UK Research and Innovation has opened a £12.05 million funding opportunity for speculative work on explainable AI, deliberately accepting the risk of failed projects in pursuit of better ways to understand how future systems reach outputs and decisions.

The programme is being delivered with the Engineering and Physical Sciences Research Council and opened on 11 August, with applications closing on 20 October. Individual projects can have a full economic cost of up to £602,500, of which UKRI will fund 80%, and successful work must begin on 1 February 2027 for up to two years.

UKRI is explicitly not requiring applicants to produce new commercial models or spinouts during the grant period. Its guidance says projects are not expected to achieve every research goal and that approaches which fail can still generate useful knowledge, provided the proposal is ambitious enough to offer a potential step change rather than incremental improvement.

That tolerance reflects the scientific difficulty underneath a familiar policy demand. Businesses and public bodies can obtain increasingly capable AI systems, but many consequential uses remain difficult when the organisation cannot establish why a system produced a result, where uncertainty entered the process, or how much confidence should be placed in the explanation generated afterwards.

Explainability is several technical problems

UKRI has defined the research scope broadly. It includes post-hoc explanations intended to help people understand model outputs, mechanistic interpretability examining how internal structures produce behaviour, process examination covering the steps and tool use involved in reaching a conclusion, techniques that make models inherently easier to interpret, and methods for quantifying uncertainty.

Those approaches solve different problems. A plain-language explanation produced after a decision may help a user without proving that the explanation reflects the mechanism which caused the result, while deeper inspection of internal model behaviour may be technically informative but unusable to the employee, regulator, customer, or citizen affected by the decision.

Agentic systems make the chain harder to reconstruct because a final action can depend on several models, retrieval systems, databases, software tools, and intermediate decisions. An error may originate in the model itself, the information it retrieved, an external tool, or the orchestration between components, leaving explainability as a systems problem rather than a feature attached to one neural network.

UKRI is allowing interdisciplinary teams where their work contributes to technical advances, although proposals consisting mainly of social-science, legal, policy, or ethical analysis are outside the programme. The intention is to improve the technology itself rather than fund another layer of commentary around existing models.

Research policy follows deployment constraints

The funding call is the first pilot under UKRI’s IS8 AI Programme and forms part of a wider national research strategy that identifies explainable and human-in-the-loop systems among the capabilities needed for future AI. UKRI has also committed substantial funding across AI during the current Spending Review period, tying research policy more closely to deployment in science, public services, and industry.

Explainability occupies an unusual place inside that agenda because it is fundamental research driven partly by practical adoption barriers. A model can achieve strong benchmark performance and still be difficult to use where a bank must justify a decision, an engineer needs to understand a recommendation affecting physical equipment, or a public authority must account for how an automated system influenced an outcome.

The requirement is not uniform across industries, and explainability should not be reduced to a single regulatory checkbox. Some environments need a user-facing reason, others need auditable system behaviour, and still others require confidence estimates or evidence that humans can detect when the system is outside the conditions in which it performs reliably.

Current methods remain imperfect. Interpretability can be computationally expensive, tied to particular model architectures, inaccessible to non-specialists, or capable of producing explanations that sound coherent without capturing the complete causal path behind an output. Improving interpretability can also introduce trade-offs in performance or system complexity.

The programme’s willingness to fund approaches that do not work acknowledges that there is no settled engineering solution. UKRI is also experimenting with distributed peer review, requiring applicant teams to participate in assessment of other submissions rather than sending every proposal through a conventional reviewer pipeline.

That combination gives the funding call a different purpose from an industrial deployment programme. It is not paying organisations to add an explanation panel to current AI products; it is trying to widen the technical options available before more capable systems become embedded in settings where a fluent answer is insufficient evidence of safe operation.

The £12.05 million fund is modest beside the capital being spent on models and compute, but the research problem reaches directly into adoption. Organisations can buy powerful AI increasingly easily. Understanding when its outputs deserve to be trusted remains considerably harder.