Industries / Industrial equipment and services / Industrial services
Safety training
Safety training is the work of teaching people who do dangerous jobs, such as miners, builders, plant and cleanup workers, how to do them without being hurt, and of keeping proof that they were taught.
1910
The new US Bureau of Mines sent railroad cars, refitted with a training room and workroom, through the coalfields to teach miners first aid, mine rescue and safe working. In its first five years about 34,000 people were trained in rescue and first aid and 11,700 training certificates were issued, by the Bureau's count as given in a history published by the Mine Safety and Health Administration (MSHA).
35
Milestones achieved across 6 domains.
10
To watch for: announced, not yet confirmed outside.
12
Open deliverables. Nobody has shipped them.
In one minute
Where it started
Teaching workers to do dangerous work safely became an organized job in 1910, when US government railroad cars began touring the coalfields to train miners in first aid and rescue. In 1919 Charles R. Allen, who trained shipyard instructors in the First World War, set out four teaching steps that instructors still follow. From the 1970s regulators wrote training into law: hours of training before a miner goes underground, 40 hours plus supervised fieldwork for hazardous waste workers, and cards such as OSHA's 10-hour course card and Britain's CSCS card.
Where it stands
The International Labour Organization (ILO) estimates that in 2019 about 330,000 workers died in work accidents and more than 395 million were hurt. Training is given at great scale, more than 6.5 million US workers through OSHA's outreach courses in five years by OSHA's count, but the evidence on what it achieves is uneven: pooled studies find that training raises test scores far more than it lowers injuries, lifting lessons do not prevent back pain, and virtual reality has beaten slides in some trials and not in others. In an Ontario study, workers in their first month on a job were more than four times as likely to make a claim for an injury that kept them off work as those with over a year in it, and in Japan foreign trainees were hurt at 3.98 per 1,000 a year in 2024 against 2.3 for all workers, by the health ministry's figures.
What is open
The trade lacks shared yardsticks. Nobody has published a test of machine-translated safety talks, a speech set recorded in site noise, a check of the AI tools that now draft inductions or choose retraining after accidents, or curves showing how fast hands-on skills fade, set against how often cards must be renewed. Each can be built today with AI models, phone video, consumer headsets and public accident records, without putting anyone near a real hazard.
Domains
AchievedWatch forOpen
Domain
Progress
Last achieved
Watch for
Open
Site orientation
Site orientation is the first safety teaching a worker gets on joining a job or entering a site: its rules, its hazards, how to raise the alarm and how to get out.
6 · 2 · 2
Last achievedOrientation renewed in eight languages2024
Watch forShorter refresher, proctored online
2
openHands-on skills
Hands-on skills are the physical routines a worker must get right under stress, such as putting on emergency breathing equipment or working in a protective suit, learned by practice while an instructor watches.
5 · 1 · 2
Last achievedRescue teams train in smoke2008
Watch forGlasses that coach the wearer
2
openSimulators and virtual reality
Simulators and virtual reality (VR) let workers rehearse dangerous jobs and emergencies in a copy of the workplace, from a replica control room to a headset that shows a computer-drawn site.
6 · 2 · 2
Last achievedMeta ends business headset sales2026
Watch forExtinguisher skills after VR
2
openCertification and records
Certification and records are the cards, certificates and databases that show a worker has had the safety training a job needs, and let an employer check it.
5 · 1 · 2
Last achievedCards without qualifications end2024
Watch forAI picks the right card
2
openLanguage and literacy
Teaching safety to workers who do not speak the site's main language well, or who read little in any language.
6 · 3 · 2
Last achievedJapan's trainees hurt more often2024
Watch forTranslation app for site terms
2
openMeasuring whether it works
Checking whether training changed what workers know, what they do at work and how often they are hurt.
7 · 1 · 2
Last achievedKnowledge fades after class2026
Watch forRetraining picked by AI
2
open
Advances to build with
4 research notes

Mistral Large 4
A trillion open parameters
Mistral published its Large 4 AI model, a trillion numbers by its count, free to download and sell, though firms that rent out models need a separate license.

Genie 4
A world that follows
Google says its Genie 4 draws a world you can walk through, hear and change by typing, and its own pages give no figures for how well that world holds up.

Muse Ember
The jobs agents repeat
Meta's Muse Ember matched Claude Haiku 5.5 on an outside firm's tests at a third of the cost per task, though both list the same prices for text in and out.

Cursor Agent SDK
Findings that cite notes
Cursor released a kit for building AI agents, and its example, a code reviewer that cites the note behind each finding, is known only from Cursor's account.
Open in safety training
3 of 12 open deliverables
Open · Site orientation
Score AI-drafted site inductions
A public test of site inductions drafted by software. Take 30 site safety plans that are already public, such as those attached to public works tenders, and have two safety professionals turn each one into a checklist of that site's hazards, rules and emergency arrangements. Then have induction builders and general AI models draft an induction for each site from the same documents, and score how many checklist items each draft covers, how many it states wrongly and how much generic material it adds.
Built with what exists
Public site safety plans; trial accounts of induction builders such as Plexa Pro; general AI models such as Claude or Meta's Muse Ember (see the Note "The jobs agents repeat"); an agent that cites the page behind each checklist item, such as one built with Cursor's agent kit (see the Note "Findings that cite notes"); a spreadsheet; Zenodo to publish with a DOI.
Open · Hands-on skills
Score self-rescuer donning on video
A program that watches phone video of a miner putting on a self-rescuer training unit, marks each step of the maker's donning procedure as done, done out of order or missed, and times the moment the miner starts breathing from the unit, together with a released set of labeled videos.
Built with what exists
The realistic training units mines already buy for expectations training; phone cameras; Google's open MediaPipe hand and body tracking; an open vision-language model such as Gemma 4 for the step checks; CVAT, an open labeling tool.
Open · Simulators and virtual reality
Hazard hunts in generated scenes
A test of whether trainees find the hazards in computer-generated site scenes as often as in real photographs. Scenes are generated from OSHA's severe injury narratives, each with a known list of planted hazards, and paired with real photos of similar work; trainees mark every hazard they see.
Built with what exists
OSHA's Severe Injury Reports data, which hold a description of each amputation, hospital stay or eye loss reported since 1 January 2015; an open image model such as Black Forest Labs' FLUX.1 or Stable Diffusion; Unity to turn scenes into walk-throughs; later, Google DeepMind's Genie 4 world model, which draws a walkable scene from typed text but is so far open only to trusted testers (see the Note "A world that follows"); a web form or PsychoPy to run the test.
Where it sits
Broader