly improves the detection performance. Our approach is even applicable to density estimation models and anomaly detectors for large-scale images. We also analyze the flexibility and robustness of Outlier Exposure, and identify characteristics of the auxiliary dataset that improve performance.

Every AI system is deployed by a human organization. In high risk applications, the combined human plus AI system must function as a high-reliability organization in order to avoid catastrophic errors. This short note reviews the properties of high-reliability organizations and draws implications for the development of AI technology and the safe application of that technology.