Editor's note.
Every clause number, bulletin number and syndicate number in this article was read at the issuing market's own domain: lmalloyds.com for the Lloyd's Market Association clause library, and ldc.lloyds.com for the Lloyd's market directory. Both are public and neither requires a login. Effective dates for LMA5400 to LMA5403 sit behind the Lloyd's Wordings Repository and are deliberately not stated here. Where a fact rests on trade or law firm analysis rather than on the market's own record, this article says so and does not state the fact.
- The Lloyd's Market Association publishes no AI clause series. Its published position is that AI is already caught by the computer system definition in its existing cyber clauses.
- LMA5566A is titled on its own face: War and Cyber Operation Exclusion No. 3, for use on commercial cyber insurance contracts, dated 18 January 2023. It has been cited in published reference lists as an artificial intelligence exclusion and as a model AI endorsement. It is neither.
- LMA5400 to LMA5403 are model cyber clauses for Difference in Conditions, Engineering, Nuclear, Power Generation, Cargo, Energy, Fine Art and Marine. They are not systemic cyber catastrophe clauses and they are not AI clauses.
- The 31 March 2023 requirement in circulation belongs to Lloyd's Market Bulletin Y5381 of 16 August 2022, for standalone cyber policies in risk codes CY and CZ. The model clauses answering it are the LMA's state-backed cyber attack exclusion series, the family LMA5566 belongs to.
- No published source names a Lloyd's syndicate for Armilla's AI liability policy. One number in circulation, 1176, is Chaucer's nuclear syndicate.
- The practical consequence for a buyer is the opposite of what the invented clauses imply: AI is not sitting inside a purpose-built exclusion. It is sitting inside definitions written for something else, and that is a harder position to read at renewal.
Section 1. What was checked, and what came back
Four identifiers, all of which appear in current market commentary and appeared in the reference lists of pages on this site until this month. Each was checked at the body that issues it.
| Identifier as circulated | What the issuing body's own record says |
|---|---|
| LMA5566 is not an artificial intelligence exclusion and not a model AI endorsement, though it is cited as both. | LMA5566A carries its title on its face: War and Cyber Operation Exclusion No. 3 (For use on commercial cyber insurance contracts), dated 18 January 2023. It is a state-backed cyber war clause. There is no AI subject anywhere in it. |
| The LMA5400 series is not a set of AI model exclusion clauses running from 2024 to 2026, and no such series exists. | LMA5400 to LMA5403 are model cyber clauses for Difference in Conditions, Engineering, Nuclear, Power Generation, Cargo, Energy, Fine Art and Marine. The date range is invented. The LMA publishes no AI clause series. |
| LMA5401 to LMA5403 are not systemic cyber catastrophe exclusion clauses effective March 2023, contrary to a pairing that circulates widely. | Wrong clause family for that requirement. The requirement that bites from 31 March 2023 comes from Lloyd's Market Bulletin Y5381 of 16 August 2022, for standalone cyber policies in risk codes CY and CZ. The model clauses answering it are the LMA's state-backed cyber attack exclusion series, the family LMA5566 belongs to. |
| Armilla is not writing AI liability through Chaucer syndicates 1084 and 1176, though that pairing is widely repeated. | Lloyd's own directory records 1176 as Nuclear Syndicate 1176, classes of business cargo, engineering, nuclear and terrorism. It writes no cyber, no professional indemnity, no directors and officers and no general liability. Syndicate 1084 does carry cyber, professional indemnity and directors and officers, but no published source names any syndicate number for this programme. |
The shape these share is worth naming, because it is the hardest kind of error to catch in this field. Each is a real identifier from a real body, given a false subject. The number checks out the moment anyone looks it up. Only the meaning is wrong, and meaning is the part a search does not verify.
Section 2. The LMA's actual position on AI
The Lloyd's Market Association has not issued an AI clause family, and its published reasoning is that one is not needed: artificial intelligence is already caught by the computer system definition used in its existing cyber clauses. That is a substantive position rather than an absence of one, and it is the single most useful fact in this article for a buyer.
It means AI exposure inside a London market cyber placement is governed by definitions written for a different technology generation. A computer system definition drafted to capture networks, hardware, software and data does capture a machine learning system, straightforwardly. What it does not do is distinguish between the two failure modes that matter most in AI liability. A system that is compromised by an attacker and a system that is simply wrong are, under such a definition, both events involving a computer system. Under the perils and triggers layered on top of it, they are usually not the same thing at all.
That is the structural gap. It is not created by an AI exclusion. It is created by the absence of one, combined with a trigger architecture built around security failure and data breach. This is why buyers who go looking for an AI exclusion in their cyber policy frequently do not find one and conclude, wrongly, that they are covered.
The same distinction appears in the one piece of published affirmative cyber wording that names AI directly. Coalition's Affirmative Artificial Intelligence Endorsement, dated 26 March 2024, on US surplus and Canada cyber policies, has exactly two limbs: security failure and data breach extended to an AI security event, and funds transfer fraud extended to fraudulent instruction by deepfake or other AI. Both limbs are about AI as the instrument or the target of an attack. Neither reaches loss caused by an AI system producing a wrong output in ordinary operation. A separate Coalition endorsement covers deepfake response, with forensics, legal takedown and crisis communications, available in the United States, the United Kingdom, Canada, Australia, Germany, Denmark, Sweden and France.
Read together, the position is coherent and it is not what the market commentary suggests. The cyber market has extended toward AI as an attack surface. It has not extended toward AI as a source of professional error, and it has not written a clause excluding it either, because the underlying definitions were never built to admit it in the first place.
Section 3. Y5381, and how one date travelled to the wrong clause
Lloyd's Market Bulletin Y5381, dated 16 August 2022, addresses standalone cyber policies written in risk codes CY and CZ and carries a requirement effective from 31 March 2023 concerning state-backed cyber attack exclusions. The LMA model clauses answering it are its state-backed cyber attack exclusion series, the family LMA5566 belongs to.
What appears in circulation instead is the 31 March 2023 date attached to LMA5401 to LMA5403, described as systemic cyber catastrophe exclusions. Nine instances of that pairing were sitting on a single page of this site. The date is real, the clause numbers are real, and the pairing is not.
This is worth a paragraph of its own because it shows the mechanism. A correct fact about the London market, that a significant exclusion requirement took effect on 31 March 2023, is remembered. The clause numbers attached to it are not, and are reconstructed from an adjacent clause family. The result carries the confidence of the true half. Anyone verifying the date confirms it and stops. Nobody opens the clause.
Section 4. The syndicate question, and what is actually published
What is published, at the parties' own sites, is this.
Armilla Insurance Services is a Coverholder at Lloyd's. Its Affirmative AI Liability Insurance is described as underwritten by certain underwriters at Lloyd's, and it was the first Lloyd's coverholder dedicated exclusively to AI liability, in 2024. Its standalone AI Liability Policy is written up to USD 25 million per organisation. Its scope includes AI regulatory violations, including defence costs and insurable fines under the EU AI Act and the Colorado AI Act. Recognised partners named by Armilla are Chaucer, AXIS Capital, Convex, Swiss Re and Greenlight Re. Armilla and Trustible announced a first-to-market collaboration on 8 October 2025.
Chaucer and Armilla launched Vanguard AI, a coordinated cyber, technology and AI liability structure, on 10 February 2026. Chaucer's own announcement of it says only that it is backed by Lloyd's of London. No syndicate number appears in it, and none appears at armilla.ai either.
So the honest statement is that the Chaucer relationship is published and the syndicate attribution is not. Where a syndicate number has been supplied in secondary coverage it is inference, and in at least one case it is wrong in a way that a single lookup would have caught: syndicate 1176 is Nuclear Syndicate 1176, managed by Chaucer Syndicates Limited, with classes of business recorded at Lloyd's as cargo, engineering, nuclear and terrorism. A nuclear syndicate is not writing AI liability. Syndicate 1084, also managed by Chaucer Syndicates Limited, does carry cyber, professional indemnity and directors and officers, which makes it the defensible part of that pairing. Defensible is not the same as published.
A buyer does not usually need the syndicate number. A broker placing the risk does, and will have it from the slip rather than from a publication. The reason to state the limit of what is public is narrower: a reader assessing whether a market is real should be able to tell which parts of the picture are documented and which are reconstructed. In this market, more of it is reconstructed than the confident tone of most coverage suggests.
Section 5. The one LMA AI document that does exist, and what it is about
The Lloyd's Market Association published an AI Adoption Toolkit on 23 April 2026, produced with Barnett Waddingham and organised in five themes. It reports that 93 per cent of respondents have or are developing formal AI frameworks, and that respondents represent firms holding over 60 per cent of Lloyd's stamp capacity.
It is a document about how firms in the Lloyd's market adopt AI in their own operations: underwriting workflow, claims handling, governance of the market's own use of these tools. It is not a wordings document and it contains no clause. Citing it as evidence of AI clause development, which happens, confuses the market's use of AI with the market's coverage of AI. Those are two entirely separate developments moving at different speeds, and the first is considerably further along than the second.
Section 6. What a European buyer should do with this
Five questions, in the order they are worth asking. They are designed for a renewal conversation rather than for a coverage dispute, which is the point.
| Ask | What you are testing |
|---|---|
| What is the clause reference and edition date of anything you are describing to me as AI wording? | Whether AI wording exists on this placement at all. If the answer is a number, ask what the clause is titled on its own face before accepting the characterisation. |
| Does this policy respond to loss caused by the AI being wrong, as distinct from the AI being attacked? | The distinction the computer system definition does not draw and the trigger architecture does. This is the single most important question on the list. |
| Where does the professional services definition end? | Whether professional indemnity reaches an output that no human reviewed. In most current wordings this is the live question, and it is a definitional one rather than an AI one. |
| Is there a sublimit, and what is the number? | Affirmative language without a stated sublimit is worth less than it reads. We do not publish sublimit percentages, because no carrier publishes one that we have been able to read at source. |
| What would you need from us to move this from silent to affirmative? | Converts the conversation into a submission question. What underwriters ask for is set out in the submission guide on this site. |
The larger point behind all five is that silence is not the same as cover, and the absence of an AI exclusion is not evidence of an AI grant. A market that has deliberately declined to write an AI clause, on the stated basis that its existing definitions already reach AI, has left buyers to work out for themselves whether the perils those definitions sit inside reach the loss they are actually worried about. For most European operators the loss they are worried about is an agent being confidently wrong in front of a customer, and that is the loss the cyber tower is least likely to answer.
Section 7. Where the regulatory clock sits alongside this
Two dates are relevant to how quickly this matters. The AI Omnibus, Regulation (EU) 2026/1744, entered into force on 27 July 2026 and moved the Annex III high-risk obligations to 2 December 2027 and the Annex I obligations to 2 August 2028. It did not move the transparency and governance rules, which apply from 2 August 2026, and it did not move Directive (EU) 2024/2853, the revised Product Liability Directive, whose national transposition deadline is 9 December 2026.
Underwriting has not followed the regulatory deadline in either direction. Specialist capacity expanded through the whole period in which the delay was uncertain, and specialist underwriters price to operator practice rather than to application dates. What the December 2026 transposition changes is not the regulatory obligation but the evidence position, and the evidence a claimant can seek disclosure of overlaps substantially with the evidence an underwriter asks for. That overlap is treated in the compliance documentation to insurance evidence chain, and the full December date cluster on agentliability.eu.
Questions
Is there an LMA artificial intelligence exclusion clause?
No. The Lloyd's Market Association publishes no AI clause series. Its published position is that AI is already caught by the computer system definition used in its existing cyber clauses, so a separate AI clause family has not been issued. Several clause numbers circulate in trade coverage as AI exclusions. Checked at lmalloyds.com, each is a cyber clause. LMA5566A states its own title on its face: War and Cyber Operation Exclusion No. 3, dated 18 January 2023. LMA5400 to LMA5403 are model cyber clauses for Difference in Conditions, Engineering, Nuclear, Power Generation, Cargo, Energy, Fine Art and Marine.
If there is no AI clause, how is AI treated inside a London market cyber policy?
Through the computer system definition and the exclusions that attach to it, rather than through anything AI-specific. That has a practical consequence for buyers: whether an AI loss falls inside or outside a cyber policy is determined by the same definitions that govern any other technology loss, and by whether the loss is a security or data event at all. An AI system producing a wrong answer with no security failure and no data breach is not obviously a cyber loss under those definitions, whatever the policy says about computer systems.
What is Lloyd's Market Bulletin Y5381 and does it concern AI?
Y5381 is a Lloyd's Market Bulletin dated 16 August 2022 concerning standalone cyber policies in risk codes CY and CZ, with a requirement effective from 31 March 2023. It concerns state-backed cyber attack exclusions. It does not concern artificial intelligence. The model clauses answering it are the LMA's state-backed cyber attack exclusion series, the family LMA5566 belongs to. Trade coverage has repeatedly paired the March 2023 date with the wrong clause family and with an AI subject; neither pairing survives a check at the LMA.
Which Lloyd's syndicate writes Armilla's AI liability policy?
No published source names one. Armilla Insurance Services is a Coverholder at Lloyd's and its Affirmative AI Liability Insurance is described as underwritten by certain underwriters at Lloyd's. Chaucer and Armilla launched a coordinated cyber, technology and AI liability structure called Vanguard AI on 10 February 2026, and Chaucer's own announcement says only that it is backed by Lloyd's of London. Syndicate numbers attached to that programme in secondary coverage are inference. One number in circulation, 1176, is Chaucer's nuclear syndicate, whose classes of business at Lloyd's own directory are cargo, engineering, nuclear and terrorism.
What should a European buyer ask a broker about AI wording at renewal?
Ask for the clause reference and the edition date of anything described as AI wording, then ask what the clause is titled on its own face. A clause that turns out to be a cyber war exclusion is not silent on AI by accident; it was never about AI. Ask separately whether the policy responds to loss caused by an AI system being wrong, as distinct from an AI system being attacked. Those are different perils and most current affirmative wording addresses only the second.
Does the LMA AI Adoption Toolkit contain AI policy wording?
No. The LMA AI Adoption Toolkit, published 23 April 2026 with Barnett Waddingham, is about how firms in the Lloyd's market adopt AI in their own operations. It is organised in five themes and reports that 93 per cent of respondents have or are developing formal AI frameworks, with respondents representing firms holding over 60 per cent of Lloyd's stamp capacity. It is a market adoption document, not a wordings document, and citing it as evidence of AI clause development confuses the market's use of AI with the market's coverage of AI.