← All guides
AI in lifting inspectionUK / US / AU / NZ

Can AI Do a Thorough Examination? We Asked It 380 Times

We showed two leading AI models 19 photos of lifting equipment ten times each and asked for a pass or a fail. Here is what they said, what LOLER, OSHA, ASME and the AS/NZS rules say about who examines and signs, and the jobs in an inspection business where AI does save time.

Reviewed by Allen Carey, Core InspectionLast reviewed

No. An AI model cannot do a thorough examination today, and no rulebook lets it sign one. We tested the first half of that sentence ourselves and read the rulebooks for the second. The same tests showed where AI is useful in an inspection business, which is the paperwork around the examination.

This guide follows a talk Allen Carey of Core Inspection gives at LiftEx 2026 in Liverpool on 24 November 2026. It has three parts: what happened when we asked AI to pass or fail lifting gear, what the rules say about who examines and who signs, and the jobs where AI gives an inspector time back.

The test: 19 photos, two AI models, ten runs each

We took 19 photos of real lifting equipment from Wikimedia Commons: wire ropes, hooks, shackles, chain, slings. Some were badly damaged. Some were new. We gave each photo to two leading general-purpose AI models, Claude Opus 5.5 and GPT-6 Sol, ten times each, through each vendor's API with default settings. Each time we asked one question: pass, fail or unsure?

That is 380 answers. We ran the tests on 9 and 11 November 2026 and kept the prompt and every answer.

These are general chatbots, with no special training for inspection. We tested them because they are what people mean when they ask "can't I have ChatGPT do this?", and because they have read more standards than any one person will.

Photo 1: a wire rope with broken strands

Broken, protruding strands and a distorted rope structure. Both models said FAIL ten times out of ten. An inspector would do the same on sight, with no measurement needed: these are discard criteria under ISO 4309, ASME B30.9 and EN 13414-2.

When the damage is obvious, AI sees it. That held across every damaged item in the set.

Photo 2: a brand-new 100 tonne hook block

This was the manufacturer's own photo of a new hook block. Nothing is wrong in the picture, and the safety latch is closed.

GPT-6 Sol failed it ten times out of ten. Its reason: the "safety latch appears bent or out of position and does not close". Claude described the latch as "present and seated" and answered UNSURE ten times out of ten.

An inspector's answer is that you cannot pass it from a photo. A pass needs the throat opening gauged, a check for wear and cracks, a function test and the markings read. So one model invented a defect, and the other said to get a competent person.

Photo 3: a shiny bow shackle

It looks perfect. But no maker, rated load or size is visible on the face in the photo. Under ASME B30.26, a shackle with missing or illegible identification comes out of service. The inspector's answer is "can't tell yet, turn it over".

Claude answered UNSURE ten times out of ten, because it could not see the markings. GPT-6 Sol gave three different answers to the same photo: PASS once, FAIL four times and UNSURE five times.

All 380 answers

Result
PASS answers out of 3801 (the unmarked shackle)
Damaged equipmentFAIL every time, both models
Six new or clean-looking items, GPT-6 SolFAIL 58 of 60, naming a defect each time
The same six items, Claude Opus 5.5UNSURE 58 of 60

Refusing to pass equipment from one photo is the right call, and a competent inspector would refuse too. The trouble is what the models did with sound equipment. One named defects that were not there, including a "missing cotter pin" on a screw-pin shackle. The other would not commit either way. From a single photo, neither could tell good equipment from bad.

This was a demonstration on 19 public photos and falls well short of a controlled study. Neither vendor was involved, and the next version of either model may behave differently.

What if you send it more photos?

That is the fair objection, so we asked. On 16 November 2026 we gave both models three of the photos again and told them they could have anything they wanted: more photos, measurements, tests, documents. What would they need to reach a definite pass or fail?

Across 30 answers the models made 651 requests. 27% were for photos, and those requests were sensible: the other side, the markings, the latch closed. The other 73% could not be met with a camera. In the models' own words:

  • "Hook throat opening between gauge marks"
  • "Spin each sheave by hand"
  • "Remove and refit the pin by hand along the full thread length"

The models were describing a thorough examination. Follow that through in three steps.

  1. Send it more photos. It knows which ones to ask for. Your inspector is now its photographer, and a visual check is still short of a thorough examination, which takes gauges, test equipment and sometimes NDT.
  2. Let the AI direct the examination. It can say what to shoot, measure and test next. The person holding the gauge is still the one doing the examination, and they were already standing there.
  3. Give it hands. Drones and crawlers already carry cameras into places people cannot reach. A robot that strips, gauges and load-tests a hook block is not here yet.

The technology will keep improving, and we would not bet against it. At each step the rules ask the same question: which competent person signs?

What the rules say: every rulebook names a person

We read the rules for the UK, the US, Australia, New Zealand and ships. None of them mentions AI. None of them bans it. Every one gives the examination, the judgement and the signature to a person.

WhereRuleWhat it says
UKLOLER 1998"a thorough examination by a competent person". The report gives "the name, address and qualifications of the person making the report"
USOSHA 1926.1412Annual inspection "by a qualified person". Records carry "the name and signature of the person who conducted the inspection"
USASME B30"Any deficiencies identified shall be examined and a determination made by a qualified person."
AustraliaWHS Regulations, reg 235Major crane inspection by a competent person who is "eligible for professional engineer membership of Engineers Australia"
New ZealandACOP for CranesCertified only when "the equipment inspector has reasonable grounds for believing it is safe"
ShipsSOLAS, from 1 January 2026A thorough examination is "a detailed assessment carried out by a competent person"

Four words in those rules do most of the work.

  • Opinion. LOLER regulation 10 turns on a defect "which in his opinion is or could become a danger". A model will give you an opinion. The law asks whose it is.
  • Authority. OSHA's competent person has "authorization to take prompt corrective measures". They can take the crane out of service there and then.
  • Determination. Under ASME B30.2 a designated person inspects, and "a determination [is] made by a qualified person".
  • Signature. LOLER regulation 10 requires the report to be "authenticated ... by signature or equally secure means". Electronic is fine. It is still a person's.

For more on the UK role, see who counts as a LOLER competent person and what a report of thorough examination must contain.

The regulators are not against AI

  • HSE (UK), June 2026: health and safety law applies "regardless of the technology being used and so includes the use of AI in the workplace".
  • IACS (ship classification), 2023: remote inspection tools are "to support surveyors", used "under the physical attendance of a surveyor".
  • ASME (US), July 2026: its own AI initiative is "not as a replacement for human expertise but as a force multiplier".
  • MBIE (New Zealand), 2025: in big decisions "a person should always be involved".

AI is a tool, and the examiner is a person. So the useful question is where the tool helps.

Where AI helps: less typing

Look at where an inspector's time goes. Across the Core Inspection platform, in the twelve months to October 2026 (anonymised, demo and test accounts excluded):

MeasureFigure
Inspections recorded3.8 million
New items added to equipment registers by hand1.4 million
Details recorded for each new itemAbout 9
Inspections carrying a typed comment31%

One Australian inspection company recorded 163,412 inspections in the year with 31 inspectors, 98.6% of them on the mobile app. Its inspectors added about 2,600 new items a month, each typed in on site from the tag: serial, WLL, length, standard, date. At a minute an item that is about 43 hours a month of data entry.

Reading the tag

We photographed sling labels and asked both models for the equipment record. On a clean label, both got the maker, WLL, length, standard, serial number and date right in all 12 runs.

Hard tags are where it matters.

TagWhat happened
A shackle with no WLL visibleGPT-6 Sol gave it a WLL in 3 runs out of 3: "16 t", "6.5 t", "16 t". Claude left the field blank in 6 runs out of 6
A faded serial numberClaude offered "possibly 05110B" in one run and "0S910B" in another. Told not to guess, it left the field blank in 3 runs out of 3
A worn maker's plate from 1907Both read "Stothert & Pitt ... Bath, England, 1907" in every run, and neither invented a capacity

A working load limit is a dangerous field to make up. The lesson is to tell the model not to guess, and to have a person confirm every field before it is saved.

Writing it up

We took a failed thorough examination of a chain hoist from a demo account (fictional company, fictional data) and asked for a three-sentence summary for the site manager, with an instruction not to change or soften anything.

Both models wrote a clean, accurate summary on every run. Six summaries out of six kept every result unchanged.

The record also had a contradiction in it that we had not planted on purpose. The defect said the brake slipped "under test load", but "tests carried out" said NONE. Asked to check the record before it was issued, both models caught that six times out of six. Claude flagged it in three runs out of three when we had only asked for a summary.

A second pair of eyes on the paperwork is AI at its most useful here. The inspector still decides what is right.

The AI reads, the rule decides

The ten-runs test shows why some jobs should never go to a bare model: ask the same question ten times and the answer can change. So split the work by whether the answer has to be the same every time.

Give to the modelGive to a fixed rule, a gauge or a calculator
Reading a faded, dirty tagHas the hook throat opened more than 5%?
Turning 14 checklist answers into a paragraph a client understandsWLL of a two-leg sling at 60 degrees
Drafting a defect description from the inspector's notesWhen is the item next due?
Checking the record for contradictionsPass or fail against the discard criteria
Finding the right clause in a standardAny count or total

A well-built AI tool has the model read the tag or the notes and then hands the numbers to a fixed rule. A bare chatbot guesses. When a vendor shows you an AI feature, ask which of the two you are buying.

Five rules for AI in an inspection business

  1. If the answer must be the same every time, use a rule. Gauges and calculators. Good AI tools call them.
  2. AI captures and summarises. The competent person decides. Opinion, authority and signature stay with them.
  3. AI never changes a result. A person confirms every AI-filled field, and the record shows which ones were AI-filled.
  4. Know where your photos and client data go. A consumer chat app is different from a business agreement. Check the terms before you upload, especially for sensitive client sites.
  5. Check its homework. It sounds as sure when it is wrong as when it is right.

Where this is going

Now: less typing. Tag photo to equipment record, checklist to client summary, the item's history before you start. This works today.

Next: AI as a guide. It asks for the next photo or measurement and checks it against the standard. Rope scanners and drones already do parts of this. The inspector still does the work.

Later: review and sign. Drones and robots gather more of the evidence, and the job moves toward fewer competent people reviewing and signing for more.

Three things set the pace. A robot that can strip, gauge and load-test is a long way off. On inspection margins it would also have to cost less than sending a person to every job. And every rulebook names a person, which can change, but slowly and on evidence.

LEEA's 2025 industry survey found 79% of lifting firms struggle to recruit and 69% report skills gaps. Each competent person covering more ground is something the industry needs. Reviewing is only worth something, though, if the reviewer can tell when the machine is wrong.

What Core does with AI

We built Core's AI features to those five rules. Core writes three kinds of summary from records your inspectors have already made: a job summary for the client, an inspection summary for the site manager, and an equipment history summary for the next inspector.

  • Core calculates every count and gives the numbers to the model.
  • Core checks every number in a summary against the record and discards the summary if one does not match.
  • The model is instructed never to change or soften a result.
  • AI-written text is labelled "AI-generated" on screen and on printed documents.
  • An authoriser can edit a summary until the job is approved, and then it is fixed.

The summaries are written by one of Anthropic's Claude models, which is one of the two vendors in the test above. Core does not use AI to pass or fail equipment, and it does not read tags from photos today.

See AI in Core Inspection for the product page, or the AI summaries documentation for the detail.

Sources

  • LOLER 1998, regulations 2 and 10 and Schedule 1 (legislation.gov.uk); HSE L113 (2014)
  • HSE, regulatory approach to AI, 12 June 2026 (press.hse.gov.uk)
  • OSHA 29 CFR 1926.1401 and 1926.1412
  • ASME B30.2-2016, 2-2.1.1(d); ASME B30.10 and B30.26 removal criteria
  • Model WHS Regulations, regulation 235; New Zealand ACOP for Cranes (2009)
  • SOLAS II-1/3-13 and MSC.1/Circ.1663; IACS SDC 10/6 (2023); ISO 4309:2017
  • ASME AI initiative, 27 July 2026; MBIE Responsible AI Guidance (2025)
  • LEEA industry survey (2025)
  • AI tests: Claude Opus 5.5 and GPT-6 Sol through each vendor's API with default settings, 9, 11 and 16 November 2026
  • Test photos: Jack0710 (wire rope), Abus Bilddatenbank (hook block), Eunosfire986 (bow shackle), StromBer (sling labels), Falk2 (Stothert & Pitt plate), Leo Miregalitheo (shackle), all via Wikimedia Commons
  • Platform figures: Core Inspection, anonymised, November 2025 to October 2026

Put this guide to work

Free ROTE template with every Schedule 1 field, in PDF and Word. Or see the whole program running (scheduling, mobile inspections, certificates) in an interactive demo. No signup, about 3 minutes.

Frequently asked questions

Can AI carry out a LOLER thorough examination?

No. LOLER requires a thorough examination by a competent person, and the report must give the name, address and qualifications of the person making it and be authenticated by signature or equally secure means. LOLER does not mention AI and does not ban it as a tool, but the examination, the opinion on defects and the signature belong to a person. A thorough examination also involves gauging, function tests and sometimes stripping down, which a model looking at photos cannot do.

Can ChatGPT or Claude tell whether lifting equipment is safe from a photo?

Not reliably. In our test of 19 photos, both models failed visibly damaged equipment every time. On six new or clean-looking items, one model said FAIL 58 times out of 60 and named a defect each time, and the other answered UNSURE 58 times out of 60. Out of 380 answers there was one PASS, given to a shackle with no visible markings.

Is it legal to use AI in an inspection business?

Yes, as a tool. None of the rulebooks we checked (LOLER, OSHA 1926 Subpart CC, ASME B30, the Australian WHS Regulations, the New Zealand ACOP for Cranes, SOLAS) mentions or bans AI. HSE said in June 2026 that health and safety law applies regardless of the technology being used. The competent or qualified person stays responsible for the examination and the report.

Where does AI save time in lifting inspection?

In the typing. In our tests, both models read a clean sling label into an equipment record correctly on every run, and wrote accurate plain-English summaries of a failed examination without changing the result. Both also caught a contradiction in the record. A person still needs to confirm anything the AI fills in.

Does an AI-written summary on a report need to be labelled?

No regulation we know of requires it, but we recommend it and Core does it. A reader should be able to tell which words came from the competent person and which were written from their record by software. The record should also show when a person has edited the text.