There is an ongoing debate in our profession about whether technology is going to make design worse: whether AI, and the younger professionals who reach for it first, will lower the bar rather than raise it. It is a good argument. It also, I've come to think, gets the diagnosis almost exactly backwards.
When two columns buckled on the 21st floor of the former Pfizer building at 235 East 42nd Street, New York, in July 2026 (floors sagging, neighbouring buildings evacuated, one of the city's largest office-to-residential conversions halted), a developer reached for a familiar phrase: a "typical construction mishap" [1][2][3]. Buckling primary columns and several sagging floors is not a mishap. It is a load-path failure. By mid-October 2026 the city had not published its findings [4], and I am not going to pre-judge it. What interests me is the question the incident forces into the open, at exactly the moment the profession is arguing about artificial intelligence: when something fails, where does the risk actually live?
Risk has never lived in the individual
Studying construction law taught me something the day job only hints at: almost none of the great design failures were about one person's incompetence. They were about systems: layers of checking, codified knowledge and assumption that everyone trusted, and that quietly let an error through.
Take MT Højgaard v E.ON, which reached the UK Supreme Court in 2017 [5]. Offshore wind turbine foundations failed, not because the engineers had been negligent, but because the internationally recognised design standard they had been required to use contained an error. The contractor had complied with that standard; the Supreme Court held it liable anyway: the contract also required a 20-year design life. The flaw lay upstream, in the industry's shared technical knowledge, invisible until experience exposed it. Or look at what the Grenfell Inquiry laid bare: not a single villain, but a certification and testing regime that had been gamed and worked around for years, generating paperwork that said "safe" while the reality drifted somewhere else entirely [6].
We are all fallible. The question that decides whether fallibility becomes catastrophe is not did someone make a mistake? It is why did the system fail to catch it? Every serious failure is, at root, an assurance failure. The person is where the error starts. The system is what was supposed to stop it.
And we have been relying on machines to do our thinking for half a century. In my experience, no serious structure since the 1970s has been designed without software running analysis that no human re-checks by hand. The dependency is not new. What is new is that we have finally started to notice it. The software we trusted for fifty years was deterministic, and our entire assurance model (independent design checks, Category 3 checking, the audit trail) was quietly built on that. It presumes the work can be re-run and interrogated:
- Run it again, you get the same answer.
- Trace the logic, you can follow it.
- Reproduce the check, you can prove it.
That is the real discontinuity with AI. Not that a machine is doing the cognition; it always has. It is that we are starting to rely on machines whose outputs are non-deterministic, non-reproducible and hard to interrogate. It is not the intelligence we should worry about. It is the loss of the audit trail.
A sign-off from someone who cannot actually check the work is not assurance.
Checker, not author
In my experience, design errors don't cluster in the logical, rule-bound parts of the work, the parts a machine handles beautifully. They cluster at the edges: in novel configurations, unstated assumptions, new storeys threaded through a sixty-five-year-old frame [7]. That is a problem of judgement and context, precisely where today's AI is weakest and an experienced human is strongest. Used well, the machine is at its most valuable as the checker, catching the human slip the current system misses.
There is an obvious objection. If the machine's output cannot be reproduced or interrogated, how can I hand it the job of checking? The answer is that the checker never holds the assurance. It signs nothing, and its verdict is not the audit trail. Its only job is to raise a flag, to say look here again, and a human still has to go and look, reproduce the concern by hand, and decide. Used that way, the non-determinism stops being a liability: a false flag costs a second look, a true flag catches a slip that would otherwise have shipped. The danger was never the machine that raises a question. It is the machine whose answer we sign without re-deriving it.
Which exposes the catch in the phrase everyone reaches for. "Human in the loop" is offered as though the human were the safeguard. But a human in the loop is only as good as the human. Put a junior who trusts the confident output, or a senior too rushed to re-derive it, and the loop launders the error, stamping a plausible wrong answer with an authority it never earned. The loop is not the control. The competence of the person in it is. A sign-off from someone who cannot actually check the work is not assurance. It is theatre with a name on it.
The same test applies to every layer of governance. Some of it is genuine defence-in-depth; some of it generates a certificate without generating safety. You usually cannot tell which until the day it fails, so the right question is whether a layer actually detects the thing it exists to catch. Most of our layers were designed to catch human error. Very few can yet see a machine error: a plausible, confident, wrong answer with no working to check.
The real regulator isn't who you think
In the UK an engineer needs no licence to practise [8], and no regulator requires professional indemnity insurance. The Engineering Council licenses the institutions that assess candidates [8]; its guidance asks only that registrants hold cover "if appropriate" [9], and the IStructE's code only requires a member to tell a client when cover is absent or inadequate [10]. Architects are different: the ARB expects registered architects to hold adequate and appropriate cover [11]. For engineers the practical obligation comes from contract: in my experience appointments routinely require £2m, £5m or £10m of cover.
Which means the entities actually governing how much risk an engineer may take are clients and insurers. Not the ICE. Not the IStructE. The market.
And the market is already moving. The Building Safety Act retrospectively extended the limitation period for defective-dwelling claims to thirty years [12], on top of a PI market where cladding and fire-safety cover was already hard to buy [11]. US insurers have begun writing AI exclusions into liability policies, E&O included [13]; some Lloyd's underwriters are instead backing affirmative AI cover [14]. The interesting part is what underwriters want to see: documented AI governance, human oversight of outputs and formal approval protocols. Firms that can show them are better placed on cover and terms [13]. Whoever writes that underwriting question is writing the real competence standard for AI in design. Right now, in the UK, that is the insurer, because in my view the profession vacated the field.
So: man or machine?
It is the wrong question. If risk lives in the system, a contest between the human and the machine re-personalises the very thing we should be de-personalising. The future is human judgement and machine capability, and the whole game is building the assurance around them that makes the pairing trustworthy.
But there is one thing the machine can never do. It cannot hold professional indemnity. It cannot be struck off. It cannot sit in the witness box and be asked why. However much of the calculation we hand across, the accountability does not move. It comes to rest, as it always has, on one person and one stamp.
Hold both of those at once, because they do not contradict. Cause is systemic: distributed across standards, checks, assumptions and the layers meant to catch each other. Answerability is singular by design: the law, the stamp and the witness box deliberately anchor it to one person, precisely so the system cannot diffuse responsibility into a fog where no one answers for anything. The stamp is not where the risk lives. It is where the buck stops.
Delegate the arithmetic. Never delegate the responsibility. The institutions' job is to become the certifiers of the systems, the competence and the governance that let human and machine work together safely; at Synaform, that is the ground our Advice work covers. As for whether we'll make mistakes along the way: of course we will. Bring me the engineer who never has, and I'll give you a pound.
The question was never man or machine. It was whether the system catches the error before the error catches us. That part is still ours.
Sources
- [1] Ashley Southall and Mihir Zaveri, "Columns in Manhattan high-rise buckle, prompting evacuations", New York Times News Service, published in the Hawaii Tribune-Herald, 8 July 2026. Two columns buckled, floors sagging between the 21st and 26th floors, the developers' "largest office-to-residential conversion in city history", and Nathan Berman (MetroLoft) calling it a "typical construction mishap". https://www.hawaiitribune-herald.com/2026/07/08/nation-world-news/columns-in-manhattan-high-rise-buckle-prompting-evacuations/
- [2] "Pfizer's former headquarters in Midtown is stable after buckled columns led to evacuations", 6sqft, 7 July 2026 (updated 8 July 2026). Buckled 21st-floor columns, evacuation orders on neighbouring buildings, collapse zone. https://www.6sqft.com/pfizers-former-headquarters-in-midtown-at-risk-of-collapse/
- [3] "Buckled columns halt work at 37-story Manhattan office-to-residential conversion", New York Construction Report, 7 July 2026. https://www.newyorkconstructionreport.com/buckled-columns-halt-work-at-37-story-manhattan-office-to-residential-conversion/
- [4] "Inside the plan to fix the nearly collapsed Pfizer Building, 1/64th of an inch at a time", Gothamist, 16 October 2026. States that the Department of Buildings has not yet released the results of its investigation into the cause. https://gothamist.com/news/inside-the-plan-to-fix-the-nearly-collapsed-pfizer-building-164th-of-an-inch-at-a-time
- [5] MT Højgaard A/S v E.ON Climate & Renewables UK Robin Rigg East Ltd [2017] UKSC 59, judgment of 3 August 2017, paras 21-25 (error in the J101 value for δ, "wrong by a factor of about ten"; trial judge rejected negligence; liability founded on the requirement that the foundations be designed for "a lifetime of 20 years"). Case page: https://www.supremecourt.uk/cases/uksc-2015-0115.html; judgment: https://supremecourt.uk/uploads/uksc_2015_0115_judgment_8ddc6bff9a.pdf
- [6] Grenfell Tower Inquiry, Phase 2 Report, Volume 1 (HC 19-I, September 2024), Chapter 2 Executive summary, paras 2.4 ("decades of failure"), 2.19 (manufacturers' "deliberate and sustained strategies to manipulate the testing processes, misrepresent test data and mislead the market") and 2.20 (failures of the BBA, LABC and UKAS). https://assets.publishing.service.gov.uk/media/66d817aa701781e1b341dbd3/CCS0923434692-004_GTI_Phase_2_Volume_1_BOOKMARKED.pdf
- [7] Council on Tall Buildings and Urban Habitat (CTBUH), Skyscraper Center, "Pfizer Building": completed 1961; renovated and replaced by 235 East 42nd Street. https://www.skyscrapercenter.com/building/pfizer-building/11898
- [8] Engineering Council, "Professional registration" FAQs: "Engineers do not require a licence to practise in the UK"; the Engineering Council has licensed professional engineering institutions to assess candidates for registration. https://www.engc.org.uk/glossary-faqs/frequently-asked-questions/professional-registration/
- [9] Engineering Council, Guidance for Institution Codes of Conduct (copy dated January 2017), item 14: "Assess relevant liability, and if appropriate hold professional indemnity insurance." Copy consulted as hosted by CIPHE, a licensed institution: https://www.ciphe.org.uk/globalassets/media/pdfs/for-the-professional/ec-code.pdf
- [10] Institution of Structural Engineers, Code of Conduct (January 2022), Guidance Notes 5.1-5.2 (Insurance). https://www.istructe.org/getattachment/About-us/istructe-code-of-conduct/IStructE-Code-of-Conduct-(2022).pdf?lang=en-GB
- [11] Architects Registration Board, Professional Indemnity Insurance: supporting guidance for the Architects Code of Conduct and Practice, published 19 June 2025: Standard 4 met when architects "ensure their liabilities are covered by adequate and appropriate professional indemnity insurance"; section "Fire safety and cladding claims" records that such cover has become "increasingly difficult" to secure. https://arb.org.uk/wp-content/uploads/Guidance-document-Professional-Indemnity-Insurance.pdf
- [12] Building Safety Act 2022, s.135, inserting s.4B into the Limitation Act 1980 (15 years for Defective Premises Act 1972 s.1 claims; 30 years for claims that accrued before commencement, "treated as always having been in force"). https://www.legislation.gov.uk/ukpga/2022/30/section/135
- [13] Mark Rosanes, "ISO's generative AI exclusion is already on thousands of CGL policies", Insurance Business (US), 16 September 2026. ISO CG 40 47 generative AI exclusion; W.R. Berkley absolute AI exclusion across D&O, E&O and fiduciary products; carriers seeking documented AI governance, human oversight and formal approval protocols. https://www.insurancebusinessmag.com/us/news/cyber/isos-generative-ai-exclusion-is-already-on-thousands-of-cgl-policies-589971.aspx
- [14] Armilla Insurance Services, "Armilla launches affirmative AI liability insurance with Lloyd's underwriter Chaucer", 30 April 2025 (company press release; Armilla is a Lloyd's coverholder). https://www.armilla.ai/resources/armilla-launches-affirmative-ai-liability-insurance-with-lloyds-underwriter-chaucer