AI model access for safety research
Lack of access for AI Safety Institutes to advanced AI models for safety research, including developers refusing access.
Source spread
Where this theme appears
This theme appears across 4 independent accountability sources, so the source mix matters as much as the headline total.
37 committee recs
1 HSSIB rec
Browse by source
Source-grouped records are useful for tracing where a concern came from. Large sections show the 50 strongest matches for that source; counts still show the full theme total.
Select committee recommendations(37)
#8 —
Recommendation: Longer term, the ONS must still evolve to address the pressures it faces, including the need to make greater use of administrative data, the difficulties of relying on traditional surveys, demands for new data in policy responses, and the impact …
#33 —
Recommendation: We appreciate that the definition of sovereignty will differ by sector and by degree. However, the general lack of clarity on definitions and objectives complicates research and development planning and investment. The UK has a strong base of talent and …
Response attribution not verified
#45 — Insufficient details provided on access to dedicated AI Research Resource and supercomputer cluster.
Recommendation: We welcome the Government’s moves to establish a dedicated AI Research Resource and a cluster of supercomputers but are concerned that it has yet to set out further details of how researchers and startups will be able to access the …
Response attribution not verified
#26 — Confirm models the AI Safety Institute could not access and name refusing developers.
Recommendation: The Government should also confirm which models the Institute has been unable to secure access to, and the reason for this. If any developers have refused access— which would represent a contravention of the reported agreement at the November 2023 …
Response attribution not verified
#24 — AI Safety Institute's limited access to unreleased models undermines safety testing capabilities.
Recommendation: However, we are concerned by suggestions that the Institute has been unable to access as-yet unreleased AI models to perform the pre-deployment safety testing it was set up to undertake. If true, this would undermine the delivery of the Institute’s …
Response attribution not verified
#28 — Require generative AI providers to share internal data with independent online safety researchers.
Recommendation: Principle 5 is crucial for addressing potential harms from generative AI, as there is currently a serious shortfall in transparency and oversight of the platforms and systems that allow users to create AI-generated content. The government should require providers of …
Response attribution not verified
#9 — Commission independent research into social media algorithms amplifying harmful content with full data access
Recommendation: There is a shortfall in data needed to accurately analyse the scale of the problem and identify policy solutions. In line with our Principle 4, the government should commission a large-scale research project into how far social media recommendation systems …
Response attribution not verified
#7 — Social media technology is insufficiently transparent and accessible to public authorities
Recommendation: The technology used by social media companies should be transparent, explainable and accessible to public authorities, as stated in our Principle
Response attribution not verified
#46 — Publish AI Research Resource action plan and undertake National Compute Cluster feasibility study.
Recommendation: The Government, or its successor administration, should publish an action plan and proposed deliverables for both the AI Research Resource and its cluster of supercomputers, and further details of the terms under which researchers and innovative startups will be able …
Response attribution not verified
#44 — Facilitate UK AI startups' access to anonymised public datasets through a National Data Bank.
Recommendation: The Government and future administrations should support the emergence of more AI startups in the UK by ensuring they can access the high-quality datasets they need to innovate. This could involve facilitating access to anonymised public data from data.gov.uk, the …
Response attribution not verified
#32 — Scientific research and development are crucial for advancing AMR diagnostics, treatments, and vaccines.
Recommendation: Taking advantage of scientific research and developments will be very important in addressing the threat of AMR. Research and development can improve tools such as diagnostics to improve prescribing of antimicrobials, as well as producing new antimicrobials and vaccines.82 It …
Response attribution not verified
#6 — Mandate health bodies to keep updated on scientific AMR developments and fund vital research.
Recommendation: It is vital that government keeps up to date with scientific developments that might help tackle the threat of AMR. Research can help identify better ways of mitigating AMR and treating those with AMR infections, including through the use of …
Response attribution not verified
#13 —
Recommendation: The government should set out how it intends to protect its sovereignty in AI, in particular, in the face of fast-changing geopolitical and technological developments. The government should work with industry and academia to define sovereign capability for the five …
Response status not verified
#12 —
Recommendation: The US’s decision to restrict some advanced AI models should be a powerful reminder to the UK government that it may not be able to count even on its allies for access to technologies. Leverage may not be enough to …
Response status not verified
#34 —
Recommendation: The Government should provide a clear, written definition of what a sovereign capability is, including different levels of sovereignty. The Government should also outline which specific existing and emerging technologies it is seeking to develop, and to what level of …
Response attribution not verified
#5 —
Recommendation: Former Chief Constable Guildford was not informed ahead of giving oral evidence on 6 January that Microsoft Copilot AI had been used to generate the erroneous information about a match between West Ham and Maccabi Tel Aviv. On this basis …
Response attribution not verified
#50 — Utilise best-suited AI models for public sector deployment when procuring solutions.
Recommendation: When procuring AI models for deployment in the public sector the Government and public bodies should utilise those best suited to the task.
Response attribution not verified
#40 — Dominance of leading AI developers stifles free and open competition in the market
Recommendation: At the so-called ‘frontier’ of AI a small group of leading developers are responsible for and accruing significant benefits from the development of advanced models and tools—thanks in part to their ability to access the necessary training data. This potential …
Response attribution not verified
#37 —
Recommendation: The Government should prioritise the diffusion and adoption of innovation and technologies across the UK. This will require greater emphasis on building skills for smaller businesses; increased advice and support on regulatory matters for smaller businesses; making finance available to …
Response attribution not verified
#41 — Require CMA to identify and address abuses of market power by AI developers
Recommendation: As the regulator responsible for promoting competitive markets and tackling anti- competitive behaviour, the CMA should identify abuses of market power and use its powers to stop them. This could take the form of levying fines or requiring the restructuring …
Response attribution not verified
#25 — Confirm models tested by AI Safety Institute, testing details, findings, and developer changes.
Recommendation: In its response to this Report, the Government should confirm which models the AI Safety Institute has undertaken pre-deployment safety testing on, the nature of the Governance of artificial intelligence (AI) 55 testing, a summary of the findings, whether any …
Response attribution not verified
#16 — NHS Resolution exploring AI to analyse negligence claims data for insights.
Recommendation: Some clinical negligence firms are reportedly using artificial intelligence to triage claims more efficiently and effectively. NHS Resolution holds almost 30 years of experience and data concerning compensation claims.33 NHS Resolution told us it is starting to explore how technology …
Response attribution not verified
#25 —
Recommendation: Natural England told us it has made progress in improving its guidance through adding real-life case studies so that businesses can understand what they need to do to comply.66 Defra also explained it has piloted the use of AI tools …
Response attribution not verified
#19 — Mandate deposition of research data in open-access repositories and encourage FAIR principles
Recommendation: Journals should collectively encourage researchers to employ the FAIR (Findability, Accessibility, Interoperability, and Reuse of digital assets) principles within their research and should mandate the deposition of research data in open- access repositories alongside the publication of research outputs.
Response attribution not verified
#18 — Mandate researchers to share open-source data and code alongside all published outputs
Recommendation: Currently, research outputs are frequently published without an associated link through to their open-source data and code. This prevents other researchers assessing work for its reproducibility. In all bar the most exceptional ethical and legal situations, researchers should share their …
Response attribution not verified
#17 — Implement 100% open access policies for scientific outputs by UKRI and funders by 2025
Recommendation: The trend towards blanket open access in the communication of scientific outputs is positive. UKRI and other research funders should continue to implement open access policies until this figure reaches 100%, by the end of 2025 at the latest.
Response attribution not verified
#4 — AI technology offers transformative potential for medical research, drug development, and personalised medicine.
Recommendation: The ability of AI models and tools to process substantial volumes of data, and rapidly identify patterns where human researchers might take months or be unable to, makes it a potentially transformational technology for medical research. Either through the development …
Response attribution not verified
#56 — Commission a review into AI's future skills and employment consequences, including retraining for automation.
Recommendation: The current Government, or its successor, should commission a review into the possible future skills and employment consequences of AI, along the lines of the 2017 Taylor Review of modern working practices which examined the landscape, suggested ideas for debate …
Response attribution not verified
#49 — AI marketplace requires diversity to support both open and closed-source models.
Recommendation: The open-source approach has underpinned many technological breakthroughs, including the Internet and AI. Whilst some providers of products and services, such as AI models and their applications, will want to keep elements of their offerings proprietary, a healthy AI marketplace …
Response attribution not verified
#3 —
Recommendation: Mr Gurr told us that fiscal constraints mean the CMA cannot expect to hire more people, and it must therefore make the most of those they have. Its 4P transformation programme is vital to achieving this, and he also pointed …
Response status not verified
#20 — Continued emphasis on data management plans as a condition of research funding is necessary
Recommendation: We welcome UKRI’s use of data management plans. A continued emphasis on their importance as a condition of research funding is necessary.
Response attribution not verified
#6 — AI offers benefits for teachers and pupils, but raises assessment questions for coursework.
Recommendation: The benefits for time-pressed teachers using AI models and tools to help prepare lesson plans are clear, and increased availability of personalised learning and tutoring tools could benefit many pupils. However, widespread use of AI raises questions about the nature …
Response attribution not verified
#3 — AI models demonstrate transformative potential in healthcare diagnostics and professional time-saving.
Recommendation: AI models and tools can transform healthcare provision, by assisting with diagnostics and, perhaps more significantly, freeing up time for the judgement of medical professionals by automating routine processes.
Response attribution not verified
#42 — High-quality input data is crucial for effective AI models and tools
Recommendation: AI models and tools rely on access to high-quality input data. The phrase ‘garbage in, garbage out’ is not new, but it is particularly applicable to AI.
Response attribution not verified
#7 — Prioritise equipping children with digital literacy and critical engagement skills for ubiquitous AI.
Recommendation: Education policy must prioritise equipping children with the skills to succeed in a world where AI is ubiquitous: digital literacy and an ability to engage critically with the information provided by AI models and tools.
Response attribution not verified
#5 — AI tools already provide useful time-saving benefits for education professionals and students.
Recommendation: AI tools are already useful time-savers for education professionals, and whilst reliable data is hard to come by, it seems highly likely that the technology is this generation of students’ calculator or smartphone.
Response attribution not verified
#2 — AI as a tool, not sentient machines, guided by human instruction.
Recommendation: Nevertheless, the technology should not be viewed as a form of magic or as something that creates sentient machines capable of self-improvement and independent decisions. It is akin to other technologies: humans instruct a model or tool and use the …
Response attribution not verified
HSSIB safety recommendations(1)
Independent reviews(26)
Fisher Review (Part 2) — Rec 19
Funding should be prioritised to support the implementation of artificial intelligence by law enforcement agencies in the proactive detection and disruption of fraud.
Justice & Legal
Hall-Pesenti AI Review — Rec 13
Government should work with industry and experts to establish a UK AI Council to help coordinate and grow AI in the UK.
Other
Hall-Pesenti AI Review — Rec 12
The Alan Turing Institute, Engineering and Physical Sciences Research Council (EPSRC), Science and Technology Facilities Council (STFC) and Joint Information Systems Committee (JISC) should work together to coordinate demand for computing capacity for AI research, and negotiate for the UK research community.
Other
Hall-Pesenti AI Review — Rec 10
The Alan Turing Institute should become the national institute for artificial intelligence and data science, becoming truly national and expanded beyond the current five universities, with a key stated aim that centres its mission on artificial intelligence.
Other
Hall-Pesenti AI Review — Rec 9
An International fellowship programme for AI in the UK should be created in partnership with the Alan Turing Institute: the Turing AI Fellowships. This should be supported by a targeted fund for identifying and recruiting the best talent, and by ensuring that the UK is open to any and all …
Other
Hall-Pesenti AI Review — Rec 3
To support text and data mining as a standard and essential tool for research, the UK should move towards establishing by default that for published research the right to read is also the right to mine data, where that does not result in products that substitute for the original works. …
Other
Hall-Pesenti AI Review — Rec 2
To improve the availability of data for developing AI systems, Government should ensure that public funding for research explicitly ensures publication of underlying data in machine-readable formats with clear rights information, and open wherever possible.
Other
Hall-Pesenti AI Review — Rec 1
To facilitate the sharing of data between organisations holding data and organisations looking to use data to develop AI, Government and industry should deliver a programme to develop Data Trusts – proven and trusted frameworks and agreements – to ensure exchanges are secure and mutually beneficial.
Other
Goldacre Review — Rec IG 8
Establish a clinic to help users who are blocked on data access MRC has previously funded excellent work at their Regulatory Support Centre to give support and guidance to individual MRC-funded groups encountering barriers to their work around data access or usage. UKRI and/or NIHR should expand this work, or …
Health & Social Care
Goldacre Review — Rec TRE 57
Address TREs for Artificial Intelligence, but as a separate workstream, funded by existing AI resource TREs with the core features described above will readily support all analysis using traditional analytical or epidemiological research techniques. Analysis and research involving techniques that fall under the heading of Artificial Intelligence – particularly unsupervised …
Health & Social Care
Nurse RDI Review — Rec 9
New research institutes and units should be considered when strategic RDI priorities best supported by focused research missions are identified by Government, UKRI and other funders. Possible examples include enhanced activities in climate change and its mitigation, antimicrobial resistance, synthetic biology, and artificial intelligence. Themes should be identified through mapping …
Other
Hall-Pesenti AI Review — Rec 18
Government should ensure that challenges addressed by the Industrial Strategy Challenge Fund (ISCF) and Small Business Research Initiative (SBRI) are designed to attract and support applications of AI across the full range of challenge areas and set funded challenges which use public sector data for AI.
Other
Goldacre Review — Rec TRE 40
Academics should use NHS data infrastructure to access NHS patient records All academic work on NHS patient records alone should always be conducted in NHS TREs, and be compliant with RAP and open working methods. Patient data should only be transferred out to other non-NHS data centres when that patient …
Health & Social Care
Goldacre Review — Rec TRE 21
The National TRE should be open to all legitimate users Any national TRE containing national NHS data should be open to all legitimate applicants following a rules-based system including: national NHS service analysts; local NHS service analysts wanting to work on their own local data in a national context; academic …
Health & Social Care
Hall-Pesenti AI Review — Rec 17
Government, drawing on the expertise of the Government Digital Service, the Data Science Partnership and experts working with data in other Departments, should develop a programme of actions to prepare the public sector and spread best practice for applying AI to improve operations and services for citizens.
Other
Hall-Pesenti AI Review — Rec 16
TechUK should work with the Royal Academy of Engineering, the Digital Catapult, and key players in industry sectors, to develop practical guidance on the opportunities and challenges of successful adoption of AI across the UK economy.
Other
Hall-Pesenti AI Review — Rec 7
Government and universities should create, at a minimum, an additional 200 PhD places dedicated to AI at leading universities. As the UK trains and attracts additional academic talent, this number should grow continually year on year.
Other
Hall-Pesenti AI Review — Rec 5
Industry should sponsor a major programme of students to pursue Masters level courses in AI, with an initial cohort of 300 students.
Other
Goldacre Review — Rec TRE 38
Review the current Registry and Audit landscape; work towards wider access and use More broadly these datasets are likely an under-used resource, and a very dispersed set of projects, of very varying scale: it would be wise to commission a deep dive to describe for each project the data flows, …
Health & Social Care
Goldacre Review — Rec Open 43
TREs themselves should be built on principles of RAP and open code It is very reasonable for TREs to use general purpose closed commercial products (such as enterprise database tools) where these reflect the best procurement choice. However, any tooling built for the TREs themselves - especially the components for …
Health & Social Care
Goldacre Review — Rec Open 42
All Trusted Research Environments for NHS data must facilitate and require code sharing All national TREs should be designed such that researchers are able to use services such as GitHub and Gitlab; and ideally require such that code is shared by default. This must be a core part of the …
Health & Social Care
Goldacre Review — Rec S23
Have a frank public conversation about commercial use of NHS data for innovation, but only after privacy issues have been addressed through adoption of TREs; ensure the NHS gets appropriate financial return where marketable innovations are driven by NHS data, which has been collected at great cost over many decades; …
Health & Social Care
Goldacre Review — Rec S4
Use the enhanced privacy protections of Trusted Research Environments (TREs) to create new, faster access rules and processes for safe users of NHS data; ensure all TREs publish logs of all activity, to build public trust.
Health & Social Care
Gauke Review — Rec 8.5
Further collaborate with industry on research and development to explore new technologies for service transformation, including advanced AI
Justice & Legal
Hall-Pesenti AI Review — Rec 15
The Department for International Trade should expand its current support programme for AI businesses.
Other
Hall-Pesenti AI Review — Rec 11
Universities should use clear, accessible and where possible common policies and practices for licensing IP and forming spin-out companies.
Other