Why Mining Safety Innovation Isn’t Reducing Fatalities Yet

Forty-two people died working in mines operated by ICMM member companies in 2024, up from 36 the year before. That increase came even as the sector poured capital into monitoring systems, artificial intelligence, and safety-focused research.

The gap between rising investment and rising deaths is the question that hung over Safemining 2026, the 4th International Conference on Safety and Occupational Health in Mining, which wrapped in Santiago on 3 September 2026. The event brought 404 delegates from eight countries to the Sheraton Santiago Hotel for more than 100 technical presentations, including 42 case studies drawn from active operations.

For investors and sector observers, that programme is a useful signal of where the safety frontier actually sits, as distinct from where press releases place it. What follows here separates the genuine progress from the unresolved, and identifies which variables most determine whether the current wave of safety investment translates into fewer deaths and lower operational risk.

Why the safety numbers are moving in the wrong direction

The long-run story mining companies like to tell is one of steady improvement. Across ICMM’s 24 member companies, which represent roughly a third of the global industry, fatalities fell from 90 in 2012 to the low thirties by the early 2020s. That trajectory is real.

The recent numbers complicate it. Fatalities reached 33 in 2022, rose to 36 in 2023, then climbed to 42 in 2024, a second consecutive annual increase.

The fatality frequency rate rose from 0.013 per million hours worked in 2023 to 0.015 in 2024. The headline trend has reversed.

The reversal is not evenly distributed across incident types. Underground operations accounted for 43% of fatal incidents in 2024, and two mechanisms did most of the damage:

• Mobile equipment interactions: 33% of underground fatalities
• Fall-of-ground events: 31% of underground fatalities
• Together, nearly two-thirds of all underground deaths

Geography concentrated the losses further. South African operations recorded 15 of the 42 fatalities, approximately 35% of the ICMM total.

Underground geotechnical hazards, particularly fall-of-ground events, remain among the most difficult mine risks to predict because rock mass behaviour depends on factors that vary across even a single stope, including stress redistribution after adjacent blasting and the presence of undetected structural features.

Yet the picture is genuinely mixed rather than uniformly worse. In the United States, the Mine Safety and Health Administration (MSHA) reported an all-injury rate of 1.74 per 200,000 hours worked in 2025, down from 1.82 in 2024 and the lowest on record.

MSHA’s 2025 all-injury rate announcement, published in April 2026, confirmed the 1.74 per 200,000 hours figure as an all-time low for U.S. operations, providing a counterpoint to the deteriorating ICMM aggregate and underscoring how jurisdiction-level conditions can diverge sharply from the global trend.

The read for investors is that safety spending is not yet producing consistent outcomes at the headline level. The concentration of deaths in underground mobile equipment interactions and unstable ground points precisely at the operational context where AI and drilling-data technology are now being marketed as the answer.

What AI and MWD technology can and cannot do

The case for the technology begins with documented deltas, not brochures. At Minera Los Pelambres in Chile, neural-network AI models lifted early warning effectiveness for small-scale slope instabilities from 43% to 82% over an eight-month evaluation, while cutting false positives by roughly 80%, according to a 2021 Australian Centre for Geomechanics study.

The pattern repeats across geographies. A Western Australian gold mine reported a 40% increase in early-stage instability detections against manual review. At a South American copper-gold operation, an AI system flagged a developing slope failure 72 hours before traditional radar alarms, buying enough time to evacuate equipment.

Measurement-While-Drilling (MWD) sits at the centre of the geotechnical use case. MWD systems capture high-resolution profiles of drilling parameters, rate of penetration, torque, rotation speed, and pressure, along every hole without slowing the drilling itself. Machine learning then maps those mechanical signals to rock mass characteristics, detecting the abnormal patterns that flag voids or weakened ground before an operator would see them.

Where the technology falls short in practice

The accumulated gains are persuasive. They are also fragile once a system leaves the pilot that produced them.

The documented limitations cluster into four failure modes:

• Data integrity: harsh downhole environments introduce noise, missing values, and inconsistencies that degrade predictive accuracy
• Generalisation: models trained on one mine’s geology frequently fail to transfer to a different deposit or equipment set
• Interpretability: deep neural networks often lack the transparency engineers and regulators need for safety-critical sign-off
• Infrastructure: real-time analytics demand bandwidth, low latency, and computing power that remote sites struggle to sustain

Dr. Alberto Fernández, presenting applied research on AI and MWD integration at Safemining 2026, positioned these as active research problems rather than solved ones. That distinction matters commercially.

For investors weighing companies or vendors making AI safety claims, the gap between a working pilot and a scalable operational system is the most material question the evidence does not yet resolve. A 43% to 82% improvement at one mine cannot be assumed transferable to another without substantial re-training and site-specific validation.

AI safety monitoring deployments in production environments consistently show a performance gap between controlled pilots and steady-state operations, a gap driven by sensor degradation, shift-change data gaps, and the accumulation of edge cases that training datasets did not capture.

Safety culture as the variable technology cannot replace

Technology only produces outcomes if a workforce uses it, reports through it, and challenges the conditions it flags. That makes culture the mechanism, not the soft counterpoint.

Workforce trust in monitoring systems is a precondition for the cultural shift safety researchers describe; operators who do not believe the system accurately represents their working environment will route around alerts, interpret false positives as evidence the technology is unreliable, and revert to informal hazard assessment.

Embedding safety as a permanent component of organisational culture is a critical challenge for the sector, Mg. Patricia Muñoz of Dimin USACH argued at Safemining 2026. Deploying a safety system and building a safety culture are not substitutes for each other.

The barriers are well documented. ICMM’s Fatality Prevention: Eight Lessons Learned names poor leadership, production pressure, and compliance-only tick-box behaviour as structural failures, the point at which safety systems become administrative burdens rather than genuine hazard awareness.

The interventions that work share a specific shape. They follow structured sequences and lean on peer-level ownership rather than top-down mandates:

• BHP Chile and Anglo American: Incident and Injury-Free and Zero Harm programmes that integrate contractors under a “one team” framing with field-based coaching
• Mount Isa Mines “Switch On”: a three-phase embedding sequence (Engage, Reinforcement, Systems) designed to sustain shifts beyond initial training
• Coeur Mining SafeStart: experienced internal miners as peer trainers, building ownership instead of relying on external consultants
• 2025 MDPI study of Polish rock mining: visible managerial commitment and non-punitive near-miss reporting identified as the two conditions most correlated with daily safety integration

The through-line is that durable change follows a structural sequence, not a campaign, and that the strongest results come where peer ownership drives it rather than leadership announcements. Muñoz’s related Chilean research for IST and Suseso treats worker-level preventive norms, not management systems, as the target unit of change.

For investors evaluating operators, the presence of a structured cultural programme is a more reliable predictor of incident rates than the presence of monitoring technology. Technology cannot compensate for a workforce that will not report a near-miss.

The research pipeline: why academia and industry are converging now

The Safemining 2026 format was itself a statement about how safety knowledge now gets made. Co-organised by Gecamin, Dimin USACH, and the University of Calgary, with Anglo American as primary mining sponsor, the conference deliberately routed academic research through operational case studies rather than leaving it in proceedings volumes. The programme drew 272 authors and co-authors across six countries and those 42 field case studies.

That model is not confined to Chile. It shows up in recurring international forums, the International Symposium on Mine Safety Science and Engineering (ISMSSE), whose 9th iteration was hosted by the University of Wollongong in 2026, and the ISSA Mining Safety Conference 2026 in Saskatoon.

What makes these collaborations work in practice

The partnerships that endure share four traits. They involve multiple stakeholders, regulators and labour, not only companies and universities. They run on sustained multi-year funding. They balance productivity gains against safety impacts explicitly rather than treating one as a by-product of the other. And they feed academic outputs directly into operational training.

The A$1.25 million, five-year Jim Askew Fellowship, announced by Evolution Mining and the University of Queensland in October 2025, illustrates the shift. It targets predictive safety and efficiency models designed for operational integration from the outset. In 2026, Jennmar Australia partnered with UNSW on advanced ground support for geotechnical safety.

Industry-funded safety research chairs have become a preferred vehicle for translating site-level incident data into publishable findings, partly because they give mining companies input into research priorities while satisfying academic independence requirements that regulators and labour unions have increasingly demanded.

The 42 operational case studies at Safemining 2026 were an application of the same principle, field-validated findings sitting alongside academic work rather than apart from it.

For investors, the implication is a compressed innovation timeline. Fellowship-funded work built for deployment reduces the development risk that usually accompanies research-stage technology.

What the evidence says about where the sector actually stands

Three threads run through the evidence, and they do not move at the same speed. Technology capability is advancing and documented. Cultural change is structurally understood but unevenly implemented. The research pipeline is now institutionalised enough to accelerate both.

The unresolved tension sits in the workforce question. Autonomous and AI-driven systems introduce a variable that safety research has not yet fully absorbed, and the sector’s own figures point in opposite directions.

Autonomous systems may reduce operator requirements by 30-50% and displace 60-80% of operators. Yet approximately 80% of mining companies planned to increase headcount while adopting AI. Both figures are real.

That contradiction is not a data-quality problem. It reflects a genuine strategic choice individual operators are making in different directions, and the safety outcomes of those choices have not been tracked long enough to judge.

Readers assessing the sector need a framework that separates two frontiers: the capability frontier of what AI and MWD can demonstrably do in controlled deployments, and the implementation frontier of what the average operation actually does. The gap between the two is where most of the sector’s current risk lives.

The two-year rise in ICMM fatalities is the benchmark against which every conference claim should be measured. The companies most likely to close the gap between safety investment and safety outcomes are those treating technology, culture, and research integration as simultaneous priorities rather than sequential ones.

This article is for informational purposes only and should not be considered financial advice. Investors should conduct their own research and consult with financial professionals before making investment decisions. Financial projections and forward-looking statements are subject to change based on market developments and various risk factors.