Distinguished colleagues and partners,

Thank you for joining the second Research 7+ (R7+) meeting in Ottawa. Our objective was straightforward in principle and demanding in practice: to move from the G7 Leaders’ commitments on “AI for Prosperity” to an actionable, near‑term programme of work that helps ensure AI strengthens (rather than erodes) the scientific enterprise. This document records both what we heard from the research community and what we, as public research funders and performers, have jointly committed to deliver.

Why the urgency? Because the pace of change in research is no longer well-modelled as linear. Consider a modest thought experiment: if AI research capability doubles every seven months (and we set aside any metaphysics about “AGI”, whatever that is), then a growing share of the quotidian work of science (search, synthesis, coding, experimental orchestration) becomes, in effect, rentable by the hour. Universities and funding systems, by contrast, are tuned to multi‑year cycles: training pipelines, tenure clocks, infrastructure procurement, and evaluation regimes. The mismatch matters. Institutions like ours struggle with “sudden”.

That is also why co‑operation matters, particularly amongst like‑minded partners committed to openness, integrity, and public value. “Clubs” are not an end in themselves; they are a means of pooling scarce talent, aligning standards, and building shared (and trusted) infrastructure, all while respecting sovereignty and legitimate constraints through approaches such as Data Free Flow with Trust. Indeed, the pragmatic joint projects to which we have committed will only succeed through sustained collaboration: technical, institutional, and cultural.

Against that backdrop, our three joint initiatives: a Joint Talent Hub framed as an AI Research Commons (ARC) pilot, a Data Free Flow with Trust (DFFT)‑grounded workstream on federated science data spaces, and joint principles for the responsible use of AI in research, should be read as practical “plumbing” for trust: lowering the friction of collaboration, raising shared assurance, and anchoring AI in scientific norms of transparency and reproducibility. Yes, the work is aspirational and not binding; but it is concrete, and it is shared.

- Mark Daley (NSERC’S Scholar in Residence in AI, Canada)

The R7+ Engagement Group took form in 2024 during Italy’s G7 presidency. The inaugural Research 7+ (R7+) meeting was a success, bringing together international leaders of research funding and research performing organizations to advance dialogue on research and innovation. Building on this momentum, and as part of Canada’s G7 presidency, the Natural Sciences and Engineering Research Council of Canada (NSERC) hosted the second R7+ meeting in Ottawa, Ontario, Canada from November 17 to November 18.

The objective of this year’s R7+ meeting was to convert high-level commitments in the G7 Leaders’ Statement on AI for Prosperity (June 17, 2025) into an actionable roadmap for the advancement of AI for Prosperity through Science Research.

The two-day event brought together Canadian scientific researchers at the leading edge of AI applications for science with AI Experts and Leaders from both public research organizations and research funders from the R7+ countries. Participants included: NSERC (Canada), CNRS (France), Fraunhofer (Germany), CNR (Italy), RIKEN (Japan), UKRI (United Kingdom), the NSF (United States), EU JRC (European Union) and CSIC (Spain) to address the opportunities and risks of AI in science in a coordinated way. The meeting provided an opportunity for insightful discussions that allowed for the identification of concrete opportunities for collaboration and engagement between R7+ organizations.

The current document, led by NSERC in consultation with participating organizations, summarizes the discussions and proposed outcomes from the 2025 R7+ meeting. While it reflects areas of general alignment, it does not capture every comment or viewpoint, and the content should not be taken as representing the official positions of participating organizations.

The R7+ AI Experts Meeting took place on November 17, 2025. This first part of the R7+ meeting included research leaders in AI from across the R7+ countries, as well as Canadian academics and researchers at the leading edge of AI who provided expertise from diverse perspectives of research, clinical practice, engineering, and ethics. The day’s discussions focused on how artificial intelligence is reshaping the practice of science in different fields and disciplines and, in turn, how public research systems should/could respond to maximize positive impacts and mitigate potentially harmful or negative unintended impacts. The discussions took place through a panel discussion followed by moderated break-out groups focusing on three disciplinary categories:

• AI for Health and Life Sciences
• AI for Earth and Environmental Sciences
• AI for Physical and Materials Sciences

Across disciplines and sectors, the general conclusion was consistent: the central challenge is not whether AI will transform research, but who will shape that transformation, to what ends, and under what norms. AI is already transforming and boosting ways of conducting science (“AI in science”) but also vast scientific areas focus now on frontier research on AI (“Science for AI”). Left unattended, the researchers expressed concern that the centre of gravity might default to large corporate actors, priorities focused on excitement regarding novel methods rather than on impact, opaque systems, and a gradual erosion of the human skills and institutions that make science trustworthy.

The outcomes of the discussions highlighted the need to build cultures, incentives and infrastructures that strengthen scientific judgement, interdisciplinarity, and public trust, and emphasized that the highest priority and the highest return on investment in AI for science lies in people rather than hardware or individual models.

Throughout the day’s discussions, six main themes emerged among the participants’ proposed actions:

1. Increase AI literacy and maintain the human edge
2. Anchor investments on impact and clearly articulated problems and communities
3. Support team science and interdisciplinarity
4. Ensure access to shared and trusted infrastructure
5. Ensure energy sustainability and transparency
6. Strengthen data access and standards

In the sections that follow, the main takeaways from discussions in relation to each of these themes are highlighted. It is important to note that this document does not intend to be a comprehensive review of these themes, nor indicate complete alignment between all participants in all topics, but rather reflect the discussions and primary highlights made by experts during the R7+ meeting.

Discussion emphasized that the highest return on public investment in AI for science comes from people, rather than from hardware or individual models alone. This theme surfaced throughout discussions. Dialogue highlighted the need for investments to center on developing increased AI literacy among researchers

The lack of AI literacy amongst researchers generally leading to either over trust on AI systems or refusal to use AI on principle was discussed, with healthcare standing out as a recurring example. It was noted that both positions represent symptoms of the same deficit: insufficient understanding of how these systems work, where they fail, and how they should be challenged and used.

In response to these challenges, participants at the R7+ meeting highlighted that AI literacy must go beyond a list of tools and should entail:

• Basic understanding of how modern AI models are trained and evaluated
• Awareness of hallucinations/confabulations, reward hacking, and bias
• Critical evaluation of the outputs of AI models and systems: comfort with checking, debugging, and demanding evidence

AI literacy should be about epistemic discipline, with an increased understanding of AI models, how they work, and their possibilities and limitations, and not just familiarity with AI systems and how to interact with and use them. Researchers argued for the need to emphasize and encourage development of socio-technical skills such as skepticism, debugging, uncertainty awareness, documentation and reproducibility – not just coding.

Additional concerns were also expressed regarding the changing role of AI in the scientific process itself whereby there could be a transition from AI-as-a-tool to AI-as-a-scientist. Historically, AI was used and understood as a tool to contribute to scientific research: a way to automate classification, regression, or pattern recognition tasks. Today, however, the role of AI in scientific research is evolving, becoming increasingly agentic1 and used in various tasks such as generating and debugging code, proposing hypotheses or mechanisms, designing experiments or simulation regimes, and orchestrating automated laboratories where simulation and physical experimentation are tightly coupled, to name a few. In some cases, AI systems have even been involved in identifying genuinely novel regularities in behavioural or biological data. Experts noted that this evolving role of AI raises the question of whether AI systems are beginning to act, in limited ways, as “scientific collaborators.” On this, experts cautioned, however, that current generative models are not reliable epistemic agents. They lack calibrated uncertainty, may “reward-hack” their training objectives, and struggle with consistency and reproducibility, especially outside narrow quantitative tasks where formal verification is possible. Consequentially, if AI is to play a larger role in research and innovation, the standards of evidence and verification around AI-generated insights should be commensurately higher, and it was noted that when used in research, AI-generated ideas should be subject to the same standards as human-generated ideas. AI could be subject to scrutiny by, for example, using automated theorem provers and proof assistants to verify mathematically oriented claims, requiring fully rerunnable code and data pipelines for computational results, or demanding explicit uncertainty estimates and robustness checks for AI-suggested hypotheses or models. Further, it was also proposed that the boundaries between human and AI contributions should be clearly outlined to support transparency, reproducibility and accountability. In short, if AI is to be welcomed as a scientific partner, it must play by scientific rules, being subject to high standards of verification.

The increased use of AI in scientific settings to support discovery and research can contribute advantages, but as AI takes on an increasingly agentic role and takes over tasks previously conducted by human researchers, additional concerns relating to deskilling were also expressed. The risk of deskilling through the gradual erosion of human competence when AI systems take over tasks that used to provide essential training and feedback was discussed while drawing parallels with earlier technologies such as calculators, spreadsheets, and autopilot systems. In these cases, experts noted that the introduction of automation changed the role of humans from performing a task to monitoring a task. While the use of technology has been beneficial in some cases, in others, particularly where human intervention is needed in rare but critical failures, overreliance can be dangerous. With AI systems that can produce plausible sounding answers to almost any question, the concern is broader. Experts raised three distinct risks:

1. Loss of critical thinking: If AI systems are routinely used to generate arguments, interpretations, or even critical reflections, students and researchers may gradually lose the habit of struggling through complexity themselves. Further, if critical thinking erodes, consequences may not be confined to science with implications in wider society, media literacy and social cohesion2.
2. Automation bias and complacency: “Human-in-the-loop” workflows can produce a false sense of security. The human role shifts from problem-solving to monitoring, but monitoring is psychologically much harder and more fragile than designers often assume.
3. Educational exposure: There was concern about the effects of unrestricted AI use among younger users including children and adolescents, who are still developing basic skills, reasoning and identity. Additionally, emotional attachment to conversational systems, noted in some studies, adds another layer of risk.

Experts urged leaders to be explicit and intentional about how AI is integrated into training and assessment and when it should not be integrated to foster the development of certain skills and abilities. The uncomfortable but important point is that convenience and competence can often be in tension. Uncritically maximizing the former can gradually hollow out the latter.

Suggestions:

• AI literacy
  • Establish education and training practices that teach skepticism, debugging, and responsible verification. This is important for all domains, but particularly so in programs that train future clinicians, engineers, and policymakers.
  • Encourage pedagogical practices where students deliberately induce and then diagnose AI failures. Learning to spot hallucinations and flawed reasoning should be part of AI literacy to minimize automation bias and complacency.
• Transition from AI-as-a-tool to AI-as-a-scientist
  • Encourage the development and adoption of tools that make AI augmented science more reproducible through formal verification, containerized pipelines, and systematic logging of model prompts and outputs used in analyses.
  • Promote a research culture in which AI generated ideas are subjected to the same, or higher, levels of scrutiny as human generated ones, particularly in safety critical domains.
  • Insist that the boundary between human and AI contributions be documented. This is important for accountability, credit, and understanding where responsibility lies if things go wrong.
• Deskilling concerns
  • Recognize that some activities are valuable because they are hard. Exercises in educational settings designed to build reasoning, problem-solving, or ethical reflection may need boundaries where AI is sometimes excluded.
  • In professional domains, favour system designs that periodically exercise human skills, rather than relegating humans to passive oversight of inscrutable systems.

In addition, and in connection to the need to develop appropriate systems that ensure that greater AI literacy is fostered, including the development of socio-technical skills, and that deskilling concerns are addressed, experts also focused on the importance of building interdisciplinary research teams capable of addressing new challenges from a socio-technical perspective. Notably, the concept of “team science” was prevalent and stood out as a critical aspect of excellence in scientific research, ensuring a wide range of perspectives, true inclusion of all members of a team and recognition of various contributions from all roles in a research team.

Current training still overemphasizes technical skills and underweights the sociotechnical aspects of AI: how systems interact with institutions, policy, incentives, and human psychology. For example, several speakers noted that while computer scientists may be well trained to optimize models, they may not necessarily be trained to reason about the social systems into which those models are deployed. Conversely, domain experts often have deep contextual knowledge but lack the vocabulary to meaningfully shape AI design. In consideration of this, team science and in particular interdisciplinary teams can be a means to strengthen research and should be treated as the default unit of excellence of AI-enabled discovery.

In relation to interdisciplinarity and team science, however, two main challenges were also noted. First, interdisciplinary work remains harder to do, harder to evaluate, and less rewarded. For instance, graduate students who straddle AI and a specific research domain may face a dual standard: they must satisfy two communities whose evaluation norms rarely align. Second, in team science there is a need to recognize contributions of all team members, including personnel with supporting roles, such as research engineers and software developers who maintain tools, pipelines, and data infrastructure. These roles are often funded precariously, treated as auxiliary, and poorly recognized in traditional academic metrics. It was strongly suggested that these research-supporting roles be elevated and that all roles within a team be appropriately and meaningfully recognized within an inclusive team science approach. Further, it was also proposed that evaluation systems recognize not only papers, but also datasets, reusable software and well-documented pipelines as first-class contributions. These are significant contributions to the scientific process which are not always appropriately reflected or recognized when using traditional markers of research excellence in evaluations.

Without a human-centered approach to the integration of AI in research, teams may be less likely to solve the evolving socio-technical challenges associated with AI. A human-centered approach may also increase the ability of all persons to effectively participate in, and benefit from, cutting-edge science.

Suggestions:

• Treat team science as the normal unit of excellence for AI-enabled research including interdisciplinary expertise and recognizing the different contributions made by all team members.
• Acknowledge and elevate research-enabling roles, such as research engineers and data stewards, as critical to the scientific enterprise. In line with this goal, evaluation systems should recognize not only papers, but also datasets, reusable software, and well documented pipelines as first-class contributions. This will further support a team science approach.
• Support interdisciplinary work by creating environments where interdisciplinarity is not a career risk. These efforts could include, for instance, developing criteria that make it possible to judge the quality of difficult-to-classify work that exists between disciplines.

Another theme highlighted a concern regarding model centric and “so-called hype driven”3 research agendas. AI experts during the meeting emphasized the need for public funders to also prioritize impact over novelty of models: A focus on the actual field impact, including applications where the problems being addressed and the users are clear and concrete, should be prioritized over the funding of larger novel models.

Size and scope of models, and particularly their fitness of use for the task sought, attracted attention and interest following a presentation from one participant who discussed the relative merits of large vs. small models. On one end of the spectrum, the presenter discussed larger models, described analogically as “elephants”, which are powerful and more general in their applications. These models tend to attract bigger headlines and investor capital but are often more expensive and more energy intensive. On the other end of the spectrum, the presenter described smaller models, presented as “bees” in the same analogy, which are more specialized. These models are designed for specific scientific or societal problems, and although seemingly more modest, they are well suited to address and solve concrete problems; additionally, these models are often less expensive and less resource intensive. While both models can be valuable tools depending on application, it was suggested that significant public value will come from a system including a few large models and many smaller, efficient, purpose-built tools.

Experts addressed the role of public research funders and emphasized the need to ensure that proposals focus on the problems being addressed and the people who will be interacting with the AI system and/or impacted by its applications rather than the models used, noting that impact is clearest when problems and users are concrete. Funders were advised to encourage evaluations that are grounded in the real-world conditions under which an AI system will operate, (considering factors such as changes in data overtime compared to the original training data source, effects of how humans will operate with the system which could influence future outputs, cost constraints, and failure consequences, etc.), and evaluations which support the selection of simpler or older models, rather than more novel systems, when these deliver better outcomes. Public research systems should avoid following “hype cycles” which often fund projects designed to showcase generic platforms rather than solve specific problems.

Some concrete examples noted during discussion included:

• Environmental modelling: In this case, it was identified that in a real-world setting, “smart regression” or simpler emulators outperformed more novel deep models.
• Industrial and municipal control problems: When considering solutions for wastewater plant odor management or the optimization of vertical farming, it was found that success depended less on a novel model architecture and more on access to data, domain expertise, and careful problem formulation.
• Health applications in prevention and primary care: It was discussed that areas relating to health applications in prevention and primary care have received less attention than areas considered like drug discovery, despite their potential for population level impact.

Suggestions:

• Reorient attention and question framing from the how (i.e., “What model shall we use?”) to first understanding the who and why (i.e., “What problem are we solving, and for whom?”).
• Funders should encourage evaluations which include older or simpler methods when these deliver better outcomes. Funders and evaluators should focus on the impact of a proposal rather than solely on the novelty of the methodology.
• Funders should also encourage evaluations that are grounded in deployment conditions (i.e., real-world conditions in which a model will be deployed).
• Resist letting benchmarks and leaderboards become proxies for value. A new state‑of‑the‑art score on a narrow benchmark is not helpful if it does not translate to safer, fairer, or more efficient systems in the field.

Across different disciplines, including environmental science, health, and physical sciences, participants emphasized that the limiting factor for the implementation of AI innovations is often not a missing algorithm but the lack of trusted, usable, shared infrastructure including:

• Data infrastructure: interoperable datasets with clear provenance, documentation, and governance. Findable, Accessible, Interoperable, and Reusable (FAIR) data for AI models and systems for on-demand access.
• Software infrastructure: mature, inference services and maintained tools for data management, modelling, monitoring, and evaluation. Enable interoperability across AI systems, data and other digital services.
• Compute infrastructure: appropriate capacity, at the right place in the pipeline (for many use cases, modest compute near the problem is more valuable than cutting-edge hardware far away), and appropriate capacity for the training and inference of models.

In particular, experts highlighted the importance of trust in addition to access. Trust represents an essential component of AI for science; this quality should be encouraged/supported by systems that are understandable, criticizable, and revisable by the communities they serve. This was highlighted as an aspect particular importance for fields where sensitive personal or proprietary data is involved. In the case of scientific and health data, for example, there is significant concern about feeding content into opaque systems that may memorize and could subsequently leak proprietary or personal information. Researchers noted that in these and many other cases, trust can begin to be built when users are involved in problem selection and system design rather than being given a product at the end of the process. Further, experts also suggested approaches such as: retrieval augmented generation (e.g., systems that optimize output by drawing from external data sources beyond the data they were originally trained on); agentic patterns (e.g. AI agents that can on- demand, use other data and services); and “compute-to-data” models which allow access to data without the need to transfer the data from its original location, as more acceptable patterns that should be considered to foster trust and protect data privacy.

In the case of the use of AI in the research process (e.g., literature review, data collection and preparation, statistical analysis, etc.) and in review processes there is a need for transparency, traceability, and clear boundaries of responsibility. In line with this, it is important to establish regulations and standards of use and experts cautioned against delegating the responsibility of responsible use of AI to individual users. Expecting each researcher, clinician, or citizen to singlehandedly understand and manage the risks of complex AI systems was described as unrealistic.

In short, experts noted that AI cannot be experienced as a “disembodied, purely technical model”; it must be addressed as a socio-technical construct, that is, existing and operating as part of a system that includes institutional incentives, regulations, and human relationships. Establishing clear guidance and regulations and ensuring early participation from stakeholders during development of applications can contribute significantly to establishing trust to continue to advance progress through the use of AI.

Suggestions:

• Approach infrastructure with sociotechnical considerations in mind. Technical considerations such as funding, storage, compute infrastructure, and tools are necessary, but not sufficient. Consideration of the social systems in which AI models will operate, including considerations such as governance, documentation, maintenance commitments, and participation from affected communities, is equally important for a comprehensive understanding.
• Encourage privacy preserving patterns such as retrieval-augmented systems, agentic AI approaches, and compute-to-data designs, especially where sensitive or high value scientific data are involved.
• Support norms for transparency when AI is used within scientific workflows, making it visible when, where, and how AI has shaped a result, and preserving audit trails.
• Support efforts to enable interoperability across AI systems, data and other systems, and a global ecosystem of scientific AI-ready FAIR (Findable, Accessible, Interoperable, and Reusable) data.

The environmental implications of AI were also widely discussed, particularly in the environmental and earth sciences break-out groups.

Participants noted that the energy usage of large-scale AI is already comparable to that of certain countries. Training and serving frontier scale models draw on vast computational resources, often powered by electrical grids heavily reliant on fossil fuels. Meanwhile, those same models are frequently deployed for marginal or cosmetic applications. In contrast, many of the smaller models, or “bee-like” applications, discussed (e.g., weather emulators, control systems, timeseries forecasting, optimization for industrial processes) could address specific goals and research needs while delivering environmental benefits since they consume relatively modest energy in comparison.

Experts highlighted that the key variable is not whether AI consumes energy but whether that consumption is: measured and transparent; proportional to the scientific or societal value produced; and directed toward applications that support sustainability. In relation to this, it was proposed that energy and emissions reporting in research contexts be normalized, especially for publicly funded projects. Although exact reporting is difficult due to imperfect tools to establish clear estimates, the goal is not to deliver a perfectly accurate measure from the beginning, but to encourage and emphasize the importance of reflecting on these questions. In short, publicly funded AI for science should be aware of the costs, considerations and inputs required to achieve measurable benefits.

Suggestions:

• Treat energy and emissions accounting for AI as part of an expanded discussion and evaluation of research excellence, including ethics and integrity. Over time, it should become as natural to ask about computational cost as it is to ask about consent procedures4.
• Pay attention to where energy comes from, not just how much is used. The same computation has different implications in different grid contexts.
• Support research and deployment choices that favour simpler models, and methods with lower energy consumption requirements when they adequately address the scientific task.

Data recurred as a central theme throughout the day’s discussions. Experts discussed primarily the challenges relating to data sharing, and the importance of sharing benefits with relevant communities and stakeholders.

Three main concerns relating to data sharing were highlighted within the context of AI for science and research:

• Data fragmentation and incompatibility: Different institutions, sectors, and countries use multiple schemas, formats, and standards. Even when there is willingness to share, there exist practical barriers that impede open sharing of data.
• Governance uncertainty: Institutions are frequently unsure what they are legally allowed to share, under what conditions, and with which safeguards. This uncertainty, due in part to a lack of clear guidance, can lead to fears of being blamed for data misuse and subsequently foster risk‑aversion which can impede progress in sharing of data.
• Misaligned incentives: Competing priorities in relation to enhancing data privacy, security, and maintaining a competitive advantage can further discourage proactive data sharing, even when there are significant benefits to be gained.

In response to these concerns, participants argued for the need to have clarified standards, governance and incentives and noted that in some sectors, task forces such as the UK’s Energy Data Task Force have demonstrated that coordinated efforts can achieve this to contribute to eliminating the barriers to data sharing. Further, it was emphasized that the quality of AI for science depends on strong data governance to ensure scientific excellence and rigour.

A particularly important thread focused on the need to clearly communicate the benefits of data sharing with relevant stakeholders. It was noted that communities that contribute data, whether patients, local communities, or companies, should see tangible benefits from responsible reuse, rather than feeling that value is being extracted and used elsewhere.

Suggestions:

• Encourage “compute-to-data” and federated approaches that allow models to be trained or run where data reside, protecting data privacy and security while reducing incentives to centralize sensitive information.
• Support the development and adoption of interoperable data standards in key sectors, with clear guidance on provenance, consent, and permissible uses.
• Treat communication of impact and sharing of benefits as integral, not optional. Whether through improved services, shared tools, or other mechanisms, contributors must be able to see the gains from data use.

In consideration of the suggestions proposed by AI experts and researchers and following discussions during the R7+ meeting, leaders of the R7+ organizations noted that public research organizations and funders remain uniquely positioned to respond to the challenges relating to the use of AI for science and research by shaping norms, building shared capacity, and anchoring AI in the values of open, impactful and responsible science. Against this backdrop, three concrete, near‑term initiatives, were identified:

1. An R7+ Joint Talent Hub for AI in Science, launched as a prototype AI Research Commons (ARC) Pilot.
2. A joint workstream on shared research data standards and federated science data spaces, explicitly grounded in Data Free Flow with Trust (DFFT).
3. Joint principles for the responsible use of AI in research.

The following sections will discuss each of these initiatives in more detail. It is acknowledged that it may not be feasible to accomplish each item in every workstream; thus, the scope is aspirational and not binding. Nevertheless, these three initiatives create a coherent policy → program → platform pathway. The Talent Hub / ARC Pilot provides the people and governance layer; the data standards work, rooted in the existing infrastructure of the DFFT, provides the information layer; and the responsible‑use principles provide the normative layer. Together, they help the R7+ to take a step from a coordination forum into a functioning multilateral research commons, designed to be credible in the OECD, legible and attractive to Global South partners, and complementary to other G7 and multilateral processes.

Leaders proposed to establish a Joint Talent Hub for AI in Science and to frame it explicitly as the first phase of an R7+ AI Research Commons (ARC) Pilot: a multilateral, standards‑driven framework within which people, problems, compute access and datasets can interoperate under common rules.

• Co-led by Canada, France, Germany, Italy, Japan, the United Kingdom and Spain, with Canada and Spain indicating possible initial resource availability.
• The Talent Hub would serve as the operational nucleus of an R7+ AI Research Commons (ARC) Pilot: a light‑weight, standards‑oriented framework in which people, problems, compute and data can interoperate under common rules.
• It would broker portable fellowships, co‑supervised PhD and postdoctoral positions, and short “researcher‑in‑residence” stays, supported by a talent‑matching system proposed by Spain, building on CSIC’s experience with large‑scale skills and mobility programs.
• France and Italy confirmed their readiness to host major AI‑for‑science workshops in 2026 under the Talent Hub umbrella.
• Framed in this way, the Talent Hub becomes a prototype multilateral research commons: open by design to partners beyond the G7, including the Global South, through portable fellowships, capacity‑building, and preferential access to shared resources.

The R7+ Joint Talent Hub is conceived as the operational nucleus of a future AI‑for‑science Research Commons (ARC) Pilot, designed to:

• Accelerate scientific discovery by connecting talent and research initiatives across borders and disciplines;
• Reduce duplication of infrastructure and program spending, by making better use of existing investments in training, data, and compute; and
• Provide a credible, non‑extractive on‑ramp for partners beyond the R7+, including Global South institutions, through portable fellowships and capacity‑building rather than one‑way talent drains.

It responds directly to the experts’ diagnosis that human capacity, not just compute, is the rate‑limiting factor in AI for science.

The Talent Hub/ARC Pilot will be built on four design principles:

• People‑first interoperability 
  Portable fellowships, co‑supervised doctoral and postdoctoral positions, reciprocal visiting‑researcher status, and cross‑appointment templates will allow researchers and research engineers to move across R7+ institutions without ad‑hoc renegotiation each time.
• Data and model commons 
  Working alongside the second agreed workstream, participants will discuss minimum viable, ISO‑aligned schemas for describing datasets, models and evaluations (dataset “cards”, model “cards”, evaluation manifests), with provenance and consent “by design”. Registries will catalogue what exists without forcing centralization.
• Compute reciprocity 
  A time‑bank mechanism would allow participants to exchange access to HPC and AI clusters, with a preference for green‑energy sites and transparent carbon accounting for AI‑intensive workloads linking directly to the “Greening AI” agenda raised by several partners.
• Open protocols, diverse implementations 
  The commons would define interfaces and rules, not a single platform; national and regional nodes remain fully sovereign over their infrastructure and the details of local implementation.

Within twelve months, the Talent Hub / ARC Pilot aims to deliver:

• Talent matching and portable fellows 
  A Spain‑led talent‑matching system would connect researchers, engineers and institutions around problem‑led cohorts (for example, AI for climate modelling, materials discovery, health), co‑funded by participating agencies and governed by a shared code of conduct and streamlined ethics and security clearances.
• A “standards starter kit” 
  An initial set of common templates (dataset cards, model cards, evaluation manifests), aligned with the data standards workstream and ISO liaisons, with at least one reference implementation per template.
• A compute access protocol 
  A time‑bank and credit system for access to participating HPC/AI systems with basic telemetry for reproducibility and energy reporting.
• Problem‑led “challenge briefs” 
  A small number of jointly owned challenge briefs, initially in areas such as climate, health and materials, co‑designed with mixed OECD-Global South teams, with pre‑agreed data governance arrangements.
• A commons registry 
  A registry (with open APIs) of participating people, projects, datasets, models and evaluations, making the ARC visible and findable.
• Workshops
  Two flagship ARC-branded workshops will be offered in 2026 by France and Italy, under the Talent Hub umbrella. Themes of the workshops are subject to final confirmation.

• Co‑led by Canada, Italy, the United States and Germany, with France, the United States and Italy identifying this as a priority area, and the United States indicating possible resource availability to support coordination.
• This initiative is linked to the DFFT principles. Coined by Japan and elevated at G20 Osaka (2019), then picked up by G7 roadmaps/action plans, DFFT promotes cross‑border data use while safeguarding privacy, security, and IP. It is not a single set of treaty obligations, but an interoperable toolkit (principles, standards, certifications, governance).
• The workstream proposes to map and align existing national and regional data spaces, then launch a pilot R7+ “Science Data Mesh” in selected domains (e.g. health, materials, etc.), combining:
  • Provenance‑rich, catalogue‑based metadata, enabling discovery and understanding of datasets even when they cannot leave their jurisdiction;
  • Tiered access and interoperability mechanisms (legal, organizational and technical) that allow queries and models to travel while sensitive data remain within local governance regimes, with differentiated access depending on sensitivity and role; and
  • Federated computation, in which models or analyses are run close to where the data reside and only aggregated updates, parameters or proofs are exchanged.
• In short, data remain governed where they live; queries, models and proofs travel.

Originally articulated by Japan and subsequently taken up by G7 and G20 roadmaps, DFFT seeks to enable cross‑border data use with robust safeguards for privacy, security, intellectual property and other legitimate public interests. This second initiative addresses concerns relating to shared research data standards, with the explicit aim of making Data Free Flow with Trust (DFFT) a lived reality in scientific collaboration.

It is important to note that this work does not attempt to create a monolithic new infrastructure. Rather, it federates and standardizes what already exists, guided by DFFT, so that science benefits from data‑driven AI without undermining national or community control over sensitive information.

1. Catalogue and label existing research datasets and data spaces 
   The first step is to map major national and regional data infrastructures (e.g. European data spaces, Germany’s International Data Spaces, Canada’s National Data Spaces, CSIC’s Science Data Space, CNR and CNRS repositories, health and climate data hubs) and to expose searchable registries and APIs that at minimum reveal metadata and access conditions.
2. Define a minimum common “science data schema” 
   Working with ISO and other standardization bodies, the R7+ will discuss a minimal set of fields (e.g., identifiers, schema, sensitivity level, lawful basis, consent conditions, and usage constraints), that participating datasets will expose. This will be tightly coupled with the “standards starter kit” in the ARC Pilot.
3. Launch a pilot R7+ “Science Data Mesh” 
   In a small number of domains (for example, materials science and health among others), participating institutions plan to implement federated analysis pipelines and demonstrate that partners can collaborate under DFFT constraints, with full audit trails of model lineage and data access.
4. Develop an assurance and certification layer 
   Building on G7 and OECD work, the R7+ will prototype science‑specific certifications or codes of conduct (for instance, for health data, environmental data, or Indigenous data), enabling researchers to demonstrate compliance and giving partners confidence in cross‑border collaboration.

• Co-led by Germany and the United Kingdom, with Canada, Spain and the European Union signaling strong interest and the UK and Germany indicating possible resource availability.
• Leaders propose to develop an R7+ statement of principles on responsible AI in research that:
  • Builds on existing frameworks (OECD AI Principles, UN, UNESCO, GPAI, regional AI acts), but focuses specifically on research settings, from AI‑assisted data analysis to AI‑enabled experimentation, simulation, and research management;
  • Sets expectations for disclosure, documentation and “safety cases” when AI tools influence research outputs, including guidance on when AI use is inappropriate or prohibited; and
  • Lays the foundation for future work on shared assurance methods, evaluation suites and lab audit practices for AI‑intensive science.

• Distil a shared understanding of “responsible AI in science”, encompassing transparency, explainability, bias mitigation, accountability, safety, reproducibility and respect for human rights and research integrity.
• Provide practical, discipline‑sensitive guidance on how AI may appropriately be used at each stage of the research lifecycle (from data collection and analysis to writing, peer review and research management), and when its use is inappropriate or prohibited (for example, in certain forms of high‑stakes inference without human oversight).
• Clarify expectations around disclosure and documentation; for instance, how researchers should report AI involvement in methods sections, and what constitutes a sufficient “safety case” when AI plays a material role in generating or validating scientific claims.
• Lay the groundwork for shared assurance and evaluation infrastructures, including common test suites for validity, robustness and reproducibility, and “fitness‑for‑purpose” assessments in regulated domains such as health and environment.

The leaders’ discussion, with input from AI experts and other participants, focused on forward-looking actions and collaborations. Main takeaways from the discussion are summarized in this appendix.

Leaders acknowledged the experts’ view that people, not compute, are the binding constraint in AI‑enabled science. They stressed the importance of:

• Treating team science as a highly valued pursuit, with explicit recognition for research engineers and data stewards;
• Normalizing increased AI literacy across disciplines, including socio-technical expertise and guidance on when and where not to use AI; and
• Guarding against deskilling, for instance by creating AI‑free spaces in education and training where deep reasoning and independent problem‑solving are cultivated.

In response to the observation that societies generally trust scientists but not AI, leaders argued that cultural norms of transparency are as important as formal regulation. Scientists should openly disclose when and how AI is used in research workflows.

Leaders discussed that the R7+ should focus on shared values and norms, rather than attempting to regulate science directly, and that the responsible‑use principles should be framed as enforceable expectations within public funding regimes.

Intergenerational differences in technology use were acknowledged as a serious challenge, with the conclusion that AI researchers alone cannot resolve them; education experts and wider societal stakeholders must be involved.

Drawing on the “elephants and bees” metaphor (large, general-purpose, foundational models vs. much smaller, specialized, AI models), leaders discussed that large general‑purpose models are not always the right tool for specific scientific questions, and that claims of “state‑of‑the‑art” performance should be viewed critically unless backed by deployment‑relevant evaluations.

Energy considerations were prominent. Some participants discussed requiring major AI heavy research proposals to estimate energy use and environmental impact, thereby encouraging smaller, more efficient models when scientifically sufficient. This perspective will be embedded in the ARC’s compute reciprocity and the data standards workstream’s carbon accounting features.

Leaders also acknowledged a substantial “evaluation and interpretability debt”. Even with open code, the internal workings of complex models remain only partially understood, and their behaviour in socio‑technical systems is difficult to predict. Public funding should support mechanistic interpretability and evaluation science as core components of AI‑for‑science programs.

The discussion on data recognized both the practical fragmentation of research data (e.g., heterogeneous formats, inconsistent documentation) and the political realities of data and compute sovereignty, including research security and Indigenous data governance.

It was suggested that DFFT be employed as the overarching diplomatic frame: ensuring that data remain governed where they reside, while queries, models and proofs move across borders under agreed standards and consent regimes. The data standards workstream is tasked with turning this into a practical reality, starting with the catalogue‑and‑mesh approach described in Section II.2.

Leaders reflected on the influence of public funders in a landscape where private AI investment is dominant. The historical experience of public funding for neural networks in the 1980s (investments that eventually enabled today’s AI wave) was invoked as evidence that public funding can shape long‑term trajectories even when not dominant in volume.

Discussions highlighted that public funders should consider the following:

• Support high‑risk, high‑reward research without immediate commercial returns;
• Make openness, reproducibility, transparency and responsibility conditions of funding;
• Invest in safety, interpretability, evaluation and governance; and
• Use their convening power to create structures such as the ARC Pilot, the DFFT‑aligned science data mesh and the joint responsible‑use principles, which can influence norms across both public and private sectors.

Leaders also acknowledged that the meeting closes not just with three specific workstreams, but with a broader sense of the R7+ as a track 2 diplomacy platform: one that can turn convergent values and complementary strengths into concrete actions towards developing shared capacity for AI‑enabled science, and that is intentionally designed to scale to a wider circle of partners in the years ahead.