[Manufacturing](https://bfcs.ai/industry/manufacturing/) · [Safer](https://bfcs.ai/safer/)

# Can AI Make Factories Safer?

Factories already have mature safety engineering: guards, interlocks and certified stop circuits. AI arrives as a new layer on top, watching for people near machines and checking isolation steps. The question is whether a probabilistic model belongs in that chain, and the evidence that it reduces injuries is still thin.

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 | Adds a watching layer; cannot replace certified guarding · Vision models can flag a person inside a robot cell or beside a forklift route, but occlusion and lighting can defeat them in exactly the places guards are meant to cover. |
| Faster | Partly · Moderate confidence | Alerts are immediate; the response is not · A warning reaches a supervisor in seconds rather than in a monthly audit. Certified guards already trigger the stop automatically, so the speed gain is in oversight, not in the stop itself. |
| Cheaper | Partly · Low confidence | No measured return on AI monitoring yet · Robot automation is associated with fewer injuries, but we found no independent evidence that AI monitoring layered on existing lines pays for itself in avoided claims. |
| Safer | Conditional · Governance-gated | Removal works; watching is unproven · Taking people away from hazards has measurable benefit. Monitoring people near hazards has almost no outcome evidence, and it raises surveillance duties that Australian law already regulates. |

## The problem, in numbers

Manufacturing does not kill as many Australians as transport or construction, but it hurts a great many. In 2024-25 the sector lodged 14,200 serious workers' compensation claims, 9.6% of the national total, at a frequency of 9.4 claims per million hours worked against 6.7 across all industries. (Source 1) It recorded 15 worker deaths in 2025, in a year when Australia's total of 222 was the highest since 2012. (Source 1)

The machine hazards are familiar. Nationally, being hit by moving objects caused 7% of worker fatalities in 2025 and being trapped between stationary and moving objects another 5%. (Source 1) In the United States, the regulator's citation list tells the same story: failure to control hazardous energy (lockout/tagout) is the third most cited standard in OSHA's preliminary 2026 list, powered industrial trucks seventh and machine guarding tenth. (Source 2) These are not exotic failures. They are known controls found missing on inspection, the same gaps that leave someone exposed when they reach into a machine.

### The evidence, in four numbers

14,200 · serious claims in Australian manufacturing · 2024-25 preliminary, 9.6% of the national total (Safe Work Australia)

9.4 vs 6.7 · serious claims per million hours worked · manufacturing vs all industries, 2024-25 preliminary (Safe Work Australia)

3rd · lockout/tagout among OSHA's most cited standards · FY2026 preliminary list; machine guarding is 10th

2 of 1,255 · AI safety studies that measured injury outcomes · records identified 2018 to 2024; neither rated high quality (Jetha et al., 2025)

## What improves, and why

Factories differ from building sites and mines in one important way: they already have engineered safety. Fixed guards, interlocked gates, light curtains and emergency stops are designed to functional-safety standards such as ISO 13849 and IEC 62061, which grade each safety function against a required performance or integrity level. AI does not start from zero here. It arrives as a layer on top.

That layer does three useful things. Cameras can track people and forklifts across a floor and warn when the two come too close, which fixed guarding cannot do on an open aisle. Vision can check that a lock and tag are on the right isolator before a maintenance task starts, turning a paper permit into an observed step. And continuous video surfaces near misses that would never have been reported, giving safety teams a picture of where risk actually concentrates.

The maintenance moment matters most. A NIOSH study of 41 robot-related deaths in the United States from 1992 to 2017 found 83% involved stationary robots and 78% a robot striking the worker under its own power, many of them while the robot was being maintained. (Source 3) That is exactly where guarding is opened and lockout is meant to take over.

> In a factory the guard is already certified. The real question is whether a model that is right most of the time should ever sit in a circuit that must be right every time.

## What the evidence actually says

The strongest evidence is not about AI watching people. It is about robots replacing them. A peer-reviewed US study found that a one standard deviation increase in local robot exposure reduced work-related injury rates by about 1.2 cases per 100 workers a year. (Source 4) That is the mining lesson again: removing people from hazardous tasks works. The same study found robot exposure was also linked to more drug- or alcohol-related deaths and mental health problems, a reminder that safety gains can carry costs elsewhere. (Source 4)

For AI monitoring itself, the outcome evidence is close to empty. A 2025 systematic review identified 1,255 records on AI tools for workplace health and safety published from 2018 to 2024; only two studies met its criteria for measuring an effect on injury or illness, and neither was rated high quality. (Source 5) The authors concluded it may be premature to recommend widespread use. (Source 5) Detection accuracy is well studied. Fewer injuries are not. In our reading, the incident reductions claimed for factory monitoring outside peer review are mostly self-reported, lack a comparison site and are chosen for publication. They may prove right, but they are not yet evidence that a plant with AI monitoring hurts fewer people than one without it.

Engineering research also explains why vision should not stand alone. In one human-robot collaboration study, the authors noted that the camera channel can fail through occlusion and be affected by lighting, and they judged neither vision nor touch sensing sufficient on its own for intrinsically safe operation. (Source 6)

## Where it falls short

-   **Not a safety function by default.** A model that learns and changes is hard to verify to a fixed reliability level. Europe's new Machinery Regulation requires third-party assessment for safety components with self-evolving machine-learning behaviour, applying from January 2027. (Source 7)
    
-   **Blind where it counts.** Occlusion and lighting can defeat vision, (Source 6) and cluttered cells, where people reach into machines, are where that is most likely.
    
-   **A warning is not a lockout.** An alert that a lock is missing protects no one unless the work stops. Without that link the system documents a hazard rather than removing it.
    

## Where governance decides the outcome

Two lines need drawing before deployment. The first is engineering: AI stays advisory unless the safety function has been designed and verified to the required level, and certified guarding is never downgraded because a camera is watching. The second is people. Continuous video of workers is workplace surveillance. In NSW it requires written notice at least 14 days before it starts, signage, and a magistrate's authority for anything covert, and Victoria's government has backed in principle 15 of 18 recommendations from a parliamentary inquiry into the issue. (Source 8) Decide in writing, with workers and their representatives, whether footage is used to fix hazards or to discipline people. A system seen as a disciplinary tool will be worked around.

## How to think about adoption

Start with the hazards the data already names: hazardous energy during maintenance, people and forklifts sharing aisles, and access into robot cells. Where a task can be automated or redesigned so nobody needs to be near the hazard, do that first; it is the only approach with outcome evidence behind it. Use AI monitoring as a second line that strengthens existing controls, wire every alert to an action that stops work, and keep certified guarding in place. Agree the surveillance terms before installing a camera. Then measure your own lost-time injuries and near-miss trends before and after, and treat any vendor's incident-reduction claim as a hypothesis to test.

## How we read this evidence

BFCS.ai does not fabricate figures. Evidence on robot automation and evidence on AI monitoring are kept separate here because they make different safety claims. No independent study we found measures injury outcomes from AI vision monitoring in factories; where that is the case we say so.

## Sources

1.  Safe Work Australia, "Key Work Health and Safety Statistics Australia 2026" (2025 fatality data; 2024-25 preliminary claims data). [data.safeworkaustralia.gov.au](https://data.safeworkaustralia.gov.au/insights/key-whs-statistics-australia/latest-release)
    
    **What we took from it:** 222 worker fatalities in 2025, the highest since 2012; manufacturing recorded 15 fatalities (1.7 per 100,000 workers), 14,200 serious claims (9.6% of the total) and a frequency rate of 9.4 serious claims per million hours worked against 6.7 for all industries; being hit by moving objects caused 7% and being trapped between stationary and moving objects 5% of worker fatalities.
    
2.  US Occupational Safety and Health Administration, "Top 10 Most Frequently Cited Standards," FY2026 preliminary list (1 October 2025 to 31 August 2026). [osha.gov](https://www.osha.gov/top10citedstandards)
    
    **What we took from it:** Control of hazardous energy (lockout/tagout, 29 CFR 1910.147) ranks 3rd, powered industrial trucks (1910.178) 7th and machine guarding (1910.212) 10th.
    
3.  Layne, L. A., "Robot-related fatalities at work in the United States, 1992-2017," American Journal of Industrial Medicine 66(6), 2023, pp. 454-461 (peer-reviewed, NIOSH). [stacks.cdc.gov](https://stacks.cdc.gov/view/cdc/230667)
    
    **What we took from it:** 41 robot-related fatalities identified in Census of Fatal Occupational Injuries research files; 83% involved stationary robots and 78% involved a robot striking the worker while operating under its own power, many during maintenance.
    
4.  Gihleb, R., Giuntella, O., Stella, L. & Wang, T., "Industrial Robots, Workers' Safety, and Health," Labour Economics 78, 2022, article 102205 (peer-reviewed). [publires.unicatt.it](https://publires.unicatt.it/it/publications/industrial-robots-workers-safety-and-health/)
    
    **What we took from it:** A one standard deviation increase in US commuting-zone robot exposure reduces work-related annual injury rates by about 1.2 cases per 100 workers; robot exposure was also linked to more drug- or alcohol-related deaths and mental health problems.
    
5.  Jetha, A., Bakhtari, H., Irvin, E. et al., "Do occupational health and safety tools that utilize artificial intelligence have a measurable impact on worker injury or illness? Findings from a systematic review," Systematic Reviews 14:146, 2025 (peer-reviewed). [pmc.ncbi.nlm.nih.gov](https://pmc.ncbi.nlm.nih.gov/articles/PMC12247322/)
    
    **What we took from it:** Of 1,255 records published January 2018 to April 2024, two studies met eligibility criteria, one of medium and one of low quality; the authors conclude it may be premature to recommend widespread use of AI for workplace health and safety.
    
6.  Mohammadi Amin, F., Rezayati, M., van de Venn, H. W. & Karimpour, H., "A Mixed-Perception Approach for Safe Human-Robot Collaboration in Industrial Automation," Sensors 20(21):6347, 2020 (peer-reviewed). [pmc.ncbi.nlm.nih.gov](https://pmc.ncbi.nlm.nih.gov/articles/PMC7664417/)
    
    **What we took from it:** The vision channel "can fail due to occlusion effects" and skeleton detection can be slightly affected by lighting (not tested experimentally); visual and tactile perception are each judged not sufficient on their own for intrinsically safe robot applications.
    
7.  European Parliament Legislative Observatory, document summary of Regulation (EU) 2023/1230 on machinery products (repealing Directive 2006/42/EC). [oeil.europarl.europa.eu](https://oeil.europarl.europa.eu/oeil/en/document-summary?id=1749182)
    
    Application date confirmed from official EU sources: [single-market-economy.ec.europa.eu](https://single-market-economy.ec.europa.eu/sectors/mechanical-engineering/machinery_en)
    
    the date set by the EUR-Lex corrigendum of 4 July 2023 to Article 54, which replaced an earlier 14 January 2027 ( [eur-lex.europa.eu](https://eur-lex.europa.eu/eli/reg/2023/1230/corrigendum/2023-07-04/oj/eng)
    
    **What we took from it:** Third-party conformity assessment is compulsory for six higher-risk categories (Annex I, Part A), including safety components with fully or partially self-evolving behaviour using machine learning approaches.
    
    the European Commission states the Regulation applies on a mandatory basis from 20 January 2027 (
    
8.  Workplace surveillance law in Australia. [arp.nsw.gov.au](https://arp.nsw.gov.au/c2005-37-workplace-surveillance-act-2005)
    
    The official Act text (legislation.nsw.gov.au) and AustLII refused automated access; the requirement for signs at each entrance is from a law-firm summary, Sprintlaw, "Workplace Surveillance Act 2005 (NSW) Explained," 31 October 2025. [sprintlaw.com.au](https://sprintlaw.com.au/articles/workplace-surveillance-act-2005-nsw-explained/)
    
    Victorian Government, "Workplace surveillance" (updated 18 November 2025): [vic.gov.au](https://www.vic.gov.au/workplace-surveillance)
    
    **What we took from it:** NSW Government circular C2005-37 (Administrative Requirements Portal), "Workplace Surveillance Act 2005": surveillance may only occur after written notice at least 14 days before it starts (less with the employee's agreement); camera surveillance needs visible cameras and signs; surveillance not notified under the Act is covert and "is an offence unless authority is obtained from a Magistrate".
    
    a parliamentary inquiry made 18 recommendations and 29 findings; the Government supports 15 in principle.
    

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 Improve Manufacturing Quality?](https://bfcs.ai/articles/can-ai-improve-manufacturing-quality/): More in Manufacturing · MANUFACTURING · BETTER · 6 min · Read →
- [Can AI Predict Maintenance Needs?](https://bfcs.ai/articles/can-ai-predict-maintenance-needs/): More in Manufacturing · MANUFACTURING · BETTER · 7 min · Read →
- [Can AI Protect Client Confidentiality?](https://bfcs.ai/articles/can-ai-protect-client-confidentiality/): Also Safer · PROFESSIONAL SERVICES · SAFER · 7 min · Read →

---

Canonical page: https://bfcs.ai/articles/can-ai-make-factories-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