King's College London

Generative and analytical AI systems offer unprecedented opportunities for innovation, efficiency and discovery, but their rapid adoption raises critical questions about integrity, governance, reproducibility, infrastructure and ethics. It also risks undermining public trust in science and discovery.

The King’s Institute for Artificial Intelligence is driving a programme of activity across King's College London to understand what happens when artificial intelligence tools become part of the research process, and to equip our research community with the tools and skills to pursue responsible, innovative, and robust scientific discovery. We will take a leading role in convening cross-sector dialogue, sharing our insights from King’s to promote best practice and address sector-wide challenges.

'The UK is a scientific nation. Scientific discovery is one of the core drivers of human progress. The UK must act decisively to maintain its scientific leadership and seize the opportunity to shape the transformation of science by AI.’ AI For Science Strategy, Department for Science, Innovation & Technology, 20 November 2025

Why does this matter outside the university?

Public trust in science is shifting. The Public Attitudes to Science 2025 survey shows that some specific aspects of trust in science have diminished:

• Only 55% of respondents believe scientists consider risks of new technologies (down from 69% in 2019)
• 42% agreed that that the speed of development in science and technology means that they cannot be properly controlled by government

As AI becomes a routine part of research, universities must show that new methods are used responsibly, transparently and with integrity.

Programme image: Yutong Liu & Digit / https://betterimagesofai.org / https://creativecommons.org/licenses/by/4.0/

Aims

1. Support the development and adoption of new AI-enabled research methods across scientific disciplines, helping researchers experiment, evaluate and integrate emerging approaches responsibly.
2. Facilitate access to the data, compute, models and tools required for high quality AI-enhanced research, so that King’s researchers can work with robust, transparent and sustainable infrastructure.
3. Strengthen metascience practices around ethics, reproducibility, publishing and novelty by shaping guidance, frameworks and sector-wide conversations on how AI transforms scientific integrity.
4. Develop the skills and capabilities of researchers and research professionals, providing training, resources and communities of practice that build confidence and expertise in responsible AI use.

King’s Autonomous Labs 

King's Autonomous Labs is a cross-university research initiative developing new approaches to automated and autonomous scientific discovery.

KAL brings together artificial intelligence, laboratory automation, robotics, sensing and data to develop self-driving laboratories. These are connected experimental systems that can perform experiments, analyse results and use emerging evidence to inform what happens next.

A self-driving laboratory is not a laboratory without people. Researchers remain responsible for defining the scientific questions, objectives, boundaries and safeguards. By combining their expertise with automation and AI, KAL aims to make experimentation more adaptive, reproducible and efficient.

Open tools, standards and frameworks

Researchers at King's also contribute to international initiatives that support transparent, reproducible and trustworthy AI-enabled research.

Evaluation Cards

AI evaluation results are published across academic papers, technical reports, leaderboards and model documentation. However, evaluation results are often reported inconsistently, with important information about methodology, benchmarks and provenance omitted or difficult to find. This can make it challenging for researchers, policymakers and other stakeholders to compare models, assess the reliability of reported results and understand the limitations of AI evaluation evidence. Evaluation Cards was developed by the EvalEval Coalition, including contributors from King's College London, to help improve transparency and consistency in the reporting of AI evaluation results.

Rather than introducing a new benchmark, Evaluation Cards provides an interpretive layer over existing evaluation data. The framework combines reported evaluation results with structured benchmark metadata and assesses them across dimensions including reproducibility, completeness, provenance and comparability. This helps users understand not only a model's score, but also how that result was produced and how much confidence should be placed in it. The open-source project already brings together more than 100,000 evaluation results spanning thousands of models and hundreds of benchmarks.

King's contributor: Dr Wm Matthew Kennedy.

Croissant

As artificial intelligence systems become increasingly dependent on large and complex datasets, ensuring that those datasets are discoverable, accessible and reusable has become a significant challenge. Dataset documentation is often inconsistent, requiring researchers and developers to spend considerable time understanding how data is structured and whether it is suitable for a particular use case. Croissant is an open metadata standard developed through an international collaboration between academia and industry to address this challenge and support more efficient and responsible reuse of datasets.

A Croissant file provides a structured, machine-readable description of a dataset, including information such as its structure, provenance, licensing and other important metadata. Because this information is recorded in a common format based on established web standards, it can be interpreted automatically by repositories, machine learning tools and AI systems. The Croissant ecosystem has since expanded to include domain-specific extensions:

• RAICroissant, which supports responsible AI documentation
• BioCroissant, which helps describe life sciences and biomedical datasets
• GeoCroissant, which supports geospatial and Earth observation data

Researchers from King's College London have contributed to the development of both the core standard and several of these extensions through the international MLCommons community.

King's contributors include Professor Elena Simperl and Dr Albert Meroño Peñuela.