Lorenzo Pacchiardi

Ensuring safe and effective use of artificial intelligence (AI) requires understanding and anticipating its performance on new tasks, from advanced scientific challenges to transformed workplace activities. So far, benchmarking has guided progress in AI but has offered limited explanatory and predictive power for general-purpose AI systems, attributed to limited transferability across specific tasks. Here we introduce general scales for AI evaluation that elicit demand profiles explaining what capabilities common AI benchmarks truly measure, extract ability profiles quantifying the general strengths and limits of AI systems and robustly predict AI performance for new task instances. Our fully automated methodology builds on 18 rubrics, capturing a broad range of cognitive and intellectual demands, which place different task instances on the same general scales, illustrated on 15 large language models (LLMs) and 63 tasks. Both the demand and the ability profiles on these scales bring new insights such as construct validity through benchmark sensitivity and specificity and explain conflicting claims about whether AI has reasoning capabilities. Ultimately, high predictive power at the instance level becomes possible using the general scales, providing superior estimates over strong black-box baseline predictors, especially in out-of-distribution settings (new tasks and benchmarks). The scales, rubrics, battery, techniques and results presented here constitute a solid foundation for a science of AI evaluation, underpinning the reliable deployment of AI in the years ahead.

Frontier AI is becoming critical societal infrastructure, but outsiders lack reliable ways to judge whether leading developers’ safety and security claims are accurate and whether their practices meet relevant standards. Compared to other social and technological systems we rely on daily such as consumer products, corporate financial statements, and food supply chains, AI is subject to less rigorous third-party scrutiny along several dimensions. Ambiguity about whether AI systems are trustworthy can discourage deployment in some contexts where the technology could be beneficial, and make it more likely when it’s dangerous. Public transparency alone cannot close this gap: many safety- and security-relevant details are legitimately confidential and require expert interpretation. We define frontier AI auditing as rigorous third-party verification of frontier AI developers’ safety and security claims, and evaluation of their systems and practices against relevant standards, based on deep, secure access to non-public information. To make rigor legible and comparable, we introduce AI Assurance Levels (AAL-1 to AAL-4), ranging from time-bounded system audits to continuous, deception-resilient verification.

AI systems increasingly incorporate continual learning mechanisms allowing their behaviour to adapt after deployment, from (1) in-context learning and (2) memory features already in wide use to (3) post-deployment weight modification under research. We argue that, by treating AI systems as frozen artefacts whose performance and safety are assessed at release, current evaluation practices structurally ignore the behavioural trajectory of a system that continues to learn from experience. Our position is that evaluation of continual learning systems should be centred on behavioural trajectories, with the complementary goals of characterising the landscape of possible behaviours and forecasting how behaviour will evolve from a given set of experiences. This can be operationalised through trajectory elicitation sandboxes and predictive monitors that forecast behavioural evolution, but may face fundamental obstacles analogous to those seen in dynamical systems. These are best addressed by (1) applying trajectory-centred evaluation to today’s continual learning systems and (2) relying on the resulting evidence to design systems amenable to it, yielding a virtuous cycle in which systems and their evaluations co-evolve.

We propose a framework for categorising AI models as General-Purpose AI (GPAI) models, based on their capabilities and generality, as defined in the European Union (EU) AI Act. Our framework breaks down the core components of the GPAI definition into measurable elements, focusing on four primary cognitive domains: Attention and Scan, Comprehension and Compositional Expression, Conceptualisation, Learning and Abstraction, and Quantitative and Logical Reasoning. We suggest using the Annotated Demand Levels (ADeLe) procedure to evaluate AI models’ capabilities in these domains, and provide a methodology for combining domain-level scores into a single measure of generality. The framework is illustrated with empirical results from existing models, and policy recommendations are made for selecting thresholds and metrics for GPAI categorisation.

Research in AI evaluation has grown increasingly complex and multidisciplinary, attracting researchers with diverse backgrounds and objectives. As a result, divergent evaluation paradigms have emerged, often developing in isolation, adopting conflicting terminologies, and overlooking each other’s contributions. This fragmentation has led to insular research trajectories and communication barriers both among different paradigms and with the general public, contributing to unmet expectations for deployed AI systems. To help bridge this insularity, in this paper we survey recent work in the AI evaluation landscape and identify six main paradigms. We characterise major recent contributions within each paradigm across key dimensions related to their goals, methodologies and research cultures. By clarifying the unique combination of questions and approaches associated with each paradigm, we aim to increase awareness of the breadth of current evaluation approaches and foster cross-pollination between different paradigms. We also identify potential gaps in the field to inspire future research directions.

Large language models (LLMs) can "lie", which we define as outputting false statements despite "knowing" the truth in a demonstrable sense. LLMs might "lie", for example, when instructed to output misinformation. Here, we develop a simple lie detector that requires neither access to the LLM’s activations (black-box) nor ground-truth knowledge of the fact in question. The detector works by asking a predefined set of unrelated follow-up questions after a suspected lie, and feeding the LLM’s yes/no answers into a logistic regression classifier. Despite its simplicity, this lie detector is highly accurate and surprisingly general. When trained on examples from a single setting – prompting GPT-3.5 to lie about factual questions – the detector generalises out-of-distribution to (1) other LLM architectures, (2) LLMs fine-tuned to lie, (3) sycophantic lies, and (4) lies emerging in real-life scenarios such as sales. These results indicate that LLMs have distinctive lie-related behavioural patterns, consistent across architectures and contexts, which could enable general-purpose lie detection.