New Funding Call: Trustworthy Evaluation of AI Tools in Science and Engineering

Up to £50,000 to verify the datasets that will evaluate AI in your domain

Registration deadline: 12 October 2026, 16:00 BST
Application deadline: 14 October 2026, 16:00 BST

Apply now

---

About this call

Apply for funding to develop and publish rigorous, reusable protocols for evaluating the integrity and trustworthiness of AI copilots and AI-enabled research tools in a defined scientific or engineering domain, and to use those protocols to seed a library of verified evaluation datasets. The overarching goal is to enable domain experts to assess whether evaluation datasets are fit for a stated purpose and to ensure that a persistent verification record is associated with each dataset, so that claims about AI tools in research rest on evidence that is transparent, reproducible and appropriate to the discipline.

Your proposed work should primarily focus on research that aligns to the mission of Responsible AI UK (RAi UK). Learn about the RAi UK programme.

You and your organisation must be eligible for UKRI funding. See the eligibility conditions below.

Call information

Back to top ↑

Who can apply

Grants are open to:

• UK higher education institutions
• public sector research establishments
• research council institutes
• UKRI-approved independent research organisations and NHS bodies with research capacity.

UK businesses, third sector organisations, standards and measurement bodies, dataset custodians and government bodies can be funded as part of the project but cannot lead it (restrictions apply, see Funding available below).

Read the guidance on institutional eligibility.

The lead applicant will need to be resident in the UK and hold a contract at the submitting research organisation that extends beyond the duration of the proposed project.

Applicants must demonstrate access to both of the following: qualified domain experts (scientists, engineers or other appropriately qualified subject-matter experts) capable of making the substantive verification judgements; and the technical capability to implement versioning, citation, metadata and registry integration. Consortia are encouraged where they bring complementary expertise in scientific or engineering domains, data stewardship, AI evaluation, measurement, standards, research infrastructure or long-term digital preservation.

We particularly encourage early career researchers to apply.

This call is not subject to a UKRI policy on repeatedly unsuccessful applications.

What we are looking for

AI copilots and AI-enabled tools are being adopted across science and engineering faster than the means of evaluating whether they can be trusted. When these tools help generate hypotheses, analyse data, write code, review literature, or interpret results, their reliability becomes a question of research integrity. Yet much of the evaluation that underpins claims about them rests on datasets and protocols whose fitness for purpose has never been examined by the people best placed to judge it. A benchmark can be well engineered and still be scientifically unsound, unrepresentative of the conditions under which a tool will be used, or contaminated by the data on which the tool was trained. Strong benchmark performance is therefore not, on its own, evidence that a tool is trustworthy in research practice. This call funds the evidence layer that sits between scientific data stewardship and AI evaluation, so that claims about AI tools in science rest on evidence held to the same standards of rigour as the science they support.

For this funding opportunity, projects should:

• develop an open, versioned and citable verification protocol.
• define evidence-based verification criteria.
• ensure verified datasets can test the properties that matter for trustworthy use in research, with criteria appropriate to the domain and intended use;
• establish a transparent and reproducible verification process.
• apply the protocol in at least one well-defined scientific or engineering domain;
• produce, for each verified dataset, a full, persistent verification record stating who verified it, against which version of the protocol and criteria, on what evidence, and with what known limitations, so that the record travels with, or remains persistently linked to, the dataset wherever it is used; and
• treat responsible AI and research ethics as substantive parts of the work, as set out in the sections below, rather than as compliance statements.

Optional elements supported by RAi UK

Projects may also propose the following. These are not required, and RAi UK will support projects that choose to take them on, including through coordination across funded projects and, where appropriate, shared infrastructure:

• a registry of verified holdings, with persistent links between each dataset and its verification record;
• an evaluation report documenting the application of the protocol, its findings, limitations and lessons, and recommendations for expansion, written to a standard open to peer scrutiny; and
• a custodianship arrangement naming the organisation that will maintain the registry over the long term.

The programme is concerned with the fitness of datasets for trustworthy evaluation. It is not intended to certify an AI system, to endorse a particular model or vendor, or to establish that a dataset is universally valid for every scientific purpose. The protocol should distinguish, where relevant, between:

• fitness for a specified evaluation task;
• fitness for a specified domain or sub-domain;
• fitness for particular classes of AI copilot or tool;
• known exclusions, biases, gaps or failure modes; and
• the evidence required to support each verification judgement, and the level of confidence it warrants.

Proposals not meeting these requirements in the judgement of RAi UK will be rejected.

Focus areas

Proposals may address any scientific or engineering domain that can demonstrate a clear need for trustworthy evaluation datasets for AI copilots or tools. Applicants must identify the relevant sector or sectors among the eight identified in the UK Government’s industrial strategy, and explain the connection between the proposed domain, the datasets to be verified and the intended evaluation use case:

• Advanced Manufacturing
• Clean Energy Industries
• Creative Industries
• Defence
• Digital and Technologies
• Financial Services
• Life Sciences
• Professional and Business Services

The Industrial Strategy sectors are broader than academic disciplines and are not expected to map one-to-one onto them. Applicants should identify the sector or sectors most relevant to the intended application or downstream use and explain the connection; where the mapping is indirect, this should be stated explicitly.

Applications that propose domains within the five UKRI priority research areas for AI adoption, namely engineering biology, advanced materials, quantum technologies, medical research and fusion energy, are particularly welcome, for the reasons set out in the next section.

For international collaborators (for example, industry or academic partners, or overseas dataset custodians), applicants should use the current NPSA Trusted Research guidance and Implementation Collaboration Checklist, alongside their institution’s own due-diligence and research-security processes. Based on that assessment, applicants may need to escalate the collaboration within their institution or department, and should allow for the tight timescale for doing so.

Further guidance on getting the most out of international collaboration while protecting intellectual property, sensitive research and personal information has been released by NPSA.

Alignment with UKRI investment in AI for science

This call is designed to complement the UKRI Enabling AI adoption across science and engineering funding opportunity, delivered through the UKRI IS8 AI Programme. That opportunity funds exploratory projects that pair AI expertise with domain expertise in engineering biology, advanced materials, quantum technologies, medical research and fusion energy, with projects starting by 26 April 2027. Among the outcomes it anticipates is the identification of the data, infrastructure and benchmarking requirements needed to enable progress.

The two calls address complementary parts of the same problem. The UKRI opportunity asks how AI can advance a scientific domain; this call asks how researchers can establish whether the data and evaluation evidence used to assess AI tools are fit for purpose. The intention is that a verified library available before the UKRI cohort begins gives those projects a trustworthy basis for evaluating the tools they build or adopt, while creating a direct route for RAi UK to connect with and learn from that portfolio.

Applicants may also wish to refer to the UK Government’s AI for Science Strategy.

Back to top ↑

Verification, responsible AI and ethics requirements

The verification protocol

Verification in this call means a documented, evidence-based assessment of whether an evaluation resource is fit for a stated evaluation purpose. It is not formal certification of an AI system, and it does not imply that a dataset is error-free or universally valid.

For the purposes of this call, an ‘evaluation dataset’ may include curated observational or experimental data, simulation outputs, reference measurements, benchmark tasks or test cases, curated corpora, or multimodal research data, provided it is used to evaluate an AI copilot or AI-enabled research tool. Applicants should define the unit of evaluation that is appropriate to their discipline.

The protocol should specify, as relevant to the domain and intended use:

• the intended evaluation task, target class of AI tool or system, and research context;
• provenance, data quality, representativeness and known exclusions or gaps;
• reference values, labels, expert judgements or other basis for assessing correctness, including uncertainty where applicable;
• metrics, evidence requirements and decision rules, including how limitations and confidence are recorded;
• contamination or prior-exposure risk, including potential overlap with model pre-training, fine-tuning or previous benchmark exposure;
• reproducibility requirements, versioning and audit trail; and
• domain-expert review, independence, handling of disagreement and triggers for re-verification.

Where model training data or development histories are unavailable, for example for proprietary models, applicants are not expected to prove the absence of contamination. They should document the checks performed, the evidence available and the residual uncertainty.

Domain-led verification

A defining requirement of this call is that substantive dataset verification is undertaken by domain experts: scientists, engineers or other appropriately qualified subject-matter experts. Applicants should:

• identify named or clearly defined domain-expert roles responsible for verification, and explain how verifiers are selected;
• describe the competence, independence and conflict of interest arrangements that apply to verifiers, including any relationship with the developers of the AI tools the dataset may be used to evaluate;
• describe how verification judgements will be checked, for example through independent review by a second expert, calibration across verifiers or sampling of decisions, where feasible;
• record in each verification record the identity or institutional role of each verifying expert, subject to appropriate privacy and governance requirements, together with the evidence reviewed, the verification date, the protocol version, the dataset version and the resulting scope of fitness; and
• record dissent, uncertainty and unresolved issues, and state the level of confidence each judgement warrants, rather than forcing a binary verdict where the evidence does not support one.

Responsible AI requirements

Each project should show how responsible AI considerations are built into the evaluation design rather than treated as a separate compliance exercise. The level of analysis should be proportionate to the domain, the intended use of the AI tool and the consequences of error.

The responsible AI plan should address, where relevant:

• assurance: the claims being made, their limits, sources of uncertainty, failure modes and the evidence needed to support use;
• evaluation: domain-relevant tasks and conditions, robustness, reproducibility and meaningful failure cases;
• datasets: provenance, quality, access, licensing, representativeness, contamination risk and versioning;
• human-centredness: where expert judgement, human oversight or affected users matter, including fairness and accessibility where relevant; and
• safety, security and dual-use: foreseeable misuse, information hazards, security implications and other material risks in the chosen domain.

Applicants should explain when a consideration is not relevant to their domain rather than force every criterion into every discipline.

Research ethics

Drawing upon recently published guidance on the ethics review and oversight of AI-related health research (Artificial intelligence-related health research: ethics review and oversight, WHO, 2026), this call aims to generate comparable guidance for AI-related research in the physical, engineering and wider natural sciences.

Each application should include a research ethics plan that addresses, as relevant to the domain:

• which of the following categories of AI-related research the project involves: (i) data-intensive research using AI; (ii) research conducted with AI tools and technologies; and/or (iii) research on AI tools and technologies; and the ethical questions each raises in the chosen domain;
• how ethical issues will be reviewed during the project, and not only at the outset, including a named point of review within the host institution or consortium and the triggers that would prompt it;
• the secondary uses to which the verified datasets may be put, including foreseeable misuse and dual-use risk, which is of particular importance in engineering biology, defence-related and security-sensitive domains;
• transparency about the AI tools used in the project itself, including their role in curation, annotation or verification;
• the reporting of negative, inconclusive or unfavourable verification findings, which carry as much value as favourable ones;
• conflicts of interest, including where verifiers, dataset custodians or partners have an interest in a particular outcome; and
• where data concern people or are drawn from health or clinical settings, compliance with the relevant health research ethics requirements, for which the WHO guidance applies directly.

Applicants should also refer to the UKRI position statement on funding ethical research and to UKRI guidance on responsible innovation.

Back to top ↑

Required deliverables

RAi UK will convene a single programme-level registry of verified holdings. All funded projects must provide the metadata, verification records and persistent links needed to include their verified datasets. Projects are not required to build or host a separate registry unless they propose this as an optional enhancement.

Minimum dataset library requirements

• A portfolio of datasets within at least one well-defined scientific or engineering domain.
• Meaningful variation in provenance, structure, scale, modality or evaluation purpose, with a rationale for the selection.
• An assessment of contamination risk for each dataset, including the likelihood that it appears in the training data of the tools it will evaluate.
• An identifiable custodian, terms of use that permit AI evaluation, and an agreed mechanism for open or controlled access.
• A version-specific verification record that travels with, or is persistently linked to, each dataset.
• A clear statement of the purposes for which each dataset is and is not fit, since verification is not endorsement.

 

Open science, citation and persistence

Outputs should follow the principle ‘as open as possible, as closed as necessary’. Protocols, templates, metadata schemas, verification records and project-generated code should be openly accessible under appropriate licences unless legal, ethical, security, commercial or third-party constraints prevent this. Underlying datasets may use controlled access where necessary.

Applicants should, where feasible:

• assign persistent identifiers to protocols, dataset landing pages, verification records and reports;
• version datasets and verification records separately and preserve links between versions;
• state licences, access conditions and recommended citation;
• publish sufficient metadata for discovery and reproducibility even when the underlying data cannot be openly shared; and
• deposit outputs in repositories expected to remain accessible beyond the award.

Back to top ↑

Partnerships, delivery, and funding

Partnerships and ecosystem engagement

Projects are expected to build a partnership broad enough to strengthen the work. Where relevant, applicants should consider partnerships with AI hubs and laboratories, Centres for Doctoral Training, domain research organisations, dataset custodians, standards and measurement organisations, and the National Physical Laboratory’s Centre for AI Measurement.

Partners should have defined roles rather than being included solely as supporters. Letters of support or collaboration should explain the partner’s contribution to protocol development, domain verification, dataset access, evaluation, infrastructure or long-term stewardship.

Project plan and timetable

The call is intended for rapid mobilisation. Applicants should propose a realistic delivery plan from mobilisation through protocol development, domain review and dataset verification, to preparation of holdings and verification records for integration into the shared RAi UK registry, final evaluation and handover.

The dates above are indicative planning assumptions. Applicants should provide their own detailed milestones and identify any dependencies, access constraints or risks that could affect delivery, particularly the time needed to agree dataset access with custodians.

Funding available

£150,000 is available and we expect to fund around three to four projects, depending on award size. We expect awards normally to be between £25,000 and £50,000. With exceptional justification, requests of up to £75,000 may be considered.

Grants will be funded at 80% of the stated Full Economic Cost (fEC). The remaining 20% is the standard contribution from the submitting research organisation. No additional matched funding is required. Cash or in-kind contributions from project partners may be included where they add value, but they will not be treated as an assessment criterion.

Project co-leads from non-academic organisations (including organisations outside the UK, with due regard for Trusted Research requirements) may be included in project costs, up to 30% of the total fEC.

We recognise that some partners may be employed by a government-funded organisation. To avoid the double counting of public funds in the costings, no salary costs will be covered for permanent employees of government bodies.

Reasonable costs of domain-expert verification time, dataset access, annotation and registry infrastructure may be included and should be justified. Where annotation or curation is contracted out, costs should reflect fair remuneration for the people doing the work.

PhD studentships, or funding associated with PhD studentships, are not eligible for inclusion in the costs sought from RAi UK.

Back to top ↑

How to apply

Applicants should ensure they are aware of, and comply with, any internal institutional deadlines that may be in place.

Applicants are responsible for registering their intent to submit during the stated registration period.

Proposals should be prepared using the provided submission template, completing all of the sections, and submitted in PDF format via the online application portal. Once you submit your proposal you will receive a confirmation email, including details of your submission.

Submission Link

Pre-registration and submission will be through Grantlounge. Pre-registration must be completed by 12 October 2026 (16:00 BST), and submissions will close on 14 October 2026 (16:00 BST).

RAi UK must receive your application by 16:00 on the day of the relevant submission deadline.

What your proposal must contain

• Case for support and statement of need.
• Description of the target scientific or engineering domains, industrial strategy alignment, and any alignment with the UKRI priority areas.
• Proposed verification protocol, including preliminary criteria and evidence model.
• Description of the domain-expert verification model and its independence arrangements.
• Responsible AI plan, addressing assurance, evaluation, datasets and human-centredness.
• Research ethics plan, as described above.
• Initial dataset portfolio, including custodians, access arrangements and the rationale for selection.
• Partnership and governance structure.
• Long-term stewardship plan for datasets, verification records and any project-hosted infrastructure.
• Project plan, milestones, risks and dependencies.
• Dissemination, open science and citation plan.
• Budget and justification.
• Letters of support or collaboration from key dataset custodians and delivery partners, where appropriate.

Back to top ↑

Assessment

Proposals that meet the assessment criteria will be considered by a panel of experts drawn from RAi UK’s network, including scientists, engineers and other domain experts, as well as AI and evaluation specialists, in order to select the final successful proposals.

Only general feedback will be given to unsuccessful candidates.

Assessment criteria

All submitted proposals will be evaluated according to the following criteria:

Fit to funding opportunity (primary)

This includes:

• the scientific and engineering credibility of the proposed domains, and their relevance to one or more industrial strategy sectors;
• the quality, relevance and diversity of the proposed dataset library;
• the value of the proposed work to the wider UK research and innovation ecosystem, including its potential use by UKRI-funded AI for science projects; and
• relationship to the RAi UK mission.

Quality (primary)

The research excellence of the proposal, making reference to:

• the clarity, rigour and generalisability of the proposed verification protocol;
• the strength and independence of domain-expert involvement in verification;
• the approach to provenance, versioning, citation, uncertainty, limitations and auditability; and
• the feasibility of delivering the minimum dataset requirements and shared-registry contribution within the timetable.

Responsible AI and research ethics (primary)

• the quality of the responsible AI plan across assurance, evaluation, datasets and human-centredness;
• the extent to which verification records function as credible assurance artefacts;
• the quality of the research ethics plan, including continuing review, dual-use and misuse risk, transparency and conflicts of interest; and
• the treatment of people involved in creating, curating, annotating and verifying data, where applicable.

Partnerships and sustainability (primary)

• the quality of partnership arrangements and access to relevant AI, measurement, domain and data expertise;
• the credibility of long-term stewardship for datasets, verification records and any project-hosted infrastructure.

Equality, Diversity & Inclusion and Responsible Research & Innovation (secondary)

• Plans to embed responsible innovation practices in the project.
• Adequate consideration of EDI in terms of the research or knowledge exchange, including the composition of the verification team.

Resourcing (secondary)

• Adequate and fully justified resourcing.
• Risks to resourcing fully identified and mitigated, including appropriate plans to manage finances during the lifetime of the activity.
• Consideration of risks of delays in completing the project within the timeline specified, including delays in securing dataset access.

Grant additional conditions (GACs)

Projects will be expected to follow open-source, open-data and open-innovation principles on an ‘as open as possible, as closed as necessary’ basis, subject to licensing, legal, ethical, security, commercial and third-party constraints affecting the underlying datasets or other project outputs.

Projects will be expected to share their verification records and domain research ethics notes with RAi UK for the purposes of programme-level synthesis.

Awards will be confirmed upon acceptance of the non-negotiable Terms and Conditions, including the fixed end date of 31 March 2027, which will be set out in the Award Letter.

Back to top ↑

Contact details

Queries regarding the submission of proposals should be directed to: info@rai.ac.uk

Back to top ↑