When Silence Becomes the Signal
Mining’s AI safety systems are teaching supervisors to read a quiet dashboard as a safe shift, and the law will not accept that reading.
Sign up for my Substack daily AI newsletter here.
Follow me on LinkedIn: https://www.linkedin.com/in/johanosteyn/
In July, Impala Platinum halted its entire Rustenburg operation, which employs about fifty-one thousand five hundred people, for a safety reset after two fatal accidents weeks apart involving underground locomotives. Those locomotives carried proximity devices designed to cut their drives before a collision. The Minerals Council’s head of safety attributed the recent rise in such deaths largely to human error, including poor maintenance of these systems and a lack of understanding of how they work. The investigation will establish what went wrong at Rustenburg. The wider question for every mining executive is what happens to human vigilance once we trust a machine to be vigilant for us.
CONTEXT AND BACKGROUND
South African mines are deploying artificial intelligence and sensor-based systems for fatigue monitoring, collision avoidance, gas detection and predictive maintenance. These tools are rightly sold as life-saving, and in many operations they are. The timing, however, should concern every board. The industry recorded forty-one deaths in 2025, the lowest in its history. Yet by mid-September 2026, the Minister of Mineral and Petroleum Resources reported forty-nine fatalities for the year so far, with platinum and gold accounting for more than seventy per cent of them. The reversal does not prove that technology has failed. It does remove any comfortable assumption that more sensors automatically mean fewer funerals. Safety outcomes depend on how people behave around these systems, not merely on whether the systems are installed.
INSIGHT AND ANALYSIS
Researchers call the underlying problem automation bias: the tendency to over-rely on an automated system, favouring its output even when other evidence contradicts it. Georgetown’s Center for Security and Emerging Technology concludes that keeping a human in the loop cannot, by itself, prevent all accidents or errors. The mechanism is subtle. When a safety system rarely produces false alarms, people learn to treat its silence as reassurance and stop checking for themselves. In one collision-avoidance simulation, participants who had adapted to how an automated detector usually erred caught only about twelve per cent of an unexpected error. They caught fifty-four per cent of the errors they had learnt to expect. Underground, this is the most dangerous moment of all: the alarm is quiet, the dashboard is green, and nobody is looking.
I have previously written about this dynamic in the context of the UK air traffic control failures. In that article, I argued that the more capable an automated system becomes, the less able its human operators are to intervene when it stops. I also argued that boards rarely measure the manual capacity they are quietly losing. Mining faces the same paradox with far higher stakes. A supervisor who once judged crew readiness by talking to people at the start of a shift now glances at a fatigue dashboard. The skill does not vanish overnight. It erodes gradually, until the day the system misses something.
IMPLICATIONS
The law has not moved with the technology, and it does not need to. Under the Mine Health and Safety Act, a mine death must ordinarily be followed by a statutory inquiry. The chief executive must take reasonable steps to ensure the employer’s safety functions are performed, and appointing managers does not relieve the employer of any duty. No inquiry will accept “the system did not flag it” as an answer. It will ask what human checks existed independently of the system.
There is also a dignity question, and it leads back to the same trap. Fatigue cameras watch a worker’s eyes for an entire shift. Research at the University of Essex found that workers under such monitoring may conceal fatigue to avoid consequences. It also found that employers may rely on monitoring to demonstrate compliance without addressing workload or staffing. At that point, the camera stops being a safety tool and becomes a compliance facade, one more source of false comfort for management. The Labour Research Service argues that biometric data should be collected only where no less intrusive option exists. It also argues that workers should be able to veto monitoring that threatens their dignity until the employer demonstrates otherwise. A camera that workers distrust is a camera they will learn to defeat.
CLOSING TAKEAWAY
AI safety systems belong underground, and the mines that use them well will save lives. The discipline they demand is uncomfortable. It means deliberate manual checks that the system does not prompt, supervisors judged on what they observe rather than on alerts they close, and regular drills in which the system is switched off. Those independent human checks, documented and tested, are what an inquiry will look for first. Accountability cannot be delegated to a sensor. When silence becomes the signal, someone must still be looking, and the law already knows whose name will appear in the inquiry report.
Johan Steyn is a prominent AI thought leader, speaker, and author with a deep understanding of artificial intelligence’s impact on business and society. He is passionate about ethical AI development and its role in shaping a better future. Find out more about Johan’s work at https://www.aiforbusiness.net
Comments