London has already handed over part of the job
Thames Water now listens to its pipes through a network of over 75,000 listening sensors, against 21,000 in 2025, according to the utility’s figures reported by New Civil Engineer. That is one strand of what Lucy Siegle, writing in The Standard, describes as an AI network spreading under and through the capital, largely unseen by the people who rely on it. Her question is not whether any single tool works. It is what happens to the city when the layer as a whole stops working, or is made to stop. This analysis takes that question seriously, and finds that London’s published risk register still lists AI only among its chronic risks, not as an emergency scenario.
Strategic Insight: The published figures for these systems report what they deliver: more leaks found, faster notes, more housing sites, more arrests. Her question is a different one: what the city does when several of them fail together.
What does London’s AI layer actually do?
The examples Siegle gathers come from water, health, housing and policing, not from AI labs.
Four systems, four kinds of dependency
At St George’s in Tooting, an AI note-taking tool was piloted in the emergency department. The trust says it cut clinicians’ documentation time by half, freeing 47 minutes per clinician per shift, and lifted the number of patients seen per shift by 13.4%. In Lewisham, the council piloted an AI mapping tool, the Small Sites AI Finder built with RCKa, which flagged over 3,000 sites where homes might be built. In Croydon, the Met mounted live facial recognition cameras on lampposts at either end of the high street; over six months of trial deployments, upwards of 470,000 people passed them and police arrested more than 170 suspects.
| Metric | Value | Strategic Implication |
|---|---|---|
| Thames Water acoustic sensors | 75,000+ (2025: 21,000) | Leak-finding runs on a network that has more than tripled since 2025 |
| St George’s A&E paperwork time | Halved; 47 minutes saved per clinician per shift | The gain is measured in patient capacity, up 13.4% per shift |
| Lewisham small sites identified | Over 3,000 | Results inform planning policy, with officers checking each site |
| People passing Croydon cameras | More than 470,000 in six months | Static deployments are continuing |
Strategic Reality: Around 99% of the food and drink feeding London, some 6,347,000 tonnes, is brought in from beyond its boundaries, according to a London Assembly release citing ReLondon. Local production accounts for under 1%, so the capital depends on the systems that move the rest in.
Why the dependency is hard to see
These systems differ in how reversible they are. Lewisham’s results still had to be checked by officers, with planners assessing each site before it could be considered, so officers’ own judgement stays in the loop. A leak network that went from 21,000 sensors to over 75,000 is different: Thames Water says the sensors have almost doubled how accurately it pinpoints leaks, and losing them would mean giving that gain back.
Siegle borrows a framing from Loomery co-founder Tim Checkley that captures the shift. In the first wave, AI helps people do existing work faster. In the second, organisations redesign the work around AI agents, and removing the AI stops the process. “Wave one was reversible. Wave two isn’t,” was Checkley’s formulation at Loomery’s Weave summit. The summit was about businesses, but the same test can be put to a hospital rota or a utility’s leak programme.
Critical Context: Professor Jason McEwen, the Alan Turing Institute’s interim chief scientist, told The Standard that “once AI is deeply integrated, it’s not the case that we can always just switch things off.” He argues systems need to “fall back to a graceful, safe state”. The institute has just released its frontier AI risks report, and McEwen puts the principle simply: “It’s the safety of the whole system.”
What the experts in the piece actually worry about
The most useful voice in Siegle’s article plays down the dramatic scenarios. Dr Stephanie Hare, who co-presents Artificial Intelligence: Decoded on the BBC, says “the sci-fi scenario is not actually what we should be worried about.” Her concern is cybersecurity, and the imbalance between defender and attacker. “If you’re an institution or company, you have to defend your entire attack surface,” she says, whereas “the attacker only needs to get it right once.” Her illustration is a bank: get in, and people “cannot access their money”.
Hare’s second worry is slower. A graceful fallback assumes there are still people who can take over. She warns that people trust AI models that “are often wrong”, and that “you have to know that they’re wrong in order to challenge them.”
- Fallback needs a human who can do the job: Hare’s 30-year scenario, a surgeon “who is using AI instead of having been trained properly”, is a resilience problem as much as a skills one.
- Accountability needs a named person: Theo Blackwell, the capital’s chief digital officer, says “if an AI system gets something wrong, there needs to be a clear way for a person to check it, challenge it and ultimately take responsibility.”
- Safety is a property of the deployment: McEwen argues models must be shown to be safe in the operating environments they are deployed into, not only in the abstract.
⚠️ Warning: The trust, utility, council and police material reviewed for this analysis reports gains: time saved, leaks found, sites identified, arrests made. None of it describes a fallback or contingency for when the system is unavailable.
Who carries the risk?
The exposure differs by system. In emergency care, the tool’s benefit is counted in patient capacity, which is what a department would stand to lose without it. Planning teams in Lewisham depended on what officers knew and on checking sites by hand before the pilot, and still review every site, so they keep a manual fallback. For people walking along Croydon high street, the more pressing question is error: how a wrong match is checked and challenged, which is the test Blackwell sets.
| Group | What changes for them | What is still unknown |
|---|---|---|
| A&E clinicians and patients | Documentation time is halved during the pilot | How a department copes if the tool fails mid-shift |
| Planning officers in Lewisham | A wider view of available land than checking by hand gave | Whether manual assessment skills are kept up |
| People passing Croydon cameras | Faces compared against a watchlist built for each deployment | How misidentifications are challenged in practice |
| London households | A food supply almost entirely imported | How far just-in-time delivery systems rely on AI routing |
City Hall’s own description of the food system is a supply dependent on “a complex set of interdependencies and just-in-time delivery systems”, as Siegle quotes it. The Mayor’s London Resilience Unit is developing a food systems resilience partnership with the charity Sustain. That work is aimed at “international shocks and crises”; whether it addresses the digital systems that move food is not stated.
Where does AI sit in London’s risk register?
Siegle reports critics’ complaint that the capital’s risk register contains no AI scenario of its own, and that its safeguards are arranged risk by risk: cyberattack, loss of power, disruption to transport. The register’s text is consistent with that. The most recent version we could find on london.gov.uk, Version 14 dated January 2025, mentions AI only in an appendix of chronic risks taken from the National Security Risk Assessment (2023). That entry warns that as AI is deployed in more and more settings, the scope for harm “magnifies substantially”.
The register describes chronic risks as continuous challenges whose responses “tend to be developed through strategic, operational or policy changes”, not through emergency planning. AI therefore sits among the risks the register expects policy, not emergency response, to address. Cyber-attacks, by contrast, are listed by what they hit, with separate entries for the health and care sector, the transport sector and government systems.
The cascade nobody owns
Read against the register, the gap sits between those two halves. The acute risks are listed one at a time. Siegle’s worry is a single fault that AI carries from one system into several others simultaneously, which a risk-by-risk list is not designed around. Her question is who would hold that overview. The article leaves it unanswered.
Reality Check: According to the document, the London Risk Advisory Group produces a new edition every year. Version 14 is dated January 2025, so on that cycle the published picture of London’s risks is now more than 20 months old.
What to watch
Three signals would show whether London is treating AI dependency as a resilience question.
- The next London Risk Register. A version that moves AI from the chronic-risk appendix into an acute scenario, or that models a failure across several sectors at once, would mean the gap Siegle describes has been taken up. Another version that keeps AI only in the chronic-risk appendix would mean it has not.
- The London Food Systems Resilience Partnership. If its published work treats routing and delivery software as a point of failure alongside international shocks, the city is planning for the dependency. If it addresses only supply origins, it is not.
- Croydon’s continuing static deployments. Police say the static deployments will continue. Published error and challenge figures from later deployments would show whether accountability is keeping pace with scale, in the way Blackwell describes.
Siegle expects no single off switch to solve this, and McEwen doubts that deeply integrated AI can always just be switched off. What the sources point to instead is each operator knowing what its service looks like when the AI is gone, and someone at city level asking what happens when several are gone at once. Planning for their failure is a question for the city as a whole.
Source: London and the AI apocalypse: Is the capital becoming dangerously dependent on artificial intelligence? (The Standard, 2026), by Lucy Siegle. Additional sources: Thames Water deploys AI, satellites and 75,000 acoustic sensors (New Civil Engineer, 2026); Emergency department clinicians able to see more patients, thanks to new AI tech (St George’s University Hospitals NHS Foundation Trust); Lewisham uses AI to identify more than 3,000 potential housing sites (Think Digital Partners, 2026); Croydon facial recognition arrests (Silicon UK, 2026); How resilient is London’s food network? (London Assembly); London Risk Register, Version 14 (London Resilience Partnership, January 2025).
This strategic analysis was written by Resultsense, a UK-focused AI news and analysis publication. We will be watching the next London Risk Register for the first sign that the city plans for AI failing across sectors at once. Read more analysis at Insights, or get in touch.