UK AI for Science Landscape: Methodology

What is mapped

The dataset maps named organisations, units, programmes, facilities, and open tools and datasets with a real UK base that do, or directly support, AI for science. AI for science means here AI used to do scientific research, and the data, compute, tools, training, and funding that support it. Entries include companies and university spinouts; university and institute research units; focused and independent research labs; research support units; doctoral, fellowship, and training programmes; consortia and networks; government programmes and agencies; private and philanthropic funders; research and compute facilities; and open tools, models, and data resources. The fields served are recorded in the domain column: Biology, Chemistry, Clinical, Materials science, Engineering, Physics, Earth and climate, Mathematics, and Cross-domain.

One row per mapped unit

Each row describes one mapped unit. A university or other large body is represented by its named AI unit, strand, or programme, not as a whole. A row called “X AI work” is the AI strand of X, and its facts (founding year, team size, funding) are the strand's, not the parent's. A national facility that is itself infrastructure for AI for science is mapped as the facility. A university AI institute is mapped once, with its AI-for-science strands as notable outputs rather than entries; a company's own named programmes are treated in the same way, so that no work is counted twice. When two rows are found to describe one unit they are merged, and the other name is kept as an alias.

A census of extant entities

The dataset records entities that exist now. Closed and dormant organisations are removed; a company bought by another and still operating stays, marked Acquired. The founding year places each current entity in time. Counts by founding year are therefore counts of today's entities, not a record of how many were formed: entities founded and since closed are absent (§9.3 and Appendix C).

UK base and placement

Any credible evidence of UK activity counts as a UK base; a Companies House registration is not required. hq_country records the country of headquarters. A UK lab or strand of an overseas company takes the parent's country; a UK company owned by a foreign group but run from its UK headquarters stays UK. An entry headquartered abroad holds, in hq_city, the city of its UK base (London where that city is not known) and is drawn there, while the stories still count it as headquartered overseas. Nothing is drawn outside the UK. At this snapshot 30 entries are headquartered outside the UK, 18 of them in the story cohort.

What the dataset is not

It is a guide to the ecosystem's organisations, not a census of the research itself. The wider shift towards AI-enabled work across research groups and university departments is substantial, but it is not organised as units that can be mapped, so a single group, grant, or project is not an entry (§4.2). It is not a register of funding flows: the funding columns describe each entry's own backers and money, not the grants it makes. It is not a ranking: relevance measures contribution to AI for science, not quality. Work next to AI for science, such as AI research in general, AI safety, or businesses adopting AI, is marked lower in relevance or left out (§4.5). Science Works, which publishes the map, is itself an entry and is placed by the same definitions.

The first taxonomy, May 2026

The first version, the UK AI4S Taxonomy V0.1, was assembled on 19 and 20 May 2026. It held 404 organisations in 27 fields, with a review-confidence annotation added in a later confidence pass. Its rows came from deep-research calls run with AI research assistants (ChatGPT and Gemini). The calls were partitioned so that no part of the landscape depended on a single search: by funding source (UK-GOV, UK-PHIL, and UK-VC), by scientific domain (DOM-BIO and DOM-CHEM-MAT), and by institutional anchor (ANCHOR-SPIN), with two general passes (CHATGPT-A-PUBLIC and CHATGPT-B-DOMAINS) and a seed-list pass that verified known organisations and expanded from each (SEED-EXPAND and SEED-EXPAND-NEW). Every call returned the same schema and recorded the call that found the row (call_id), the collection date, its sources, and a confidence level. Calls overlapped by design, so that an organisation found by several calls was cross-confirmed; the outputs were merged on name and country of headquarters.

The gate used to generate the list was a single test: AI or machine learning is central to the organisation's scientific work, in the sense that it would not exist in roughly its current form without it. Who the AI work was for was recorded separately and was not used to gate. Each row carried a first-pass verdict, threshold_classification: Definitely UK (233 rows), Maybe UK (18), International benchmark (84), or Out — captured for audit (69). Out-of-scope rows were kept so that later passes would not find them again. Relevance (High 249, Medium 94, Low 61) was seeded by an AI pass and adjusted by hand. Of today's 587 entries, 243 entered through these passes (Table 1).

The first map, June 2026

The first map was built from the taxonomy on 22 June 2026 by a single build script that reads the dataset and writes self-contained pages, with every statistic computed at build time. In late June the curators added funders, policy bodies, and regional lists for the north of England and Scotland. On 25 June the inclusion gate (Include?) and the AI role column were introduced; the data stories were restricted to entries rated High from the start.

The corrected workbook, September 2026

In September the working sheet was copied into a corrected workbook, with every correction logged in a changelog tab. The first corrections filled lead investors for 82 venture-backed organisations, removed seven duplicate rows, and refreshed stale funding records. A build guard was added: rows whose normalised names collide are merged, programme budgets are kept out of funding comparisons, and a QA report is written on every build.

A full refresh (11 September), a funding and openness review (18 September), and a data-completeness audit and a pilot of the rule for university units (22 September) followed. The funding review recorded how firm each main funder type is (Verified, Provisional, Estimated, or Unresolved) and separated an unknown funder from a mixed one; the openness review separated Not applicable from Unknown. A research application on 23 September removed closed organisations, merged duplicate programmes, consolidated EMBL-EBI's sub-resources into the institute, and replaced the threshold columns with inclusion_rationale. It left 446 rows in 52 columns: 324 entries on the map and 188 in the story cohort. Proposed changes to the value chain, AI role, and AI methods were held for review rather than applied.

One definitions document and an evidence base, October 2026

From 2 October 2026 the work moved to a column definitions document. Since 3 October it has been the only source of what each column means and which values it may take, with a dated change log of every decision. On 3 October the 135 rows not on the map were moved out of the dataset into an internal evidence base, with eleven working and legacy columns; from then the dataset holds only current data for the entities on the map, and a row taken off the map goes to the evidence base with its reason. Since then 20 further rows have left the map, the latest on 22 October 2026. Between 2 and 22 October the columns were audited against the definitions (§6, Appendix B). The evidence base keeps every version of every cell and every piece of research, proposal, review, and decision, in a dossier for each organisation considered: 1,410 organisations by 19 October 2026, on the map or not.

The data chain

1. The live working sheet, “Newest AI4S”. It is read-only and is never edited; it still contains errors corrected downstream.
2. The corrected workbook and its changelog tab, the dataset of record in September 2026.
3. The repository dataset, data/uk-ai4s-taxonomy-v0.1.csv. Until October it was an export of the workbook's first sheet. Since October, decisions have been applied to it directly, each change logged in the evidence base and in version control; the corrected workbook is to be brought into line from lists of the changes.
4. The build. build_map.py reads the dataset and two page templates and writes processed.json, aggregates.json, qa_report.json, the map page, and the landing page with the stories. A check script (22 automated checks) runs after each build (§6.5).

The site reads only the build outputs embedded in its pages. The raw dataset, the internal notes, and the evidence base are not published.

Research passes, May 2026

The May passes are described in §2.1. Partitioning the search along independent axes gives each organisation several chances to be found; an organisation found by one call only is either niche or easily missed. The method note of June 2026 recorded the known limit: generic search misses organisations that do not surface easily, such as smaller spinouts, recent programmes, and groups within larger institutes. The seed list and the curators' additions were the correction.

Additions and sweeps, June to September 2026

The curators added funders, policy bodies, and regional lists in June. In September a sweep of recently announced organisations, with particular attention to new independent labs, was run with an AI assistant; its candidates, and those of a discovery sweep on 11 September, were researched, appended for review, and audited on 18 and 21 September. A sweep of ARIA's funded programmes followed on 24 September.

Domain sweeps, 2 October 2026

Three sweeps looked for organisations missing from the dataset in physics and Earth and climate science (84 candidates found, 70 added), in engineering and materials science (134 found, 78 recommended, 59 added), and in chemistry and biology (141 found, 91 recommended, 83 added). Each sweep used several searchers, each covering a different part of its field, and a list of every row ever entered, so that nothing already considered was proposed again. Relevance was judged against a yardstick of the curators' earlier placements. Each High candidate was researched into a full row and checked by a second reviewer briefed to refute it; Medium rows were researched one by one and reviewed in batches; Low rows were filled in batches without a second review, with funding detail and team size left blank where not readily found. Every row was checked by script against the controlled vocabularies. The engineering sweep applied a boundary: AI that serves engineering research, design, and simulation is in scope; AI for running operations (process control, maintenance, inspection, autonomous products) is Low or out.

The May passes had leaned towards the life sciences and towards companies. The October sweeps were aimed at the fields that had been under-mapped, so search effort has not been equal across fields (§9.1).

Targeted sweeps, 19 to 22 October 2026

• AI for mathematics. New UK entries were researched and Mathematics was added as a field (19 and 20 October).
• ARIA. ARIA's programmes were re-checked against the relevance definition of 17 October (20 October), and four sweeps of its programmes, opportunity seeds, and partners listed about 170 award holders, from which candidate companies were researched (21 October).
• University AI institutes. A sweep of 54 universities (21 October) was followed the same day by a second, substantive reading of 42 further units, judged by what each institute does rather than by counting its programmes (§4.2).
• Focused and independent research labs. Fifty-two entries were checked against the five tests for the type (§5.2).
• Google DeepMind's science teams. Each team's existence and UK base were checked from its papers' affiliations, and each was tested for whether it is an entity of its own (22 October). Candidates await a decision.
• Fundamental AI and AI hardware. Evidence was gathered on the entries rated Fundamental AI and on organisations that may be missing (22 October). The boundary is open (§4.5).

Public submissions

The landing page has forms for suggesting an addition or a correction. Submissions go to a private response sheet and are reviewed by hand against the organisation's own pages and the source given. Accepted changes are applied with their sources, and the site is rebuilt. Nothing submitted enters the dataset automatically.

The review step

Every discovered row enters through a review step. A candidate becomes a card on a private review tool, carrying its proposed row, its evidence, and the reasoning; the curators accept, amend, or reject it; only accepted rows enter the dataset. Candidates not added are kept, with the reason, in the evidence base.

The inclusion gate

Only rows with Include? = yes appear anywhere. In scope is an organisation with a real UK base that does, or directly serves, AI for science. Named, lasting labs, centres, facilities, programmes, and widely used open resources qualify. Left out are industry-adoption programmes, funders with only a minor or historical AI line, and closed or dormant organisations. Each row's inclusion_rationale says why it is in at its relevance, naming a comparable entry where it can.

University units and institutes

The map represents organisations and lasting research infrastructure, not every group that uses machine learning; requiring every activity to appear would make coverage depend on arbitrary lab boundaries and favour universities with more visible websites. University research enters at the most coherent official level that passes four tests, set on 22 September 2026:

1. Entity: an official named unit with an identifiable institutional owner and a public presence.
2. Durability: a standing centre, institute, or facility, a multi-year programme, or an operational unit with continuing staff and activity.
3. AI for science: AI is central to its research in the natural, medical, environmental, or engineering sciences, or it materially supplies data, compute, or experimental infrastructure for that work.
4. Distinctness: a named lab, centre, facility, or other distinct unit, not an invented aggregate of unrelated activity.

A whole university or department goes in only when AI for science is its explicit focus. Rows such as “University X AI work” that would aggregate unrelated projects are not created. A rule for university AI institutes was added on 21 October 2026: an institute goes in when AI for science is a substantial part of what it actually does, judged from its own pages (its research, labs, projects, programmes, outputs, and people), not by counting programmes. Its AI-for-science strands then become its notable outputs rather than entries of their own. A convening umbrella of members with no research of its own is typed Research support unit.

Strands, facilities, and open resources

A national science body, company, or institute is normally represented by its named AI strand, with that strand's own founding year, team size, and funding where known. A facility that is itself infrastructure for AI for science is mapped as the facility, with its AI work as the notable output. Open models, tools, and data resources that researchers use are entries in their own right.

Exclusions

Rows leave the map, or never enter it, for these reasons: a duplicate of another row; no current UK base; closed, absorbed, or dormant; no AI-for-science component on the organisation's own pages; below the gate on relevance; or folded into a parent unit that is mapped instead. A folded row's contribution modes, methods, sources, and name pass to the receiving row. Every row taken off the map is kept in the evidence base with its reason (§2.4).

Relevance

Relevance measures how much an organisation matters to AI for science. Two things count: the scale of its contribution to AI for science, and the proportion of its work oriented that way (definition of 17 October 2026). Either can carry an organisation to High: a large contribution makes it High even where AI for science is a small share of what it does, and a unit whose work is mainly AI for science is High even when it is small. An organisation need not be primarily AI-oriented to be High, and it need not use AI itself: data infrastructure that is critical for AI for science is High without it. The mapped unit's own work is judged, not its size or fame; whether it is running yet does not change its relevance. The examples under each level illustrate a judgement and are not thresholds; no row is placed for consistency with another, and each is judged on its own pages.

High includes companies whose core is AI-driven discovery or design; named AI strands of national science bodies; compute, data, and open models that researchers use for AI now; companies whose models or tools scientists use, open or not; laboratory automation and self-driving labs; funders, fellowships, and doctoral programmes whose purpose is AI for science, and funders whose AI-for-science funding is large in scale; and think tanks and policy bodies that contribute to AI for science. Medium includes research into future AI technology itself; doctoral programmes broader than AI for science; entries with a thin UK base; companies where AI is one tool in a conventional business; and institutes, centres, and funders whose AI-for-science contribution is modest and one part of broader work. Low includes taught master's courses; commercial analytics sold into markets outside science; AI for running operations rather than for research; data-science centres that mainly convene; AI-safety organisations; and organisations whose contribution is small and a minor part of their work.

At this snapshot 282 entries are High, 188 Medium, and 117 Low. Open: taught master's courses are still placed at Low by type; the boundary between fundamental AI and AI for science is under review, and the entries it affects may move.

Populations

The map shows every entry; its relevance filter starts at High, the level the stories use, and can be widened. The stories use the story cohort, except where a story says otherwise (Table 2).

Identity and placement

name is the mapped unit's name; aliases holds other names and acronyms, including a merged duplicate's name; website is the mapped unit's own page. status is Active or Acquired (bought, still operating). uk_presence, hq_country, and hq_city follow §1.4; hq_city holds one place name from the map's place list. founded is the year the mapped unit started, not its parent; for the AI strand of an older body, the year its AI programme began, where evidenced. team_size_band counts the staff of the mapped unit, not the parent; doctoral centres count staff, not students. parent_institution records the university or research institute an entry came from or belongs to: for a spinout, the institution it was spun out of; for a lab, centre, strand, doctoral centre, course, fellowship, or single-lead hub, the institution it is part of or that hosts or leads it; for a programme run by several institutions, the lead institution, or both where there are genuinely two leads. Companies, funders, and independent organisations are blank. The map draws only recorded links; nothing is inferred from names or cities.

Organisation type

org_type takes one of fifteen values and is judged by what the organisation does rather than its legal form; where in doubt, by what it does. The main rules:

• University spinout needs evidence that the company was spun out of a university. A spinout from a research institute that is not a university is a company.
• Owner rule. A wholly owned subsidiary, in-house lab, product line, or acquired spinout is typed by its owner, Company-public if the owner is listed.
• Fellowship programme covers doctoral centres, taught degrees, and fellowships.
• Focused and independent research lab (added 21 October 2026) must pass five tests: it does research itself; it is independent of any university, national lab, government body, or company; it is anchored by public or philanthropic money; it was not built as a conventional business; and it has no shareholders or investors expecting a financial return, its own purpose being research. Legal form does not decide it.
• Academic institute or centre is any research institute that conducts science, whatever its legal form.
• Research support unit is a unit of a university, institute, or national lab whose main job is helping other researchers use AI rather than conducting science itself.
• Research facility is physical infrastructure whose main job is to be used by other researchers, including research and technology organisations that lead with laboratories or test facilities offered to industry; Compute facility is computing infrastructure for researchers. A unit that does its own science around its labs stays an Academic institute or centre.
• Private or philanthropic funder is a body whose primary purpose is giving money for research or companies and that does not do the science itself. Government funders stay Government programme or agency; learned societies are not funders unless funding is their primary purpose; a body that does the work keeps its own type even if it also gives money.
• Open tool or data resource is an open tool, model, or data resource; a data service is typed this way whoever runs it. Other is reserved for entries awaiting a type decision (none at this snapshot).

The stories group the fifteen types into five families (Table 3).

AI role

ai_role records what people doing AI for science in the UK can use from the organisation today, with at least a reasonable degree of ease, judged by what predominates in its work. An Integrator provides something the field can use: infrastructure, compute, and data; training, funding, strategy, and governance; publicly funded research; open models and tools; or paid products and services that outside scientists can buy or license. A company still building a product for scientists is an Integrator. A User applies AI mainly to its own work, as an AI drug-discovery company developing its own pipeline does, even when it partners with others. Fundamental AI is research into AI itself, where that predominates. Each row holds one role, except OpenAI and Anthropic, which hold both Integrator and Fundamental AI; where one role must be picked, such as a dot's colour, Integrator is used.

Contribution modes

value_chain_position lists every contribution mode the unit substantially does, or is genuinely in train to do (funded and scheduled, or being built): usually one to three, always in the order of Table 4. Aspirations, incidental work, and what only a parent or customer does are not counted. Because the order is fixed, the first-listed mode is not the main one.

Fields and AI methods

domain records the field of science served, not where the method came from. Physics includes astronomy and fusion, particle, and quantum physics. Earth and climate covers the Earth system, weather, climate, the ocean, geoscience, and environmental science. Clinical applies only where the work serves clinical or patient research; drug discovery up to the clinic is Biology or Chemistry. Engineering is the field for robotics. Mathematics (added 19 October 2026) is the field served when the work is for mathematicians: formal proof and theorem proving, proof assistants and their libraries, conjecture generation and mathematical discovery, and their funding and training. An entry can list several fields; no main field is defined.

ai_work_type records the methods the unit itself demonstrably uses or teaches: Predictive ML, Generative model, Simulation + ML (emulators, surrogates, hybrid physics-ML models, and digital twins), Foundation model, Robotics + ML, NLP, Agentic, Computer vision, and Other. N/A is for funders, compute, and generic training; every training programme (doctoral training centres, taught master's courses, training schools, and fellowship schemes) is N/A, because such programmes train people rather than do AI work themselves. The cell is left blank where no method could be evidenced (7 entries).

Openness

openness_posture records how freely the field can get hold of the organisation's own main outputs (research and publications, data, models, software, and tools), judged as a whole and only on what the mapped unit itself produces. Policy research counts as an output. Charging for something does not make it proprietary, and the absence of a public download is not proof of restriction. Papers about a closed platform do not make it open, and a funder's open-access policy for its grantees is not the funder's own openness. Outputs released only for non-commercial or academic use count as open; a drug-discovery pipeline counts as a proprietary output. The reason for each value is recorded and is internal.

“Fully” means nothing restricted at all; any restricted main output makes the value “Mostly”. Mixed is not a fallback for uncertainty; uncertainty is Unknown. Research institutes registered as charities are judged on their outputs. Open: whether the Not applicable ruling for government departments and large charities reaches a department's own research directorate or a charity's own research team.

Funding type and investors

funding_source_primary records the type of funder that gives the unit the most money. Any reasonable margin is enough; there is no fixed percentage (an 80% threshold was replaced on 2 October 2026). Mixed applies only when no type leads, for example two types at about 40% each; Unknown applies when there is no information about who funds the unit. Equity investment from anyone is VC, including corporate venture arms and state investment funds. A funder paying from its own money takes the kind of money it is: a government's is Government, an endowment's or a person's wealth is Philanthropy, and a company's profits are Corporate or self-funded. A company owned by another company, or a company's in-house lab, is Corporate or self-funded with the owner as lead, as is an organisation that lives mainly on its own trading income. Further rules cover companies backed by UKRI grants whose private funding is undisclosed, and entries funded by ARIA.

lead_investors names the primary funder; for a venture-backed company, the lead of each round where the lead is evidenced. A lead needs evidence of shares or structure and is never taken from the order of a list. other_investors lists every other identifiable funder, and funding_sources every funder type present, main type first. funding_model_basis records how firm the main type is: Verified (a complete quantified split; two entries), Provisional (main type supported by sources; 514), Estimated (Mixed, or a best estimate; 53), or Unresolved (Unknown; 18). funding_model_note is a short public summary of how the unit is funded, written from internal notes and required to agree with the funding cells. funding_stage records where the unit is in its funding life, from separate lists for companies and other bodies.

Funding amounts

funding_total_gbp_m records all the funding the unit has received from every source, including money from its own university or institution: for companies and young non-profits, everything received to date; for established bodies with a running budget, the latest year's income; for programmes, centres, consortia, and funders, the current award or settlement. Money granted counts, in full once allocated; pledges, earned income (trading, fees, and investment returns), and in-kind support, including compute time, do not. Research grants and contracts count; fees for services do not. The total is split into private, philanthropic, and public money, state-backed equity counting as public at the amount the state invested, and funding_total_complete marks whether it is the unit's whole funding for its period or a minimum. Further columns record what the figure is, what it covers, and when; funding_date is the date of the most recent disclosed funding event. Currencies convert at fixed rates: US dollars at 0.78 GBP, euros at 0.85, Canadian dollars at 0.56, and Swiss francs at 0.91.

At this snapshot 422 entries have a recorded total; 135 totals are complete, and 452 entries have a minimum or no total. No amount is shown for any entry on the map, and the private, philanthropic, and public split is not shown for any entry; the stories use amounts only as described in §8.4. Open: a few counting questions, among them whether funds only partly backed by government count as private money.

Descriptions and sources

Two descriptions are shown on the map. Primary activity (New primary_product) follows one house style, “<kind of unit> that <what it does for AI for science, and for whom>”; Notable output (New notable_output) follows “<named output> (<year>): <what it is>”. Both were rewritten from each organisation's own pages between 3 and 6 October 2026, and rows that rested on secondary pages were re-checked on 19 October; the older primary_product and notable_output are kept and not shown. source_urls holds two to five pages actually used, primary sources first, and is shown on the map. source_confidence (High or Medium) is blank on 221 rows that predate it. collected_date records when a row was entered or last refreshed, and call_id the search or batch it came from.

Retired columns

Seven columns have been removed, every value kept in the evidence base. The main ones are threshold_classification and threshold_rationale (23 September 2026, replaced by Include?, Relevance, and inclusion_rationale); success_framework_tag (2 October: it had no written definition and was applied inconsistently); and contribution_mode (19 October: it overlapped the contribution modes and was never audited). Eleven working and legacy columns, among them review flags, proposed values, and the funding shares of the retired 80% rule, moved to the evidence base on 3 October.

Audit design

From 2 October 2026 audits followed one design.

1. A brief gave the readers the column definition verbatim, as it stood that day.
2. First readers worked in batches, reading each row from the organisation's own current pages and, where relevant, from Companies House filings, funder announcements, and UKRI's Gateway to Research. Each proposal carried its evidence, sources, a confidence level, and a close-call flag.
3. A second reader tried to refute every proposed change from the organisation's own pages.
4. Larger audits were piloted on known cases first. An audit whose method failed was withdrawn and rerun: the first relevance audit of 3 October was withdrawn the same day and rerun from fresh page reads.
5. Results went to the review tool as cards: changes both readers confirmed, grouped by reasoning so that a group could be approved together, and close calls and rule questions separately.
6. The curators decided every card. Readers applied nothing. Decisions were applied to the dataset with a ledger of every changed cell.

Format-only fixes (capitalisation, spacing, the form of a link) were applied directly. Readers were AI research assistants; the curators' decisions are the record (§9.4).

Rules and decisions

When a card raised a question the definitions did not answer, the curators settled it, and the rule went into the definitions document with its date; the change log records each such decision. A decision on one row is not turned into a general rule without the curators' agreement. Worked examples in the definitions illustrate judgements and are not counts or thresholds.

Coverage of the audits

In all, 52 dated audits and reviews ran between 2 and 23 October 2026 (Appendix B). By 19 October every column that the map or the stories use had been checked at least once against its current definition (inclusion_rationale only on newer rows), and several were re-audited after their definitions changed: relevance (3 October, then a re-reading on 18 October under the definition of 17 October), contribution modes (2, 17, and 21 October), organisation type (19 and 21 October), and openness (2 and 19 October). Whole-dataset checks covered duplicates (2 October), British spelling (3 October), and allowed values, links, and websites (19 October); every source link was checked by script, and each flagged row was read by two readers. By 19 October the curators had decided more than 2,800 cards.

Corrections keep their history

A corrected public note keeps its history in the cell: the new note first, marked current with its date, then why it changed, then the old note, marked superseded. The site shows only the current note. A row taken off the map is moved, whole, to the evidence base with its reason. Every version of the dataset is kept in version control, and the evidence base holds every value each cell has held, with the research, proposals, reviews, and decisions behind it.

Build checks

Every build writes a QA report: rows merged as duplicates (none at this snapshot), programmes and funders whose amounts are kept out of funding comparisons, cities without coordinates, entries not drawn, and the composition of the story cohort. A check script then runs 22 automated checks. Among them: every main funder type is an allowed value; the openness denominators reconcile with the chart; the map's default filter matches the stories; the dataset holds only included rows and current columns; every dot is in the UK; programme budgets are never counted as money raised; medians use both middle values of an even count; methods are counted across every field an entry lists; the geography and team-size denominators reconcile; and the headline claims of stories 01, 02, and 08 hold on the current data. The build itself stops if a recorded parent institution has no coordinates, or if the file behind story 04's age-matched comparison lacks an older company it needs.

Known limits of the review

Two readers looked at every proposed change, but only one at every value kept. In the identity audit of 19 October, 194 of 519 organisation types were kept at medium or low confidence and not read again. There has been no systematic cross-row consistency check, and catch-all values (Other, Cross-domain, Unknown, and Mixed) passed unless a reader objected. A random-sample audit with a reported error rate has not been run. inclusion_rationale has not been audited for older rows, by the curators' decision of 19 October. Some sites were unreachable or blocked automated reading; such rows were marked unverifiable rather than changed. Readers worked within fixed search budgets, and some runs exhausted them; the limits each reader reported are kept with its output. Values a reader recalled rather than read were flagged and checked separately.

Reading the dataset

The build keeps rows with Include? = yes. It merges rows whose normalised names collide (lower case, punctuation removed, trailing legal suffixes such as plc and Ltd dropped), keeping the row with more filled cells, and logs every merge. Relevance is read regardless of case. Multi-valued cells are split on commas; N/A and Unknown are read as no value. The story cohort is every included row rated High with a UK base: hq_country = UK or uk_presence = Yes.

Multi-valued columns

Contribution modes, fields, AI methods, and funder types hold several values per entry. In the story cohort, 230 entries list more than one contribution mode, 144 more than one AI method, 102 more than one field, and 102 more than one funder type. The stories count an entry under every value it holds, or highlight one value at a time and light every entry that lists it. Shares are shares of entries, not of entry–value pairs, so rows can sum to more than 100%. The first-listed value is not a ranking. Two views still colour each entry by the first field it lists: the by-field view of story 03's flow chart and the whole-UK view of story 07.

Groups used by the stories

Main funder types are grouped as in Table 6: Government UK and Government other combine as Government, and a cell holding more than one value or an unlisted value reads as Unknown. Openness values are grouped as in Table 5: Unknown stays in the denominators and Not applicable is left out. Organisation types are grouped into families as in Table 3. Two places count focused and independent research labs within research bodies, as shorthand: the sentence above story 01's landscape chart and the bars of its founding chart. Everywhere else they are kept apart.

Funders, programmes, and money raised

Comparisons of money raised use funding_total_gbp_m and leave out three kinds of entry, so that the same money is not counted twice and budgets are not read as money raised: government programmes and agencies and research and compute facilities, whose amounts are budgets (53 with a recorded amount); funders, which the build identifies as entries whose first-listed contribution mode is Funding, that is, entries whose modes are only Funding, Strategy, or Governance (38 entries, 7 of them with a recorded amount); and entries marked funding_amount_applicable = no (80). Only positive amounts count. An incomplete total is a minimum and is shown as “at least”. Medians use both middle values of an even count. Story 03's chart of where each kind of money goes also leaves out funders whose money is their own, that is, funders with no recorded backer other than themselves (BIST, Wellcome Trust, and XTX Markets at this snapshot): they are the source of a kind of money rather than a destination for it.

Geography and clusters

Each entry is drawn at the coordinates of its hq_city; entries in the same place are spread by a fixed spiral of about two kilometres, so that each is visible. An entry with no known UK city is not drawn (one at this snapshot). For the stories, places are grouped into clusters: Hinxton (the Wellcome Genome Campus) with Cambridge, and Harwell, Abingdon, and Didcot with Oxford; every other place is its own cluster. The Golden Triangle is the London, Oxford, and Cambridge clusters. Geography comparisons use UK-headquartered entries in two groups, the Golden Triangle and the rest of the UK, and set entries headquartered overseas apart. Swindon, the registered address of UKRI and several of its councils, is set aside in story 07 and counted with the rest of the UK in story 05.

Founding years and windows

Founding-year charts place entries founded in 2010 or earlier in one bar and mark the latest year as a part year. Story 01 compares two five-year windows, 2015–2019 and 2020–2024; story 02 compares three three-year windows, 2017–2019, 2020–2022, and 2023–2025. Entries without a known founding year are left out of founding-year charts and counted everywhere else.

AI methods

Methods are counted across every field an entry lists. Shares of a method within a field are shares of the field's entries with an identified method; N/A and blank cells are left out of that denominator. In story 08, Other counts as an identified method. The legacy spelling “Computer Vision” is read as Computer vision.

The map

The map shows every entry, with a relevance filter that starts at High. Dots can be coloured by a chosen column and sized by team size or by recorded funding: a dot's radius grows with the square root of the recorded amount, and programmes, facilities, and funders are not sized. The page carries each dot's size, not its amount. An entry's panel shows its Primary activity and Notable output, and its funding (main funder type, backers, lead and other investors, and the public funding note) under a notice that the funding information is provisional. Institutions are grouped from recorded parent links only.

Story 01: founding years and the landscape today

“Of today’s organisations and programmes, more than twice as many were founded in the five years from 2020 compared with the five years before.”

Population. The story cohort (282). The founding chart counts the 274 entries with a known founding year; eight undated entries are left out of it. The landscape chart shows every entry in the cohort, by family.

Measure. Entries by founding year. The headline compares the number of today's entries founded in 2020–2024 with the number founded in 2015–2019; the windows are equal, so the ratio of totals is the ratio of annual averages, and a build check fails if it drops below two. The bars can be coloured by family, by main funder group, or by one field at a time, an entry lighting if it lists the field. The main-funder text compares two three-year windows (2020–2022 and 2023–2025); its statement on government among companies' funders reads funding_sources, in which state-backed equity counts as venture investment (§5.7). The landscape chart can colour entries by main funder group, or highlight one contribution mode or one AI method at a time.

Caveats. The chart counts today's entities, not formation. A scoping sweep on 21 October 2026 found at least 15 High-relevance entities founded since 2011 that have since ended, 12 of them founded in 2015–2019 and nearly all fixed-term public programmes and facilities that ended as planned (Appendix C). Counting them would have taken the comparison, on the data of that day, from 53 against 124 to about 66 against 126; the headline is therefore stated as a count of today's organisations and programmes, not as a rate of formation. Counts for the most recent years are lower bounds, because new organisations take time to be found, and 2026 is a part year. Announcement, incorporation, and launch can fall in different years; the year recorded is the one the evidence gives for the unit's start, and a few are close calls. The family colours and the sentence above the landscape chart count independent labs within research bodies (§7.3).

Story 02: the newest wave

“The newest wave is heavy in fellowships, networks, and programmes, which barely featured before 2020.”

Population. The story cohort with a known founding year (274).

Measure. One bar per group of organisation types (Table 3), growing with founding year and stacked by main funder group. The headline and standfirst compare fellowship and training programmes, consortia and networks, government programmes and agencies, and research and compute facilities across the windows 2020–2022 and 2023–2025, and against 2017–2019; facilities are included because until 21 October 2026 they were typed as government programmes. Build checks fail if these comparisons, or the steady count of new companies the standfirst describes, stop holding on the current data. Every name and number in the findings panel is computed from the data.

Caveats. Survivorship bears hardest on this story: most of the High-relevance entities found to have ended were fixed-term programmes, networks, and hubs founded in 2015–2019 (Appendix C), so earlier programmes are undercounted by construction. Company counts in the latest window will rise as recent start-ups are found; the headline makes no claim about companies for that reason. A reorganised body can carry a recent founding year and so place existing activity among new entries. Open: whether “barely featured before 2020” should be restated in the present tense, as story 01's headline was.

Story 03: who funds what

“Venture capital builds models and tools; public and philanthropic money builds the shared foundations.”

Population. The flow chart counts story-cohort entries with a known founding year that list at least one of its rows (274). The heat map counts the cohort less the three funders whose money is their own (279 entries; §7.4). The philanthropy table counts High and Medium entries with a UK base (470).

Measure. The flow chart shows entries by contribution mode as founding years accumulate, coloured by main funder group or by field; its rows are the modes held by at least three entries, and Design and Governance have no row of their own. The heat map gives, for each of the three main funder groups (venture or investment, government, and philanthropy), the share of its entries working in each of six parts of the value chain: models and tools (Model building, Tools and platforms, and Design), data (the three data modes), compute and facilities, talent and training, funding, and strategy and governance. Groups with fewer than 20 entries show counts rather than shares. The philanthropy table profiles the three philanthropies that back the most entries (a philanthropy needs at least four to be considered), counting an entry when the philanthropy, a fund it runs or co-founds, or its founder's giving is named as lead or other investor and the row records philanthropic money; cells give each philanthropy's share of its entries, and a cell is marked when the philanthropy is at least 1.5 times as likely as other research funders to support that kind of entity, with at least three entries. The comparison pools what other research funders support: the 301 entries backed by public money or by a philanthropy other than the three profiled.

Caveats. The main funder is the single largest type of funder; most entries have other backers too. Entries work in several parts of the value chain, so a row of the heat map sums to more than 100%. A philanthropy's equity stakes, and investments by its venture arm, are investments rather than giving and are not counted. Named investors establish that a backer recurs, not ownership or control. In the flow chart's by-field view each entry is coloured by the first field it lists (§7.2).

Story 04: the newest companies

“A few new companies, several founded by people from DeepMind, have raised more in three years than most older ones in a decade.”

Population. UK-headquartered companies and university spinouts at High or Medium relevance with recorded funding: 149 of 176. This story counts Medium entries, as its scope note says, because companies one step removed from AI for science compete for the same money and people. Companies founded by former DeepMind employees come from a separately researched list of 65 companies (23 in the UK, 9 on the map), compiled on 20 October 2026.

Measure. Recorded funding (funding_total_gbp_m: all funding from every source, a minimum where the total is incomplete), on a logarithmic scale and against founding year: the share held by the largest companies, the companies founded since 2023 and those among them with at least £20m recorded, and teams of 25 or fewer staff. The comparison with older companies uses what each company founded in 2015–2018 (31 with recorded funding) had raised within three years of founding, worked out from its dated rounds (an upper bound where the dates do not separate); for the newest companies it uses the years since founding. The DeepMind list counts a company when at least one co-founder was employed at DeepMind or Google DeepMind before founding it; internships, scholarships, consultancy, and advisory roles do not count. Its amounts are disclosed equity and grants, not valuations, and the dataset's own amounts are used for companies on the map. Rounds reported as in progress are drawn apart, as predicted amounts rather than money raised.

Caveats. Totals are lower bounds where funding is incomplete, and amounts convert at fixed rates (§5.8), so a total close to a threshold can depend on the rate. The older cohort holds only companies that still exist; companies that failed, which are likely to have raised less, are absent. Most companies on the DeepMind list are not on the map; its research is kept with its sources in the evidence base. Headcounts come from team pages, company accounts, and other public sources and can lag; a small team at a new company may not yet have hired.

Story 05: geography

“Companies cluster in the Golden Triangle; government is a particularly important investor for the rest of the country.”

Population. The 264 UK-headquartered entries of the story cohort; the 18 headquartered overseas are set apart.

Measure. For each main funder group, the share of its UK-headquartered entries in the Golden Triangle and in the rest of the UK (§7.5). The standfirst gives the same split for all entries and for companies and spinouts, and the main funders of the entries outside the triangle and in Edinburgh, the largest cluster outside it. Mixed and Unknown funder groups are not drawn.

Caveats. Location is the city of headquarters, one per entry; organisations with several sites and national programmes are placed at their headquarters. National funders registered at UKRI's Swindon address (5 in the cohort) count as outside the triangle in this story, while story 07 sets Swindon aside. Oxfordshire sites other than Harwell, Abingdon, and Didcot, such as Culham, count as outside the triangle.

Story 06: openness

“Government and philanthropy fund the open layer of AI for science.”

Population. The story cohort (282). The 59 entries whose openness is Not applicable are left out of this output-based comparison, among them pure funders; the remaining 223 include 17 with unknown openness, which stay in the denominators.

Measure. For each main funder group, the share of its entries whose main outputs are Open, Proprietary, Mixed, or Unknown (Table 5). The main view shows venture or investment, government, and philanthropy; an expanded view adds university or institutional, corporate or self-funded, and Mixed. Entries whose main funder is Unknown count in the coverage figures but are not drawn.

How main funders are assigned. The main type is the type of funder that gives the organisation the most money; there is no fixed percentage. Mixed means that no type leads, for example two types at about 40% each. Venture or investment includes equity from anyone, including corporate venture arms and state investment funds. A company owned by another company, or an organisation living mainly on its own trading income with little outside funding, is corporate or self-funded. How firm each assignment is (verified, provisional, or estimated) is recorded; 24 cohort entries use a best estimate. Other backing is kept in each entry's profile.

Suppliers. Of the 119 companies and university spinouts in the cohort, 85 provide tools, platforms, data, or infrastructure that others use (they list Tools and platforms, Data infrastructure, Physical infrastructure, or Compute infrastructure); 70 of these are mostly or fully proprietary. Supplying a research tool does not by itself make the underlying work open.

Caveats. Openness is judged on the mapped unit's own main outputs as a whole: one open release does not make an organisation open, and a funder is not credited with its grantees' outputs. Unknown is an evidence gap, not a finding that an organisation is closed. The grouping uses the main funder only. Openness and funding were assessed in different audits; both are recorded judgements, and the reasons for openness values are internal.

Story 07: clusters

“Beyond the big three, each cluster has a character of its own.”

Population. High and Medium entries with a UK base, drawn at their UK city: 469 (282 High and 187 Medium). This story counts Medium entries so that smaller clusters have enough to describe; entries headquartered overseas count at their UK base. Profiles cover the 21 clusters with at least three entries (424 entries). Swindon (10 entries, mostly national funders at UKRI's address) is set aside, and 35 entries sit in places with fewer than three.

Measure. Each profile's counts by family, its three most common fields (an entry counts in every field it lists), its most common main funder group, and its entries holding Physical, Compute, or Data infrastructure are computed from the data. The sentence on each cluster's character is written by hand and carries no numbers. The field highlight lights every entry that lists the field; the robotics and agents highlights light entries listing those AI methods.

Caveats. Location is the headquarters or UK base, not every site. On the whole-UK view each dot is coloured by the first field its entry lists; no main field is defined (§7.2). Small clusters rest on a handful of entries. Entries with no known UK city are not drawn (one High or Medium entry at this snapshot).

Story 08: fields and methods

“Most of the organisations doing AI for science are in biology, and each science uses AI in its own way.”

Population. Story-cohort entries that do AI-for-science work themselves: those listing any of Data generation, Data curation, Data infrastructure, Model building, Tools and platforms, Design, Physical infrastructure, or Compute infrastructure (229 of 282). Entries that only fund, train, or set strategy are left out (53).

Measure. One row per field; each entry is a dot in the column of its founding year, in every field it lists, so the rows overlap. Undated entries (7) are not drawn but count in the shares. The steps give each field's share of the companies and spinouts in scope (119); the share of each field's entries holding Design; and the share of each field's entries with an identified AI method (206 in scope) that list Simulation + ML, or Agentic. Comparisons between entries working in one field and in several set Cross-domain apart. A build check fails if the headline's claim stops holding on the current data.

Caveats. Methods are those a unit demonstrably uses, identified from public descriptions; the use of agents in particular is likely to be undercounted. Fields record the science served, not where a method came from. Shares for small fields rest on few entries.

Coverage and discoverability

The dataset is a best effort, not a complete census. Organisations that publish little, are new, or sit inside larger bodies are harder to find. Search effort has not been equal across fields: the May passes leaned towards the life sciences and companies, and the October sweeps targeted the fields that had been under-mapped (§3.3). Field shares partly reflect this. A standing policy, set in September 2026, is that refresh sweeps give every field the same query budget.

Recent years are lower bounds

A newly founded organisation can take months to appear in the sources the searches use. Counts for the latest founding years will rise as the dataset is refreshed, and the latest year is a part year.

Survivorship

Because the dataset holds only entities that exist today, earlier founding years undercount what was there, and fixed-term programmes that have finished never appear. A scoping sweep on 21 October 2026 (142 web searches across six kinds of entity, a full pull of UKRI's Gateway to Research, and the 2014 and 2019 cohorts of doctoral training centres) found at least 15 verified High-relevance entities founded since 2011 that have since ended, 12 of them founded in 2015–2019 and almost all fixed-term public programmes and facilities that finished as planned; three were company closures (Appendix C). It also found 11 borderline and 11 Medium-relevance entities that have ended. The 2014 cohort of doctoral centres held no High-relevance centre that has ended. The count is a floor: archived pages were blocked for several searchers, one institute's site refused automated reading, and a name filter skipped four candidates that remain unverified. Open: whether to keep ended entities in the dataset, with a status and an end year, so that they count in the founding-year stories while staying off the map.

AI-assisted research

Discovery and audits were carried out with AI research assistants: deep-research modes of ChatGPT and Gemini in May 2026, and Claude agents from September 2026. Assistants proposed; they did not decide. Every proposal carried its sources; in the audits every proposed change was checked by a second reader; and every decision was taken by the curators. Assistants can misread pages, lean on secondary sources, or recall rather than read; the controls are primary sources first, refutation by a second reader, flags on recalled values, and the curators' review. Errors will remain, and corrections are welcome (§10.4).

Categorical boundaries

Relevance, organisation type, AI role, contribution modes, and openness are judgements against written definitions. Close calls were decided by the curators, and the definitions record the rules that came out of them, but the boundaries are not sharp: a different reader could place some entries differently. The boundary between fundamental AI and AI for science is under review. No public claim rests on a fine distinction between neighbouring categories.

Multi-valued columns

Fields, contribution modes, methods, and funder types are multi-valued, so counts by value overlap and do not sum to the number of entries. A single colour per entry cannot show several values; where a view does so, it uses the first-listed value, which is not a ranking (§7.2).

Funding

Funding amounts are incomplete: 452 of 587 entries have a minimum or no total. Amounts come from announcements, company filings and accounts, award databases, and funders' reports; a figure from an aggregator with no primary source never fills a row on its own. Main funder types are mostly provisional: supported by sources rather than by a complete quantified split, which two entries have. Currencies convert at fixed rates rather than the rate of the day. Public funding notes and investor names are shown on the map as provisional.

Openness

Openness is judged from what an organisation publishes about its outputs and their licences. Organisations that release little, or say little about access, are Unknown (17 in the cohort at this snapshot). Judging main outputs as a whole requires weighing open and restricted parts, and the line between Mostly and Mixed involves judgement.

Methods from public descriptions

AI methods are recorded only where a unit's own pages, papers, or releases show it uses them. Organisations that describe their methods loosely, or not at all, are under-recorded, and newer methods such as agents spread faster than descriptions are updated.

Snapshot dates

Each row records when it was entered or last refreshed. The figures on the site are computed at each build and move with every review. Differences between snapshots mostly reflect changes in definitions and coverage, not in the field: a comparison between snapshots should apply the same definitions to both before a change is read as a trend.

The definitions document

The column definitions document is the only source of what each column means and which values it may take. It is updated whenever a definition changes, and its change log records each decision with its date. Earlier records of how the definitions were reached are kept in the evidence base as history. The build's groupings for the stories follow the document (§7).

The evidence base

The evidence base holds everything researched, proposed, reviewed, or decided about every organisation considered, including those excluded from the map and those never added: every earlier version of the dataset, every research run with its brief and its readers' outputs, every card shown to the curators, every decision with its time, and the working files of each session. It is rebuilt after every audit and is not published.

Refresh cadence

The dataset is refreshed roughly every two months. Measuring the same cohort twice is the only comparison that shows a change in the field rather than in the method. A finding that moves between snapshots is tested by applying the earlier definitions to the new data before it is called a change.

Public corrections

Anyone can suggest an addition or a correction through the forms on the landing page (§3.5). Suggestions are reviewed against sources; accepted changes are applied, logged, and published at the next build.

Versions

Build outputs are committed with the dataset, and every version of both is kept in version control; earlier versions of the site can be restored. This document is rebuilt from the dataset and the build outputs, and its figures are filled in at build time.