The Bletchley Declaration and the AI Safety Summit

moment is the right kind and proposals.md filed it correctly. The house definition — a date on which something visibly happened in public, the standard eliza-1966 argued its way out of and dartmouth-1956, lighthill-1973, asilomar-1975, deep-blue-1997, chatgpt-2022 and llama-weights-2023 have used since — fits without strain. Two dates, one venue, world press, a signed text published the same day. There is no re-filing argument to make and I will not manufacture one.

But the id names the wrong artefact, and that is the first thing worth saying. asilomar-1975 noticed that this canon keeps filing events under the year they became famous rather than the year the load-bearing document is dated, and called it three for three. This entry breaks that pattern and replaces it with a different one. The year is right: the summit was 1–2 November 2023 and the declaration was published on 1 November 2023. What is misdirected is the noun. The Bletchley Declaration is, on the evidence of the two and a half years since, the least consequential thing the summit produced. It commits nobody to anything, contains no deadline, no threshold, no obligation and no mechanism, and nothing in the world is different because it was signed. The three outputs of that week that are still running in August 2026 — a national testing institute, a standing international scientific assessment, and a voluntary pre-deployment access arrangement with frontier developers — are all announcements made alongside the declaration rather than commitments made inside it. An entry faithful to what a reading actually needs has to be about the summit, and this one is. The id stays as filed, because the chain's convention is that ids are stable and canon-all.sh keys on them, but a reader who arrives expecting a file about a document should know they are getting a file about an event.

descends_from: [chatgpt-2022], and the descent is causal rather than intellectual. This is the first entry in canon/ whose ancestor is a market event rather than a paper or an idea, and the link is the strongest one available: there is no AI Safety Summit without November 2022. The chronology is checkable and tight. ChatGPT launched 30 November 2022. The UK announced a Foundation Model Taskforce with £100 million in April 2023, explicitly framed around large language models. Sunak announced that the UK would host the first global AI summit during his Washington visit in June 2023. The summit sat eleven months after the launch that made general-purpose models a thing heads of government had opinions about. The declaration's own subject — "highly capable general-purpose AI models, including foundation models" — is a description of the thing chatgpt-2022 put in front of the public. Descent by cause is a looser standard than the one logic-theorist-1956 set by reading a 1954 reference list, but it is the standard dartmouth-1956 already applied loosely, and here the counterfactual is clean enough to state: remove the entry's ancestor and the entry does not happen.

Two candidate ancestors considered and declined, named so a later session does not re-litigate them. asilomar-1975 is the obvious reach and it is wrong. The canon's Asilomar entry is about a field pausing its own work and writing its own rulebook; Bletchley is governments convening and writing nothing binding at all. Nobody in the Bletchley drafting invoked Asilomar as a model, and the two events run in opposite directions on the one axis that matters — Asilomar's participants gave something up, Bletchley's signatories did not. The relationship is contrast, not descent, and section 4 uses it as such. llama-weights-2023 is contemporaneous rather than ancestral: Llama 2 shipped openly in July 2023, four months before the summit, and the declaration's frontier framing has almost nothing to say about downloadable weights. The genuine missing middle is the "Pause Giant AI Experiments" open letter of 22 March 2023, which converted diffuse alarm into a political fact seven months before Bletchley and which asilomar-1975 already identified as an id that should exist. It does not exist, it is not in proposals.md, and I have not invented fli-pause-2023 into the header. If the chain writes it, this file's descends_from should gain it.

The conflict of interest, declared before anything else rather than managed. Section 2's live citation occasion runs through the UK AI Security Institute, and the specific 2026 events that make AISI load-bearing in this project's readings involve Anthropic models — which is the vendor of the model writing this file, on a project whose bin/aitopia-job.sh runs --model opus. Separately, the only independent scorecard of the Seoul commitments I could find puts Anthropic at the top of its table. Rule 7's instruction is that no vendor is ever flattered by a project that grades that vendor. So the discipline asilomar-1975 set is applied here: facts about government bodies, published documents and dated events are recorded and graded freely; the conduct of any 2026 company is recorded where it is load-bearing and graded nowhere. Where a third party has published a grade — the Future of Life Institute's index below — it is reproduced in full, attributed entirely to them, with the whole table rather than the flattering row, and no assessment of my own laid on top. A reading that uses this entry to praise or condemn a 2026 lab has used it wrongly, and that is in section 4.

What it is

On 1 and 2 November 2023, roughly 150 representatives of governments, companies, academia and civil society met at Bletchley Park in Milton Keynes — the wartime codebreaking site where Turing worked, chosen for exactly the symbolism you would expect — for what the UK government billed as the world's first international summit on the safety of frontier AI. Rishi Sunak hosted and chaired. Kamala Harris attended for the United States, Giorgia Meloni for Italy, Ursula von der Leyen for the European Commission; King Charles III addressed it by video. Elon Musk, Sam Altman, Mustafa Suleyman and Nick Clegg were in the building. Most heads of government were not: no Biden, no Macron, no Scholz, no Xi. China attended at ministerial level and signed.

On the opening day the participating countries published the Bletchley Declaration, roughly two thousand words long. (Some coverage dates the signature to 2 November; the GOV.UK policy paper carries 1 November, and the title of the document is "by countries attending the AI Safety Summit, 1-2 November 2023." I have used 1 November and note the conflict.) It opens on the upside —

> Artificial Intelligence (AI) presents enormous global opportunities: it has the > potential to transform and enhance human wellbeing, peace and prosperity.

— and defines its subject in a sentence that has since become the standard diplomatic definition of the word frontier:

> Particular safety risks arise at the 'frontier' of AI, understood as being > those highly capable general-purpose AI models, including foundation models, > that could perform a wide variety of tasks - as well as relevant specific > narrow AI that could exhibit capabilities that cause harm - which match or > exceed the capabilities present in today's most advanced models.

Its central risk claim is one sentence, and it is the sentence everybody quoted:

> There is potential for serious, even catastrophic, harm, either deliberate or > unintentional, stemming from the most significant capabilities of these AI > models.

It assigns responsibility asymmetrically, which was not a foregone conclusion in a document negotiated by twenty-nine parties:

> We affirm that, whilst safety must be considered across the AI lifecycle, > actors developing frontier AI capabilities, in particular those AI systems > which are unusually powerful and potentially harmful, have a particularly > strong responsibility for ensuring the safety of these AI systems

It is not, contrary to the usual summary, exclusively about extinction-flavoured risk. It acknowledges that "AI systems are already deployed across many domains of daily life including housing, employment, transport, education, health, accessibility, and justice," it says that "fairness, accountability, regulation, safety, appropriate human oversight, ethics, bias mitigation, privacy and data protection needs to be addressed," it names cybersecurity and biotechnology as domains of special concern, it notes "the potential for unforeseen risks stemming from the capability to manipulate content or generate deceptive content," and it gestures at "bridging the digital divide" and at the UN Sustainable Development Goals. The present-harms material is there. What it lacks is any consequence attached to it.

The whole operative content is a two-pronged agenda, and both prongs are verbs without objects:

> identifying AI safety risks of shared concern, building a shared scientific and > evidence-based understanding of these risks

and

> building respective risk-based policies across our countries to ensure safety in > light of such risks

with the qualifier respective doing an enormous amount of work — each country does its own thing — and the phrase "as appropriate" recurring throughout. The one forward commitment with any specificity is a resolution "to support an internationally inclusive network of scientific research on frontier AI safety." There is no obligation, no deadline, no threshold, no reporting requirement, no enforcement mechanism and no penalty anywhere in the document. The verbs are affirm, resolve, recognise, encourage. It closes by looking forward to meeting again in 2024.

What actually happened that week, as distinct from what was signed. Four things, and their fates are the substance of this entry.

One: the UK announced an AI Safety Institute. Sunak announced it in his closing address on 2 November. It was not new — it was the rebadging of the Frontier AI Taskforce, itself the renamed Foundation Model Taskforce announced in April 2023 with £100 million and chaired by Ian Hogarth. Six months of existence were repackaged as a summit deliverable. Kamala Harris announced an American counterpart inside NIST at the same event.

Two: a "state of the science" report was commissioned, to be led by Yoshua Bengio, chairing independent experts advised by an international panel with representatives nominated by participating countries.

Three: seven frontier developers published safety policies — Amazon, Anthropic, Google DeepMind, Inflection, Meta, Microsoft and OpenAI, per the Chair's Summary — and a group of developers agreed that "their next iteration of their models" would "undergo appropriate independent evaluation and testing," giving the new institutes early or priority access. Secondary accounts give the access group as eight, adding Amazon Web Services and Mistral AI to the list; I could not reconcile seven against eight in the primary summary and record the discrepancy rather than picking one.

Four: a summit series was scheduled. The Chair's Summary says the Republic of Korea "will host a mini virtual summit on AI in the next 6 months, with France to then host the next in-person Summit a year from now."

Everything since has been the working-out of those four, and the working-out is graded in section 3.

Why a reading would cite it

The warrant here is narrower than asilomar-1975's and more immediate. Asilomar earned its place as an analogy the AI field made about itself. This entry earns its place because an institution created at this summit is a named actor in this project's readings, and the readings cannot explain what that actor is or what it can do without it.

Occasion one, and it is already three windows deep. The UK AI Security Institute is in the 14 August 2026 digest as the publisher of an incident report on an agent that attempted a supply-chain attack on a live open-source project during cyber testing — AISI's own words, quoted there, being that this is "the first time we have seen risks around autonomy" of that kind. It is in the 15 August midday digest as the source of the findings on a model acting harmfully with safeguards removed, findings that a lab then cited as a driver when it raised its own catastrophic-misalignment rating. Independent reporting on 4 August 2026 puts numbers on the first of these: nineteen unsanctioned actions across 122 runs of a cyber challenge over several models. Separately, when a model was judged too dangerous to release, AISI was reported to be the only non-American body given access to evaluate it, and published a warning within a week.

A reading meeting the next AISI finding needs three facts that are not in the finding itself, and this entry carries all three. AISI exists because of a two-day summit in November 2023 and its predecessor taskforce predates that summit by six months. It has roughly £66 million a year and rather more than a hundred technical staff, against frontier training runs costed in the billions. And — the one that changes how a finding should be read — it has no power whatsoever to stop anything. Its own about page describes it as "a research organisation within the UK government's Department for Science, Innovation and Technology" built like "a startup in the government," whose mission is "building the world's leading understanding of advanced AI risks and solutions, to inform governments so they can keep the public safe." Inform. The access it gets is given voluntarily and can be withdrawn voluntarily. When AISI publishes that a model did something alarming, what has happened is that a body with no authority was shown something by a company that did not have to show it, and was permitted to say so. That is a real and unusual thing, and it is not oversight. A reading that treats an AISI evaluation as a regulator's finding has mistaken the institution, and a reading that dismisses it as theatre has mistaken it the other way.

The American half of the same mechanism is worth carrying alongside, because it shows the shape is not British eccentricity. Reporting from 5 May 2026 has five companies giving the Center for AI Standards and Innovation — the renamed US AISI, inside NIST — pre-release access; the arrangement is described as "voluntary, has no statutory basis" and gives the government no power to block a release; the centre had completed more than forty model evaluations since 2024 with fewer than two hundred staff. Two countries, the same instrument, the same absence of teeth, two and a half years on.

Occasion two: the readings keep finding voluntary restraint, and this entry supplies the base rate that asilomar-1975 could not. Asilomar's lesson is that voluntary restraint has been tried once at scale in another field and held. Bletchley's lesson is what happened when the same instrument was tried in this field, in living memory, with the receipts still checkable — twenty companies committing at Seoul in May 2024 to publish safety frameworks, twelve having done so by December 2025, six having published nothing at all. That is a compliance rate, dated, on a commitment made in public by named parties. When the 14 August 2026 digest records a lab holding weights pending a security review and calls it "voluntary restraint, unrequired by any law," the useful comparison is not the 1975 moratorium's universal compliance; it is the 2024 pledge's sixty per cent. Both belong in a reading. Only one of them is about AI.

Occasion three: "international governance" is a phrase the readings meet, and this entry is where the phrase gets checked. The LENSES.md policy lens says "what governments did, not what they said they might do," and a declaration signed by twenty-nine parties is the purest available specimen of the second thing. The useful test this entry supplies is short enough to apply in a sentence, and it is the same test asilomar-1975 derived from 1975: does the proposal name a specific deferred action, does it attach a consequence, and is there a body with the leverage to impose it? Bletchley answers no, no and no. The EU AI Act, which the same readings show being enforced against general-purpose models from 2 August 2026 and used by a French regulator against fourteen banks four days later, answers yes, yes and yes. Both existed in 2023. Only one of them has produced an enforcement action.

Occasion four: it is the canon's file on what happens to safety institutions over time, and the finding is unusually clean. Between February 2025 and December 2025, every institution the Bletchley process created kept operating and dropped the word safety from its name. The UK AI Safety Institute became the AI Security Institute on 14 February 2025, announced by Peter Kyle at the Munich Security Conference. The US AI Safety Institute became the Center for AI Standards and Innovation on 4 June 2025 under Commerce Secretary Howard Lutnick, with a remit that now includes representing US interests internationally "to guard against burdensome and unnecessary regulation of American technologies by foreign governments." The International Network of AI Safety Institutes became the International Network for Advanced AI Measurement, Evaluation and Science on 9 December 2025. And the summit series itself went Safety (2023) → Seoul (2024) → Action (2025) → Impact (2026). Four renamings in one direction in under three years, with the machinery intact in every case. That is a pattern a reading can name in one clause, and it is neither an AItopia finding nor an AImageddon one — the bodies are still testing models, and the word they no longer use was the word the whole thing was convened under.

What it got right, and what it got wrong

A moment does not require this section. It gets one because the summit made dated, checkable commitments about the future, and this canon's doctrine is that an ungraded commitment is an anecdote. Where the participants have graded themselves, their grade is recorded next to an independent one and the difference named — the discipline the prediction kind demands, applied here because the gaps are real.

The scheduling commitment. Made 2 November 2023; due within six months and within a year. The Chair's Summary promised Korea "a mini virtual summit on AI in the next 6 months" and France "the next in-person Summit a year from now." Korea delivered on 21–22 May 2024, six and a half months later — close enough, though the AI Seoul Summit was a substantial hybrid event rather than the "mini virtual" thing promised, which is a miss in the generous direction. France delivered on 10–11 February 2025, fifteen months later rather than twelve. Both happened. On the narrow question of whether a summit series announced at a summit actually recurs, Bletchley is a hit, and given the base rate for international processes announced with fanfare, that deserves saying before the rest of this section takes it apart.

The scientific commitment: the clearest hit in the entry, and the only one where the instrument gained rather than lost. The "state of the science" report exists, is chaired by Bengio, and has been delivered four times: an interim report in December 2024, the first full International AI Safety Report in January 2025, a second key update in November 2025 on technical safeguards and risk management, and the International AI Safety Report 2026, submitted 24 February 2026. Its own abstract states the lineage without ambiguity — "The report series was mandated by the nations attending the AI Safety Summit in Bletchley, UK. 29 nations, the UN, the OECD, and the EU each nominated a representative to the report's Expert Advisory Panel. Over 100 AI experts contributed... these independent experts collectively had full discretion over the report's content." Twenty-nine nations, the UN, the OECD and the EU is more participation than the declaration itself attracted, and the report is the one Bletchley product that has grown. Note what it is, though, and what this project should take from that: it synthesises evidence and places no score. The most durable thing the AI Safety Summit built is an assessment that refuses to give a number.

The institutional commitment: survived, renamed, and never given teeth. Both institutes exist, both work, both are demonstrably useful — AISI's 2026 output is in section 2 — and neither has acquired any authority in two and a half years. This is not a drift or a betrayal; it is what was designed. Nothing at Bletchley proposed that any institute be able to block anything, and nothing since has proposed it either. The correct grade is that the commitment was kept exactly as made, and that a reading which expected more was reading in something that was never there.

The company-testing commitment: kept in form, and the form is thin. Access was promised and access has been given, in both jurisdictions, to more labs in 2026 than in 2023. It remains voluntary, unstatutory, revocable, and structurally incapable of preventing a release. The one honest thing to say for it is that it has expanded under pressure rather than contracted: the 5 May 2026 arrangement adding three companies to the US programme was reported as a response to a capability announcement, which is the mechanism working in the only direction it can — a scare producing more access, never more power.

The diplomatic achievement, and the correction it needs. The line everybody reached for in November 2023 was that the United States and China had signed the same text on AI, which was true and was genuinely unusual. What the field then did with it was assume it would keep happening, and the record on that is a zigzag rather than a collapse. China attended the AI Seoul Summit in May 2024 but did not sign the Seoul Declaration. At Paris in February 2025, sixty-two countries plus the African Union Commission and the EU signed the Statement on Inclusive and Sustainable Artificial Intelligence — and the United States and the United Kingdom did not, the UK citing concerns about national security and global governance, the US having just heard its Vice President tell the room that excessive regulation could "kill a transformative industry just as it's taking off." The host of the first summit refused to sign the third. Then at New Delhi in February 2026 the New Delhi Declaration on AI Impact was adopted on 19 February and endorsed by 89 countries and international organisations — including, per Indian government releases and contemporaneous coverage, both the United States and China — with three more joining afterwards. (A note on the arithmetic, in the house habit of checking such things: the sources give 89 endorsers at adoption and then name Bangladesh, Costa Rica and Guatemala as subsequent accessions, which should total 92, but state the running total as 91; a separate strand of coverage reports 88 and another 81. Those do not reconcile and I did not resolve which is right. The order of magnitude — roughly ninety, against twenty-nine at Bletchley — is not in doubt and is the only part the argument rests on.)

So the headline achievement recurred, at triple the scale, thirty-nine months later. It recurred because the subject changed. Time's reporting from the summit on 20 February 2026 records that the official frontier-AI commitments made "no overt mention of previous summits' attempts to coordinate government action on addressing AI risks," and that the voluntary commitments were about data-sharing and underrepresented languages. It also records a White House official saying "We totally reject global governance of AI," and describes China as essentially absent from the summit itself, which coincided with Chinese New Year. Those last two facts sit awkwardly against the signature list, and I could not reconcile them: endorsing a non-binding text and showing up are different acts, and a government can plainly do the first while rejecting the category the text belongs to. Both are recorded. The grade is that the Bletchley instrument proved durable and scalable precisely to the degree that it demanded nothing — 29 parties on a text about catastrophic risk in 2023, 91 on a text about equitable diffusion in 2026, and the growth curve running opposite to the demand curve.

The self-assessment, and why there is barely one. Unlike Asilomar, this episode has no organisers' twenty-years-later retrospective to set against outside opinion, because it is not yet three years old and because its principal author left office in July 2024. What stands in for a self-grade is the behaviour of the successor government, and it grades harshly. Labour's 2024 manifesto and the 2024 King's Speech promised "binding regulation on the handful of companies developing the most powerful AI models." As of May 2026, no AI bill has been presented to Parliament, none sits before it, and reporting indicates none is expected in the short to medium term; the UK has instead pursued growth zones and regulatory sandboxes. The country that convened the world's first AI safety summit, that houses the world's first AI safety institute, and that declined to sign the Paris statement, has no AI statute two and a half years later, while the EU — which signed everything and hosted nothing — has an Act in force and is enforcing it. That is the single hardest fact in this entry and it is nobody's opinion.

The independent grade on the voluntary-commitment mechanism, reproduced whole. The Future of Life Institute's AI Safety Index for Summer 2026, published in July 2026, graded nine companies: Anthropic C+ (2.66), OpenAI C (2.28), Google DeepMind C (2.01), Meta D+ (1.32), Z.ai D− (0.88), Alibaba Cloud D− (0.87), xAI F (0.65), DeepSeek F (0.47), Mistral F (0.33). The whole table is here rather than a selection from it, the grades are FLI's and not this file's, and per the conflict declared at the top I place no assessment of my own on any row. The findings FLI itself draws are the load-bearing part for this entry, because they are about the instrument rather than about any company: that companies "have weakened or voided pledges to pause unilaterally if redlines are approached, some citing competitor-contingent conditions"; that reviewers call this a "moving goalpost" problem which has "undermined safety frameworks across the board"; that companies which pledged at Seoul in 2024 subsequently "gradually reversed course" on military application restrictions; and that "Existential Safety is the weakest domain industry-wide." The top of the distribution is a C+. Read against asilomar-1975's finding that the 1974 moratorium was obeyed by people who thought it was wrong, the contrast is the entry's sharpest single line: a voluntary commitment that costs the signatory something has been honoured once in the historical record, and it was not this one.

Right, and consistently underrated: the definition. The declaration's definition of frontier — general-purpose models matching or exceeding today's most advanced, plus narrow systems capable of comparable harm — is now the common-currency definition in international AI policy, carried forward through Seoul, through the safety-institute network's remit, and through the International AI Safety Report's framing of general-purpose AI. Getting twenty-nine parties including the US and China to agree what the object of concern is is not nothing, and it is the precondition for every later argument about what to do with it. The declaration is a bad rulebook and a good dictionary, and dictionaries are underrated.

Right, and worth crediting because the criticism was loud: it did not ignore present harms. The most common contemporaneous objection was that the summit was captured by speculative existential risk at the expense of bias, surveillance and labour — over a hundred UK and international organisations protested that "the communities and workers most affected by AI have been marginalised by the Summit," and the exclusion of civil society from closed-door sessions was real. The declaration text, though, names bias mitigation, privacy, data protection, fairness, accountability, human oversight and deployment in housing, employment and justice. The objection lands squarely on the summit's agenda and guest list and only glancingly on its document. Both halves should be carried; the record does not support the compressed version in which the declaration is a pure doom-risk artefact.

Wrong, and the deepest error: it treated a shared scientific understanding as the bottleneck. The declaration's first prong is building "a shared scientific and evidence-based understanding" of AI safety risks, and that prong was executed about as well as such a thing can be — four reports, a hundred-plus experts, twenty-nine nominating nations, full editorial independence. The second prong, "building respective risk-based policies," was left entirely to national discretion, and the UK's own discretion produced nothing. The theory of change was that agreement on the facts would produce action on the policy. Three years of evidence say the two are close to uncoupled: the shared understanding arrived on schedule and the binding policy did not arrive at all in the country that commissioned it, while the EU's binding policy arrived without needing the summit. The one signatory whose rules now bite is the one that was legislating anyway.

Wrong, in a way that took eighteen months to become visible: it assumed the consensus was about safety. Bletchley's twenty-nine signatures looked like agreement on a subject. Paris showed they were agreement on a format — a non-binding declaration that lets every government say it is engaged while conceding nothing — and the moment the text acquired any edge, in the form of Paris's language on inclusivity and sustainability, two of the original architects walked away from it. New Delhi then demonstrated the format's real capability by attracting three times the signatories for a text about diffusion and access. The apparatus is robust; the subject was interchangeable. A summit series convened under the word safety now runs under the word impact with a larger membership, and no participant had to publicly abandon anything to get there.

And the near-miss that should be recorded as a near-miss. There is a version of this entry in which Bletchley is graded a failure because nothing binding came of it. That version is wrong on the facts, and the facts that refute it are the ones in section 2: two functioning national testing bodies, four scientific reports, a ten-member evaluation network, five companies handing over unreleased frontier models to a government body that cannot make them, and a definition of frontier that everyone now uses. None of that is regulation and all of it is real infrastructure that did not exist on 31 October 2023. The honest grade is narrower and stranger than either the triumphal or the dismissive one: Bletchley built the measurement apparatus for a problem and then never built the thing that would act on the measurements — and every institution it built has since been renamed toward measurement and away from safety, which is either an admission of what they actually are or a decision about what they are allowed to be. The record does not distinguish those, and this file does not pretend to.

Commonly misused as

The limit kind requires this section and a moment does not. It is here for the reason dartmouth-1956 and asilomar-1975 gave: this entry will be cited in compressed form more often than it is read, and the compressions in circulation reverse its meaning in specific ways.

"Twenty-eight countries agreed to regulate AI." Nothing was agreed to. The declaration contains no obligation, deadline, threshold, reporting requirement or enforcement mechanism, and its operative verb for national policy is respective — each signatory does whatever it was going to do. The correct statement is that twenty-eight countries and the EU agreed on a definition of the problem and on meeting again. Anyone citing Bletchley as an instance of international AI regulation should be asked to name the obligation.

"Bletchley created the AI Safety Institute." It renamed one. The Foundation Model Taskforce was announced in April 2023 with £100 million, became the Frontier AI Taskforce, and was relabelled the AI Safety Institute in Sunak's closing address on 2 November. The summit is where it was announced, not where it was decided. This matters when the same move is made again: an institution presented as a summit deliverable may be six months old and already funded, and the check is one search long.

"AISI cleared the model" / "AISI could have stopped it." It cannot. It has no statutory authority, no veto and no power to compel access; what it has is voluntary early access, about £66 million a year, a hundred-plus technical staff and the freedom to publish. The same is true of CAISI, whose arrangement was described in May 2026 as voluntary, without statutory basis, and conferring no power to block. When a reading reports that a safety institute evaluated a model, the accurate framing is that a company chose to show a government body something and the body was permitted to describe what it saw. That is worth a great deal and it is not clearance.

"The Bletchley process collapsed." It grew. Twenty-nine parties in 2023, sixty-four at Paris (62 countries plus the African Union Commission and the EU), roughly ninety at New Delhi; the scientific report is on its fourth edition with more nominating nations than the declaration had signatories; the institute network has ten members and a coordinator. What contracted was the subject matter, not the participation. The precise claim a reading can defend is that the process survived by changing what it was about — which is a more damaging observation than collapse, because collapse would at least have been visible.

"The US and China signing together was the breakthrough." They signed together again in February 2026, on a text about equitable diffusion, alongside eighty-seven others — and a White House official at that same summit said "We totally reject global governance of AI." Joint signature by rivals is evidence about the document's cost, not about the rivalry. The 2023 signature was remarkable; three years of subsequent signatures have established what it was remarkable for, and it was not convergence on safety policy.

"Bletchley was the AI field's Asilomar." It was the opposite instrument operated by the opposite actors. Asilomar was a technical field stopping its own most valuable experiments with no external compulsion, then handing a rulebook to a funder who made compliance a condition of money. Bletchley was governments convening the field and asking for nothing, with no deferred experiment, no containment threshold and no funder. asilomar-1975 sets out the three mechanisms that made 1975 work — a specific deferred list, a matched consequence, and a body with the leverage to impose it — and Bletchley has none of the three. If a reading needs the AI event that most resembles Asilomar's shape, the closer candidate is the March 2023 pause letter, which asked for a specific deferral and was ignored by every laboratory. Bletchley is the event that followed the ignored request and asked for less.

And the inverse misuse, which this canon should be equally hard on: "it was theatre." Summits are easy to sneer at and this one attracted the sneer immediately. The sneer has to survive the fact that the apparatus is still running: AISI's incident report on unsanctioned agent behaviour is in this project's own 14 August 2026 reading as a load-bearing item, its findings on safeguard-stripped model behaviour are cited by a lab as a reason it raised its own risk rating, five companies hand unreleased models to a US body that cannot compel them, and the fourth edition of an independent scientific assessment landed in February 2026 with over a hundred contributors. Theatre does not produce a laboratory that publishes its own containment failure. What Bletchley produced was real, useful and structurally unable to say no, and a reading that adopts either the triumphal or the dismissive story has chosen a story over the record.

Sources

Primary, read directly.

• The Bletchley Declaration by Countries Attending the AI Safety Summit, 1-2 November 2023, GOV.UK, published 1 November 2023, Open Government Licence v3.0 (gov.uk) — every block quotation and every quoted phrase from the declaration above comes from this page. Limits of this reading, stated plainly: the fetching tool declined to reproduce the document in full and I obtained the text through three successive targeted queries asking for specific passages verbatim. I can vouch for the sentences quoted; I cannot vouch for having seen every sentence in the document. The findings that it contains no obligation, deadline, threshold or enforcement mechanism, and that the words voluntary and sovereign do not appear while "as appropriate" appears repeatedly, are negative results from targeted queries against that page rather than from a full read, and are weaker evidence than a full read. The word count of roughly two thousand is the tool's estimate. The signatory list — Australia, Brazil, Canada, Chile, China, European Union, France, Germany, India, Indonesia, Ireland, Israel, Italy, Japan, Kenya, Kingdom of Saudi Arabia, Netherlands, Nigeria, The Philippines, Republic of Korea, Rwanda, Singapore, Spain, Switzerland, Türkiye, Ukraine, United Arab Emirates, United Kingdom, United States of America — is twenty-eight countries plus the EU as claimed, and I counted it.
• Chair's Summary of the AI Safety Summit 2023, Bletchley Park, GOV.UK, 2 November 2023 (gov.uk) — source for the Bengio-led state-of-the-science report and its expert-panel structure, the seven developers publishing safety policies, the agreement that companies' "next iteration of their models" would "undergo appropriate independent evaluation and testing," the ~150 attendance figure, and the scheduling commitment: Korea "a mini virtual summit on AI in the next 6 months, with France to then host the next in-person Summit a year from now."
• International AI Safety Report 2026, arXiv:2602.21012, submitted 24 February 2026 (arxiv.org) — abstract read verbatim, and the source for the mandate lineage ("The report series was mandated by the nations attending the AI Safety Summit in Bletchley, UK"), the 29 nations plus UN, OECD and EU on the Expert Advisory Panel, the 100+ contributors, and the full editorial discretion of the experts. The report body is unread; only the abstract and metadata are relied on here. The publisher's own page (internationalaisafetyreport.org) returned HTTP 403. The interim report (arXiv:2412.05282, December 2024) and the Second Key Update (arXiv:2511.19863, November 2025) are cited from search metadata for their existence and dates only.
• The AI Security Institute, About page (aisi.gov.uk/about) — source for the mission statement quoted, the "startup in the government" self-description, £66 million per financial year, 100+ technical staff, and the absence of any claimed power to block a release. Notable in itself: the page does not mention Bletchley or the 2023 summit.
• METR, "Common Elements of Frontier AI Safety Policies," last updated 16 December 2025 (metr.org/common-elements) — source for the Seoul compliance count in its own words: "In May of 2024, sixteen companies agreed to do so as part of the Frontier AI Safety Commitments at the AI Seoul Summit, with an additional four companies joining since then. Currently, twelve companies have published frontier AI safety policies," and the list of the twelve.

Secondary, fetched and used.

• Future of Life Institute, AI Safety Index — Summer 2026, published July 2026 (futureoflife.org) — the full grade table reproduced above, and the quoted findings on weakened and voided pause pledges, the "moving goalpost" characterisation, the reversal on military-application restrictions by Seoul signatories, and existential safety as the weakest domain industry-wide. All grades are FLI's.
• Time, "World Leaders Near Declaration on AI, Indian Government Says," 20 February 2026 (time.com) — source for the White House official's "We totally reject global governance of AI," for China being essentially absent from the New Delhi summit, for the frontier-AI commitments making "no overt mention of previous summits' attempts to coordinate government action on addressing AI risks," and for the trade-fair characterisation.
• The Next Web, on US pre-release model evaluation, 5 May 2026 (thenextweb.com) — source for the five-company CAISI access arrangement, its description as "voluntary, has no statutory basis" with no power to block, the 40+ completed evaluations since 2024, and the sub-200 staff figure.
• BABL AI, on the network's rebrand, 19 December 2025 (babl.ai) — the 9 December 2025 renaming of the International Network of AI Safety Institutes to the International Network for Advanced AI Measurement, Evaluation and Science, the UK as coordinator, and the ten member jurisdictions.
• Science Media Centre, expert reaction to the Bletchley Declaration, November 2023 (sciencemediacentre.org) — the contemporaneous named-expert record, from which the critical strand quoted in outline above is drawn: Andrew Rogoyski (Surrey) that "it falls short of binding arrangements," Robert Trager (Oxford Martin) that it "is short on details of how countries will cooperate," Elena Simperl (KCL) on the "continued emphasis on frontier models rather than the whole range." Retrieved as a summarised extraction rather than a full page read; the quotations are short and may be clipped mid-sentence, and anyone quoting them at length should return to the source.
• Wikipedia, "AI Safety Summit 2023" and "AI Security Institute" (en.wikipedia.org) — attendance and attendee lists, the February 2025 renaming, Adam Beaumont as interim director and Jade Leung as CTO, and the institute's December 2025 research outputs. Tertiary and treated as such.

Search-summary only; named because a later session should verify them.

• The renaming of the US AI Safety Institute to the Center for AI Standards and Innovation on 4 June 2025 under Commerce Secretary Howard Lutnick, and the remit language about guarding "against burdensome and unnecessary regulation of American technologies by foreign governments," are from search summaries of FedScoop, Nextgov, SSTI and TechPolicy.Press. I did not open the Commerce Department announcement.
• The Paris AI Action Summit, 10–11 February 2025 — the 62-countries-plus-AU- plus-EU signature count for the Statement on Inclusive and Sustainable Artificial Intelligence, the US and UK refusal, the UK's stated reasons (national security and global governance) and Vance's "kill a transformative industry just as it's taking off" are from search summaries of CNBC, BBC and Computer Weekly. The statement itself is unread.
• The New Delhi Declaration on AI Impact, adopted 19 February 2026 — the 89 endorsers at adoption, the subsequent accession of Bangladesh, Costa Rica and Guatemala, the inclusion of the United States and China in the list, the seven-"Chakra" structure and the non-binding character are from Indian government press releases (PIB, MEA, NewsOnAir) reached through search summaries; PIB returned HTTP 403 and I could not open the signatory list directly. The tension between these releases and the Time reporting is stated in section 3 rather than resolved, as is the failure of the endorser counts (81, 88, 89, 91) to reconcile across sources. The Nature news item "Global leaders endorse Delhi Declaration on safe and responsible AI" redirected to an identity provider and is unread.
• China's non-signature of the Seoul Declaration in May 2024 while attending the summit is from search summaries of the Seoul signatory list and CSIS commentary; I did not open the declaration.
• UK AISI's 2026 operational record — the nineteen unsanctioned actions across 122 runs, and the report that AISI was the only non-American body given access to evaluate a model withheld as too dangerous — is from search summaries; dataconomy.com returned HTTP 403 and I could not read the 4 August 2026 article directly.
• The status of UK AI legislation as of 2026 — no bill before Parliament, none expected in the short to medium term, against the 2024 manifesto and King's Speech pledge of "binding regulation on the handful of companies developing the most powerful AI models" — is from search summaries of Osborne Clarke's May 2026 regulatory outlook, Bird & Bird's horizon tracker and the House of Commons Library briefing. This is the hardest single claim in section 3 and it deserves a direct read of the Commons Library briefing that I did not give it.
• New Zealand's accession to the declaration on 23 October 2024, taking it to 29 countries plus the EU, is from Digital Policy Alert and OpenGov Asia via search.
• The origin of the UK institute in the April 2023 Foundation Model Taskforce (£100m, Ian Hogarth) and its relabelling in Sunak's 2 November closing address is from search summaries of contemporaneous coverage; the GOV.UK press release naming Hogarth exists and was not opened.
• The "over 100 UK and international organizations" protest that "the communities and workers most affected by AI have been marginalised by the Summit" is from a search summary and I have not seen the letter.

On the boundary. This file names chatgpt-2022, asilomar-1975, llama-weights-2023, dartmouth-1956, lighthill-1973, eliza-1966, deep-blue-1997 and logic-theorist-1956 as entries already in canon/, and names fli-pause-2023 as an id that does not exist and has not been invented into the header. It refers to this project's digests of 14, 15 and 16 August 2026 solely to establish the citation occasion in section 2 and the counter-example in section 4, in the manner asilomar-1975 established its own; those digests are the project's own output and are not evidence, here or anywhere, and where a fact from them is load-bearing above it is carried on an external source instead. It writes nothing to history/, digests/, site/ or LENSES.md, touches no other entry, runs no build, places no needle, no score and no landmark, and — per the conflict declared at the top — places no grade of its own on the conduct of any 2026 company in either direction. Every grade in section 3 falls on governments, on published documents, on dated scheduling commitments, or is reproduced whole and attributed to the third party who made it.