ARIA (UK)
Last reviewed
May 18, 2026
Sources
No citations yet
Review status
Needs citations
Revision
v1 · 4,490 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
May 18, 2026
Sources
No citations yet
Review status
Needs citations
Revision
v1 · 4,490 words
Add missing citations, update stale details, or suggest a clearer explanation.
The Advanced Research and Invention Agency (ARIA) is a United Kingdom government research funding body established by Act of Parliament in 2022 and made operational in January 2023. Headquartered in London and sponsored by the Department for Science, Innovation and Technology (DSIT), it is modelled on the United States Defense Advanced Research Projects Agency (darpa) and seeks to fund "high-risk, high-reward" scientific and technological research that conventional grant systems are seen as too cautious to back. ARIA's founding Chief Executive Officer was Ilan Gur, who served from 2022 until June 2025; its Chair is the technology entrepreneur Matt Clifford. Its initial budget was approximately £800 million over four years.[^1][^2][^3]
ARIA is frequently described in British and international media as the "British DARPA." It must not be confused with the United States nist aria programme (Assessing Risks and Impacts of AI), an evaluation and benchmarking initiative run by the National Institute of Standards and Technology. The two agencies share an acronym but differ fundamentally in purpose: UK ARIA is a funding body that commissions speculative research across many scientific domains, while NIST ARIA is a sociotechnical testing programme aimed at measuring AI risk in deployment contexts. This article concerns the United Kingdom agency only.[^4][^5]
ARIA's portfolio spans neurotechnology, climate, robotics, plant biology, energy and computing, with several programmes that are directly relevant to artificial intelligence. The most prominent of these for the AI safety community is the Safeguarded AI programme, led from its launch in April 2024 by David "davidad" Dalrymple as part of the "Mathematics for Safe AI" opportunity space, and backed by approximately £59 million in committed funding. ARIA is institutionally distinct from the UK AI Security Institute (uk aisi), which sits inside DSIT and tests frontier AI systems for capability and risk, although both organisations are part of the same broader UK government science and technology architecture.[^6][^7][^8]
The intellectual origin of ARIA lies in long-running advocacy by Dominic Cummings, a political adviser who became chief adviser to Prime Minister Boris Johnson in July 2019. Cummings had written in detail on his personal blog from 2017 onwards about the history of the United States Advanced Research Projects Agency (ARPA, later DARPA), and in particular about the period between 1962 and 1975 under directors such as J. C. R. Licklider, Robert Sproull and Stephen Lukasik, which he treated as a model of unusually productive state-funded science. Cummings argued that the United Kingdom's research funding ecosystem, then centred on UK Research and Innovation (UKRI) and its predecessor research councils, was risk-averse, peer-review-bound and structurally unable to support contrarian research bets.[^9][^10]
After Cummings entered Downing Street, the "ARPA project" became one of his publicly stated priorities. Press accounts at the time reported that his messaging-app profile summarised his agenda as "Get Brexit done, then ARPA." The Conservative manifesto of 2019 contained a commitment to "a new agency for high-risk, high-payoff research, at arm's length from government." Cummings left Downing Street in November 2020, but the idea survived his departure and was carried forward by the Department for Business, Energy and Industrial Strategy (BEIS) under successive ministers.[^9][^10]
The agency was publicly named on 19 February 2021 by BEIS, then under Secretary of State Kwasi Kwarteng, with an initial budget allocation of £800 million over the spending review period. The Advanced Research and Invention Agency Bill was introduced to Parliament on 2 March 2021. After passing through both Houses and surviving a series of "ping-pong" amendments concerning freedom of information and accountability, the Bill received Royal Assent on 24 February 2022, becoming the Advanced Research and Invention Agency Act 2022.[^11][^12]
The Act creates ARIA as a statutory corporation with a small board, gives it a ten-year statutory protection from dissolution, exempts it from the Freedom of Information Act 2000, exempts its procurement from the standard public procurement rules, and assigns its sponsorship to the Secretary of State for Business, Energy and Industrial Strategy (later transferred to the Secretary of State for Science, Innovation and Technology when DSIT was created in February 2023). It also requires ARIA to have particular regard to scientific research and inventions that may bring about social, economic or environmental benefit.[^11][^13]
Although the legislation passed in February 2022, the agency had no permanent leadership at that point. On 20 July 2022, BEIS announced the appointments of Ilan Gur as the first CEO and Matt Clifford as the first Chair. ARIA was formally established as an operational body on 26 January 2023, when its statutory functions came into force and the agency began hiring. The first cohort of Programme Directors, numbering eight, was announced in October 2023 and a second cohort of eight followed in April 2025. ARIA's first programme funding calls opened in early 2024.[^1][^2][^14]
ARIA is a non-departmental public body, sponsored by DSIT through a framework agreement that sets out its accounting and reporting obligations. Its accountable officer is the CEO. The board is appointed by the Secretary of State and includes the CEO, the Chair and a small number of non-executive directors. Unusually for a British public body, ARIA is statutorily exempt from the Freedom of Information Act 2000 and from the Public Contracts Regulations, design choices justified by ministers on the basis that high-risk research requires confidentiality and procurement flexibility, and criticised by transparency campaigners and some parliamentarians during the Bill's passage.[^11][^12][^15]
The Act also imposes a "tolerance to failure": ARIA is explicitly directed to expect that many of its bets will not work out, and the Secretary of State is required to have regard to that fact when reviewing its performance. ARIA is permitted to commission research, give grants, take equity, enter contracts and form subsidiaries, but it cannot itself confer degrees.[^13]
ARIA's operating model is borrowed directly from DARPA and other ARPA-style agencies such as the United States ARPA-E (Advanced Research Projects Agency-Energy) and the more recent ARPA-Health (ARPA-H). The unit of execution is the Programme Director, an individual technologist or scientist hired on a fixed-term contract, typically three to five years, who is given an opportunity space and a budget and asked to design and run a single multi-million-pound programme of work. Programme Directors decide which teams to fund, how to structure milestones, and when to terminate projects.[^14][^16]
ARIA staff have publicly described their use of the Heilmeier Catechism, the set of nine questions popularised by former DARPA Director George Heilmeier in the mid-1970s, as a discipline that every Programme Director is expected to be able to answer for their programme. These questions include "What are you trying to do?" (in plain language), "How is it done today and what are the limits of current practice?", "What is new in your approach and why do you think it will be successful?", and "How will progress be measured?" The catechism is part of the conceptual core of the ARPA model that ARIA has imported.[^17]
ARIA's funding instruments avoid traditional academic peer review. Programmes are designed by Programme Directors and approved internally; teams are selected by the Programme Director, often in collaboration with a small panel and with input from "Activation Partners," a network of nonprofits, research labs and venture firms that ARIA uses to source and support funded teams. This absence of peer review is one of the most significant departures from UKRI practice.[^16][^18]
Programme structure is typically broken into "Technical Areas" (TAs), themselves divided into work packages and phases with milestone gates. Funded teams are referred to as "Creators." A programme that begins with broad scoping work in a Phase 1 may be narrowed, expanded or terminated at the discretion of the Programme Director.[^7][^16]
Ilan Gur was announced as ARIA's founding CEO on 20 July 2022 and joined later that year. Gur holds a PhD in materials science from the University of California, Berkeley. He served as a Programme Director and Senior Adviser at ARPA-E between 2011 and 2014, then founded Cyclotron Road, a fellowship for science-based entrepreneurs hosted by Lawrence Berkeley National Laboratory, in 2014. Cyclotron Road's model was later spun out into the independent nonprofit Activate, of which Gur was founder and CEO until his appointment at ARIA. Gur's background combined direct experience of the ARPA model in the United States with experience of fellowship-style talent funding, both of which shaped ARIA's early operating choices.[^19][^20]
Gur announced in June 2025 that he would step down to return to the United States. Kathleen Fisher was announced as the next CEO in November 2025 and took up the post in February 2026. Fisher had previously served as Director of the Information Innovation Office at DARPA. Her appointment further deepened ARIA's operational connection to the DARPA model.[^2][^14]
Matt Clifford CBE has chaired ARIA since 2022. He is the co-founder of the talent investor Entrepreneur First. From late 2024 onwards he was also the United Kingdom government's independent reviewer for the AI Opportunities Action Plan, a strategy document commissioned by the Sunak government and published by the Starmer government in January 2025. Following publication, Clifford was appointed the Prime Minister's Adviser on Artificial Intelligence in 10 Downing Street, a role he held until stepping down in June 2025 for family reasons. He retained the chair of ARIA.[^21][^22]
Clifford's dual role placed ARIA at the centre of British AI policymaking during 2024 and 2025, since the Chair of the agency was simultaneously the principal external author of the government's AI strategy.[^22]
ARIA's initial allocation was £800 million over the four-year period to 2025-26, announced as part of the 2020 Spending Review and confirmed in the agency's framework documents. Spending was deliberately back-loaded: ARIA had to be set up, staff hired, Programme Directors recruited and programmes designed before significant sums could be committed. Reporting in 2025 by Research Professional News found that ARIA was on track to spend only a fraction of the £800 million by the end of the original spending review period, reflecting the time required to stand up new programmes rather than any reduction in commitment.[^15][^23]
In the June 2025 Spending Review, the Treasury confirmed a forward commitment of at least £1 billion for ARIA over the 2025 to 2029 period, reaffirming political support for the agency through to its statutory ten-year review. ARIA's annual accounts for 2024-25 reported around 53 staff and a year of operational spending of approximately £27.6 million.[^14][^23]
Individual programmes are sized in the range of £30 million to £100 million each. Examples include Precision Neurotechnologies (approximately £69 million), Massively Scalable Neurotechnologies (approximately £50 million), Safeguarded AI (approximately £59 million), Scaling Compute (approximately £50 million), and the climate cooling programme (approximately £57 million across a £50 million experimentation envelope and a separate £11 million scoping element).[^7][^14][^24]
The most significant AI-focused programme in ARIA's portfolio is Safeguarded AI, which sits inside an opportunity space ARIA calls "Mathematics for Safe AI." The programme was designed and originally led by David "davidad" Dalrymple, who joined ARIA as a Programme Director in September 2023, and who in October 2023 published a programme thesis outlining the approach. ARIA officially opened the programme for proposals in April 2024 with a committed budget of approximately £59 million over a roughly three-to-five-year period.[^6][^25]
The thesis of Safeguarded AI is that conventional approaches to AI safety, including reinforcement learning from human feedback, red-teaming and behavioural evaluation, do not provide the quantitative safety guarantees that society demands from safety-critical engineered systems such as civilian aviation or nuclear power. Safeguarded AI proposes to build a "gatekeeper": a verifier system that uses formal mathematical proofs against an explicit world model to certify, before deployment, that an autonomous AI system will not violate specified safety properties.[^6][^26]
Dalrymple drew the term "guaranteed safety" partly from earlier writings by AI researchers including Stuart Russell and Yoshua Bengio. In an influential 2024 position paper, Dalrymple, Bengio, Russell, Max Tegmark and other co-authors argued that future advanced AI systems should be constrained by mathematically verified contracts and that the path to credible safety lay in combining probabilistic world models with formal verification. Safeguarded AI is the practical research vehicle for that vision.[^26]
ARIA divides Safeguarded AI into three Technical Areas:
Yoshua Bengio, a Turing Award laureate and longstanding advocate of guaranteed-safety approaches, joined Safeguarded AI as Scientific Director in August 2024.[^6][^27]
By 2025, Safeguarded AI had funded a range of TA1 and TA2 teams across British and international academic institutions, startups and non-profits, including teams at the University of Oxford working on formally verifiable AI for business applications, and teams associated with research organisations such as mila institute in Montreal, far ai and others. Programme materials emphasise that Safeguarded AI is interested in tool-building, not in deploying frontier general-purpose chatbots.[^28]
In November 2025, ARIA announced a strategic pivot for Safeguarded AI: TA1 was expanded in scope and TA3 was refocused. In February 2026, a further pivot was announced in which TA3 was narrowed to cybersecurity and microelectronics applications, and the previously planned TA3 Phase 2 solicitation for cyber-physical applications was cancelled. Dalrymple transitioned from Programme Director to Technical Advisor and Nora Ammann, originally one of ARIA's Programme Directors for AI safety and assurance, was confirmed as Programme Director of Safeguarded AI. The £59 million programme envelope was retained.[^6]
Safeguarded AI is one of around fifteen opportunity spaces that ARIA had publicly identified by early 2026. Opportunity spaces are framing categories: each may contain one or several discrete programmes run by individual Programme Directors. The following list, drawn from ARIA's website, summarises the main spaces and their associated programmes.[^29]
Of the listed programmes, Safeguarded AI, Scaling Compute, Hypersensory Intelligence and Robot Dexterity have the most direct AI content. Other programmes, such as Forecasting Tipping Points and Precision Neurotechnologies, rely on AI methods as components but are not classified as AI programmes.[^7][^29]
ARIA is more closely modelled on DARPA than any other British research body. The similarities are deliberate: empowered Programme Directors with hire-and-fire authority over their portfolios, fixed-term tenures, an explicit "tolerance to failure," the Heilmeier Catechism as an internal discipline, the absence of academic peer review, the structuring of programmes into Technical Areas, and the rhetoric of "high-risk, high-reward" research.[^16][^17]
There are however four substantive differences. First, DARPA is part of the United States Department of Defense and exists primarily to deliver national-security capability; ARIA is not attached to any procurement customer. Critics, including the Wonkhe and War on the Rocks commentariat, have argued that this is the model's biggest weakness, since the disciplining effect of an end customer (such as the United States Air Force or Navy) is absent. Second, ARIA is much smaller than DARPA, both in absolute budget and as a share of national research spending. Third, ARIA includes climate, biology and neurotechnology programmes that DARPA does not run in comparable form. Fourth, ARIA's statutory FOI exemption is a stronger transparency carve-out than the equivalent exemptions DARPA enjoys under United States law, where DARPA contracts often eventually become subject to disclosure through Department of Defense reporting.[^10][^15]
ARIA staff have publicly argued that the appropriate comparators are not DARPA in isolation but the broader family of "ARPA-style" agencies, including ARPA-E, ARPA-H, the Defense Innovation Unit and Israel's chief scientist office.[^16]
ARIA is part of, but distinct from, a constellation of United Kingdom government and government-adjacent bodies concerned with AI. The most important of the others is the UK AI Security Institute (uk aisi), known as AISI. AISI was created in November 2023 immediately after the ai safety summit at Bletchley Park as the AI Safety Institute, then renamed the AI Security Institute in February 2025. AISI sits inside DSIT, not at arm's length like ARIA. AISI conducts pre-deployment safety testing and red-teaming of frontier AI systems through agreements with leading AI developers including OpenAI, Anthropic and Google DeepMind, and runs a grants programme of more than £15 million to fund external AI safety research. AISI is part of an international network of ai safety institute bodies that includes the United States us aisi and equivalents in Singapore, Japan and elsewhere.[^34][^35]
The functional differences between ARIA and AISI can be summarised as follows. AISI is principally an evaluator of frontier AI systems and an advisor to government on AI risk; its in-house technical work tests existing systems against capability and misuse benchmarks. ARIA is principally a funder of speculative research and tools, with Safeguarded AI being the most prominent AI line of work. AISI's outputs are evaluations, reports and a small grant programme; ARIA's outputs are programmes, funded teams and (the agency hopes) deployable technologies. There is operational coordination between the two bodies, since both report to DSIT and Matt Clifford, ARIA's Chair, was the Prime Minister's Adviser on AI during 2025.[^21][^34]
Other adjacent institutions include UK Research and Innovation (UKRI), which remains the principal funder of mainstream British research; the Alan Turing Institute, the national institute for data science and AI, which is a partner organisation to several ARIA programmes; the Department for Science, Innovation and Technology itself, which sponsors ARIA and contains AISI; and several non-governmental AI safety bodies such as the Centre for the Governance of AI in Oxford. ARIA's place in this landscape is sui generis: it is the only United Kingdom body whose statutory remit is high-risk research with explicit tolerance to failure.[^11][^34]
ARIA was politically contested from the moment of its proposal. Concerns expressed during the Bill's passage included:
A non-exhaustive timeline of major announcements:
By early 2026, ARIA had committed several hundred million pounds across its opportunity spaces, with the bulk of programmatic spending still ahead under the renewed 2025 to 2029 envelope. Its long-term significance for British research and for global AI safety research will not be assessable until well into the 2030s.