[Engineering](https://bfcs.ai/industry/engineering/) · [Safer](https://bfcs.ai/safer/)

# Can AI Make Structural Reviews Safer?

Structural failures in Australia and abroad have been traced to calculation errors and to reviews that never looked at the right element. AI can screen inspection images for cracks the same way every time, where human ratings are known to vary, but the failures that matter most were not missed for lack of a flag. The sign-off, and the duty behind it, stays with a registered engineer.

Updated **October 2026** · 6 min read · Evidence-graded · 8 sources

## The BFCS verdict

Every article scored on the same four dimensions

| Dimension | Verdict | Why |
| --- | --- | --- |
| Better | Partly · Moderate confidence | Consistent at spotting surface damage; blind to calculations · Image models classify cracks with high accuracy on curated datasets, and visual inspection by people is known to vary. But both failures examined here involved under-designed elements, which no camera can see. |
| Faster | Partly · Low confidence | Triage of images is quicker; the judgement is not · Screening thousands of inspection photos is the obvious gain. We found no independent field study measuring how much review time it saves once an engineer checks every flag. |
| Cheaper | Partly · Low confidence | Defects are costly; the return on AI review is unmeasured · Serious defects are common and expensive to fix, so earlier detection has obvious value. We found no independent evidence that AI-assisted structural review lowers rectification costs. |
| Safer | Conditional · Governance-gated | A second pair of eyes, never the signature · AI helps only if it widens what a review covers and every flag reaches an engineer with authority to act. Registration law in NSW and Queensland places the responsibility on a person, and nothing in it transfers to a tool. |

## The problem, in numbers

Structural defects in Australian apartments are not rare. A 2025 survey of strata managers covering 520 NSW apartment buildings registered between mid-2018 and mid-2024 found 53% had a serious defect in common property within six years of construction, and 18% had a serious structural defect in foundations, floors, walls, roofs, columns or beams. (Source 1) The regulator itself cautions that newer buildings may simply not have shown their defects yet. (Source 1)

The national diagnosis came earlier. The 2018 Building Confidence report found that three jurisdictions did not register structural engineers, and that building surveyors often accept an engineer's self-certification, so "large parts of the design are not substantively reviewed by another qualified practitioner". (Source 2) Its Recommendation 17 called for genuine independent third party review of specified design components or building types. (Source 2) When the Opal Tower cracked at Christmas 2018, investigators found hob beam assemblies under-designed against the National Construction Code and the concrete standard, and recommended independent checking of critical elements by a registered engineer. (Source 3)

### The evidence, in four numbers

53% · of recent NSW apartment buildings with serious defects · 520 Class 2 buildings registered July 2018 to June 2024, as reported by strata managers (Building Commission NSW, 2025)

18% · of the same buildings with a serious structural defect · foundations, floors, walls, roofs, columns or beams; against 15% of 642 buildings in 2023

68% · of bridge condition ratings predicted within one point of the average · 49 US state bridge inspectors; ratings ranged across up to five values (FHWA, 2001)

Nearly 2x · load demand on the failed node, recalculated after the collapse, versus the design figure · Florida International University bridge collapse, 2018; the peer review missed it (NTSB)

## What improves, and why

A structural review means two different jobs. Design review checks calculations and drawings before anything is built. Condition review inspects what has been built, from a new tower to an ageing bridge. AI offers something to each, but on different terms.

Condition review is where AI has a real case, because human inspection varies. In a US federal study involving 49 inspectors from 25 state agencies, condition ratings spread across up to five values, and only 68% were predicted to fall within one point of the average. (Source 5) Visual inspection alone was judged unlikely to find the specific defects an in-depth inspection is meant to target. (Source 5) A model that screens every drone or camera image the same way, at three in the afternoon as at nine in the morning, addresses that variability directly.

Design review is different. An AI tool can compare one element's loads against similar elements, or flag a connection whose capacity looks out of line. In our reading, that kind of anomaly flag is promising, but we found no peer-reviewed study measuring whether it catches errors that independent human checking misses.

> At Miami, the cracks were seen and photographed. What failed was the calculation behind them, and the judgement about what they meant.

## What the evidence actually says

Crack detection on curated images is well studied. On SDNET2018, a public dataset of 56,092 labelled concrete images, a baseline model classified bridge deck images with 90.45% to 91.92% accuracy. (Source 6) That figure needs care. By our arithmetic about 85% of the deck images contain no crack. If the test images share that mix, which the paper does not state, a model that never flagged anything would already score about 85%, and the study reports no figure for the share of cracks missed. (Source 6) For safety, misses are the number that matters.

The failure record points somewhere else. In March 2018 a partly built pedestrian bridge at Florida International University collapsed onto a live road, killing six people. (Source 4) The US National Transportation Safety Board found the probable cause was load and capacity calculation errors in one node: post-collapse calculations put the demand on it at nearly twice the design figure. (Source 4) An independent peer review did take place, using finite element software. It modelled the completed structure only, not the construction stages or the nodal connections, and the reviewer told investigators the budget and time for that analysis had not been agreed with the designer. (Source 4) Cracking at the node was observed before the collapse; the board found the engineer of record failed to recognise its significance. (Source 4)

In our assessment, that sequence is the core lesson for AI. Spotting the cracks was not the gap. A crack-finding model would have added another photograph. What was missing was a review scoped to the critical element, a correct reading of the cracking, and a decision by those with authority to stop work and close the road. An AI review tool inherits the same weakness as any reviewer: it checks only what it is pointed at.

## Where it falls short

-   **It sees surfaces, not sums.** Both the Miami and Opal investigations found under-designed elements. (Source 3), (Source 4) Image models inspect what is built; they cannot see a load factor applied wrongly in a calculation.
    
-   **Accuracy hides misses.** When most images show no crack, headline accuracy can look high while real cracks go unflagged. (Source 6) Ask for the miss rate on your own structures.
    
-   **Scope blindness.** A review covers only what it is scoped to cover. (Source 4) A tool that reports "no anomalies" on the elements it was given says nothing about the ones it was not.
    

## Where governance decides the outcome

The law already decides who signs. In NSW, a registered design practitioner must declare that regulated designs for apartment and similar buildings comply with the Building Code before work starts, and any varied design needs a fresh declaration. (Source 7) In Queensland, only a registered practising professional engineer may carry out professional engineering services, or someone working under their direct supervision, which means directing, overseeing and evaluating the work. (Source 7) In our reading, a tool cannot hold that registration, so responsibility for using its output stays with the engineer. Engineers Australia has said those implementing AI must hold themselves accountable, and that developers cannot hide behind disclaimers such as "use with caution". (Source 8) Keep the independence the Building Confidence report asked for: (Source 2) an AI check run by the designer is still self-checking.

## How to think about adoption

Use AI where the evidence supports it: screening inspection imagery for cracks so engineers spend their time on the flagged areas, with every image retained for audit. Before trusting a model on your structures, test it against defects your engineers have already confirmed and record what it misses, not just its accuracy. For design review, treat anomaly flags as prompts for a human check, never as clearance. Write down which critical elements each review covers, so a "pass" cannot be read as wider than it is. Route every flag to someone with authority to stop work, and keep the registered engineer's signature, and the independence of the reviewer, exactly where the law puts them.

## How we read this evidence

BFCS.ai does not fabricate figures. Failure investigations, defect surveys and AI detection studies answer different questions and are not pooled here. We found no independent study measuring whether AI-assisted structural review reduces failures or defects; where that is the case we say so.

## Sources

1.  Building Commission NSW, "2025 Strata Defects Research Report" (Research Report, November 2025; produced with Strata Community Association (NSW) and NSW Fair Trading). [nsw.gov.au](https://www.nsw.gov.au/departments-and-agencies/building-commission/building-and-construction-resources/2025-strata-defects-research-report)
    
    **What we took from it:** Survey of strata managers covering 520 Class 2 strata buildings of four or more storeys registered 1 July 2018 to 30 June 2024 (42% response rate, margin of error plus or minus 3.3% at 95% confidence). 53% of buildings had serious defects in common property in the six years after construction; structural defects (foundations and footings, floors, walls, roofs, columns and beams) were reported for 18% of all buildings in 2025 against 15% of 642 buildings in 2023. Where strata managers knew the accumulated costs of serious defects, the average was $215,400 per building. The report cautions that lower defect rates in newer buildings may reflect a lag before defects are detected.
    
2.  Shergold, P. and Weir, B., "Building Confidence: Improving the effectiveness of compliance and enforcement systems for the building and construction industry across Australia," expert assessment for the Building Ministers' Forum, February 2018 (published April 2018; now hosted by the Treasury). [treasury.gov.au](https://treasury.gov.au/publication/c2018-682884)
    
    **What we took from it:** Notes that three jurisdictions did not register structural engineers; that self-certification is often accepted, so "large parts of the design are not substantively reviewed by another qualified practitioner"; and Recommendation 17, that each jurisdiction require genuine independent third party review for specified components of designs and/or certain types of buildings.
    
3.  Carter, J., Hoffman, M. and Foster, S., "Opal Tower Investigation: Final Report," independent advice to the NSW Minister for Planning and Housing, 19 February 2019. [flatchat.com.au](https://www.flatchat.com.au/wp-content/uploads/2019/02/OPAL-TOWER-INVESTIGATION-FINAL-REPORT-Rev0-1.pdf)
    
    **What we took from it:** Findings include that the as-constructed hob beam and panel assembly was under-designed under the National Construction Code and AS3600 at a number of locations; recommendations include independent third party checking and certification of the design of critical elements by a Registered Engineer. We found no copy on an official NSW Government site; this copy is hosted by Flat Chat.
    
4.  US National Transportation Safety Board, "Pedestrian Bridge Collapse Over SW 8th Street, Miami, Florida, March 15, 2018," Highway Accident Report NTSB/HAR-19/02, adopted 22 October 2019. [ntsb.gov](https://www.ntsb.gov/investigations/AccidentReports/Reports/HAR1902.pdf)
    
    **What we took from it:** One bridge worker and five vehicle occupants died. Probable cause: load and capacity calculation errors by the designer in the 11/12 nodal region; contributing: an inadequate peer review that failed to detect them, and the engineer of record's failure to identify the significance of the cracking observed before the collapse. Post-collapse calculation found demand on the node "nearly twice" the design figure. The peer reviewer modelled the completed structure only, not construction stages or the nodal connections, and told investigators the budget and time for that analysis were not agreed.
    
5.  US Federal Highway Administration, "Reliability of Visual Inspection for Highway Bridges," FHWA-RD-01-020, June 2001. [fhwa.dot.gov](https://www.fhwa.dot.gov/publications/research/nde/01020.cfm)
    
    **What we took from it:** 49 inspectors from 25 state agencies; Routine Inspection condition ratings varied across up to five different ratings, with only 68% predicted to fall within one point of the average and 95% within two; In-Depth Inspections using visual inspection alone are "not likely to detect or identify the specific types of defects."
    
6.  Dorafshan, S., Thomas, R. J. and Maguire, M., "SDNET2018: An annotated image dataset for non-contact concrete crack detection using deep convolutional neural networks," Data in Brief 21, 2018, pp. 1664-1668 56,092 labelled images of bridge decks, walls and pavements; the bridge deck subset holds 2,025 cracked and 11,595 uncracked images. [pmc.ncbi.nlm.nih.gov](https://pmc.ncbi.nlm.nih.gov/articles/PMC6247444/)
    
    **What we took from it:** Baseline AlexNet accuracy was 90.45% (trained from scratch) and 91.92% (transfer learning) on bridge decks, 87.54% and 89.31% on walls, and 94.86% and 95.52% on pavements. Only accuracy is reported, not the share of cracks missed, and the class mix of the test sets is not stated. The 85% uncracked share of deck images (11,595 of 13,620) is our arithmetic.
    
7.  Practitioner law. [nsw.gov.au](https://www.nsw.gov.au/housing-and-construction/compliance-and-regulation/professionals-working-on-regulated-buildings/design-and-building-practitioners/design-obligations)
    
    Queensland, Professional Engineers Act 2002, section 115 (current consolidation): [legislation.qld.gov.au](https://www.legislation.qld.gov.au/view/whole/html/inforce/current/act-2002-054)
    
    **What we took from it:** NSW Government, "Design practitioner obligations when working on regulated buildings" (last updated 8 September 2026): for Class 2, 3 and 9c buildings a registered design practitioner makes the design compliance declaration, regulated designs must be lodged before work starts, and varied designs need a new declaration lodged within one day.
    
    a person who is not a practising professional engineer must not carry out professional engineering services (maximum penalty 1,000 penalty units) unless under the direct supervision of one who is responsible for the services, which requires that engineer to direct, oversee and evaluate the work.
    
8.  Engineers Australia, "Responsible AI in Australia: An engineering perspective," submission to the Safe and Responsible AI in Australia discussion paper, July 2023. [engineersaustralia.org.au](https://www.engineersaustralia.org.au/sites/default/files/2023-09/Responsible-AI-in-Australia-an%20engineering-perspective.pdf)
    
    **What we took from it:** Engineers, developers and companies implementing AI should "hold themselves accountable for their work", and those developing AI cannot rely on disclaimers such as "use with caution"; regulation should apply where the technology is high-risk and requires public protection.
    

Putting this to work

## Working through this in your own organisation?

BFCS.ai is published by Inhouse CX. We help small and mid-sized organisations decide where AI earns its place, and put the governance in place that makes it safe to use.

[Talk to Inhouse CX →](https://inhouse.cx/engage/)

Continue the thread

- [Can AI Reduce Engineering Errors?](https://bfcs.ai/articles/can-ai-reduce-engineering-errors/): More in Engineering · ENGINEERING · BETTER · 3 min · Read →
- [Can AI Accelerate Design Cycles?](https://bfcs.ai/articles/can-ai-accelerate-design-cycles/): More in Engineering · ENGINEERING · FASTER · 7 min · Read →
- [Can AI Make Lending Fairer?](https://bfcs.ai/articles/can-ai-make-lending-fairer/): Also Safer · FINANCE · SAFER · 7 min · Read →

---

Canonical page: https://bfcs.ai/articles/can-ai-make-structural-reviews-safer/
This is the Markdown version of the page, generated from its HTML at build time.
Site guide for agents: https://bfcs.ai/llms.txt · Every page: https://bfcs.ai/sitemap.xml