Moving from Prediction to Action requires Causal AI: Response from the CHAI Hub to the AI for Science Strategy
Executive Summary
The CHAI (Causality in Healthcare AI) Hub welcomes the UK’s AI for Science Strategy and strongly endorses the shift “from prediction to action”. Importantly, action requires causality: intervening in a system requires understanding of the underlying cause-and-effect relationships to avoid unintended consequences. Causal AI enables the UK to deploy systems that are intervention-ready (predicting outcomes of actions), robust and generalisable (transportable to new settings and data), scientifically grounded (integrating physical laws and domain knowledge), and safe by design (avoiding hallucinations). Therefore, a firm causal foundation for the UK’s sovereign AI ambition is essential to ensure that we develop robust and accurate models.
We outline how CHAI is ready to contribute to the goals of the AI for Science strategy, how we can interact with DSIT (Department for Science, Innovation and Technology) and UKRI (UK Research and Innovation), and how we can support others in responding to the strategy’s actions. CHAI’s core causal AI research (motivated by but not restricted to medical research) is a unique ingredient that unlocks AI for causal discovery. Causal agents can integrate scientific understanding, knowledge from literature, and data at-scale to identify hypotheses by building internal world models for reasoning and exploration. CHAI is ideally placed to offer the required national leadership in causal AI, and act as a foundational partner in the ecosystem to deliver scientific and robust next-gen AI at scale. We are pleased to see CHAI already recognised as a supporter of the strategy (through our involvement in testing the Isambard AI resource). We are ready to scale our contribution to address the actions and missions outlined in the strategy.
CHAI support for the missions and actions
Action 1: Accelerating AI driven science: CHAI offers an integrated scientific knowledge creation framework using causal AI. Our flagship programmes: causal agents and problem fingerprinting, directly support this. We take a collaborative approach, focusing on building inclusive, interdisciplinary teams. We expand on this further in addressing the actions below.
Benefits: Accelerated discovery while ensuring scientific validity and robustness.
Action 2: Implications of AI for science: We welcome the focus on understanding the implications of AI in science. CHAI would welcome inputting into the National AI in Research Survey and supporting the goals of the UK Metascience Unit, by welcoming secondees into CHAI using our secondment programme.
Benefits: Collaboration could better surface the implications of AI in science.
Action 3: Scaling data storage: We very much welcome the move to scale up storage of simulated and real data. Intelligent scientific discovery algorithms, inspired by causal methods, can identify the right evidence and data to prioritise.
Benefits: Maximising return by ensuring focused storage investments and data use.
Action 4: Developing high-value datasets: High-value datasets should not automatically imply high volume and veracity. There is a need to actively be identifying the right data to collect and share. In sensitive domains this implies also sharing data without divulging unique information. Causal AI, via optimal experimental design and active sampling, can offer direct mechanisms for identifying the right composition of data assets.
Benefits: Enables efficient use of complex, multi-modal datasets by UK scientists.
Action 5: Dark data and negative experiments: CHAI can provide the methodological tools to understand dark data and distinguish it from missing data – through explicit causal modelling. This is essential for using the data we have with an understanding of selection biases, and robust data augmentation strategies to fill in the blanks. This is particularly important to secure full exploitation of high-value real-world data.
Benefits: More robust inferences and better use of data leading to maximal gain.
Action 6: Large-scale data infrastructure: We can help design infrastructures that allow for important causal metadata (such as known confounders and selection biases) to contextualise databases. This will allow safer and more consistent inference and more reliable synthetic data generation.
Benefits: Directly contribute to further valorisation of important, leading data assets.
Actions 7-8: Compute and simulation: We will expand our existing work with Isambard AI to showcase and realise the potential of the compute. With increased and dedicated compute allocation on these clusters we can unlock, at unprecedented scale, determinants of risk when models are stress tested with our causal approaches – thereby providing the ‘safety rails’ for wider AI deployment.
Benefits: Build on our impactful and high-stakes use-case to offer highly visible impact of compute investment.
Action 9: Doctoral talent: CHAI can contribute to a network of CDTs (Centres for Doctoral Training) to deliver training in causality – appreciation of which is critical regardless of specialism in AI. Our causality training is technically complete yet truly interdisciplinary, including governance, ethics and socio-technical systems thinking.
Benefits: AI researchers well-versed in causality will ensure robust and ethical development in AI.
Action 10: Upskill scientific researchers with AI: CHAI will expand its existing secondment programme to include an ‘inward’ secondment scheme where scientific researchers will spend a period embedded in a CHAI partner group/organisation – developing skills in causal AI and delivering a specific project.
Benefits: More experienced researchers in causal AI.
Action 11: Training programmes: We are already developing ‘causal education agents’ that provide tailored training for individuals. This activity can be scaled up and used as a blueprint for broader flexible on-demand training.
Benefits: Embedding fluency in causal reasoning, an area where the UK can lead.
Action 12: Interdisciplinary Research Teams: AI hubs provide a stellar foundation for bootstrapping the activity in this space. CHAI, due to its causal AI expertise, offers a unique ingredient that unlocks AI for scientific discovery. We are developing causal agents that combine scientific understanding of the world with knowledge extracted from the literature and data at scale to rapidly identify hypotheses for further investigation. This integration also allows the development of world models that provide core infrastructure for generating synthetic data and running in-silico trials in a robust and generalisable way. A world model built using non-causal AI is risky, as learnt associations often do not generalise: a world model is only as good as its mechanistic or causal assumptions. Agents can also support the design of studies using real-world data – such as optimising pipelines in meta-analysis and designing target trial emulations.
These current flagships of CHAI epitomise the ambition and belief that causal AI not only creates a safer AI but is also an integral component of the next-gen AI that AI for scientific discovery needs. In addition, CHAI offers a unique organisational and governance model that incorporates multidisciplinary expertise across all sciences to realise the true potential of causal AI. And hence by extension, CHAI’s model can address the missions outlined in the strategy.
Benefits: CHAI can offer the next-gen AI that AI for scientific discovery needs.
Action 13: Research Technical Professionals: We support the creation of an attractive professional pathway in academia that allows career progression for technical professionals.
Benefits: Retention and professional development of technical staff in academia.
Action 14: Community-driven benchmarking: Most existing benchmarks fail to anticipate how methods will perform in real-world settings. CHAI can build upon experiences in running communities of practice and has already worked with the scientific community on the problem fingerprinting flagship. Hence, we are ideally placed to deliver benchmarking tests and datasets that cover transportability, fairness, robustness (through stress-testing and adversarial testing), and performance under regime changes.
Benefits: Benchmarks that truly reflect real-world performance.
Action 15: Further AI for Science missions: We stand ready to generate proposals for these missions, and address others that arise.
Benefits: Causal grounding of future missions.
Mission 1: Drug Discovery: To meet the ‘100-day’ target we must bridge the ‘valley of death’ where successful molecules in the lab fail in the clinic. This is a causal transportability problem: ensuring lab results (including in-silico and automated labs) remain valid in the real-world environment. Causal inference allows us to mathematically map and therefore infer how and when efficacy will transport. This will reduce the failure rate in late-stage trials by filtering out false positives early.
Identifying potential targets requires rapid synthesis of fragmented knowledge: including existing literature, wet-lab experiments and molecular data. Our causal agent flagship can accelerate this by distinguishing true biological relationships from data correlations. This, combined with our work in the problem fingerprinting flagship, ensures the challenges are accurately framed and addressed. Realising the goal of this Mission will bring together the benefits of our contribution in each of the Actions we outlined above.
Benefits: Bridge the ‘valley of death’ by ensuring that lab-based efficacy will transport to real-world environments.
Conclusion and Recommendations
We propose the following actionable steps, which CHAI stands ready to support:
Fund causal AI as essential scientific infrastructure, because causal world-model development is core for AI for scientific discovery and Mission delivery.
Embed new individualised training tools in causal training pathways.
Ensure high-value datasets are created and made easily accessible, to maximise efficiency and directly empower benchmarking.
We see a huge opportunity for the UK to take an international lead in Causal AI for Science. CHAI hosts interdisciplinary research teams ready to drive scientific breakthroughs. CHAI is a ready and willing partner with all entities responding to the AI for Science strategy, offering causally-grounded science for research, training, and implementation, as well as an established and inclusive AI-focused ecosystem. We welcome approaches.
Contact: chai@ed.ac.uk