A Master Mariner Asks What the Cameras Are Actually Measuring

Alex Byelyavtsev, a serving master mariner with Maersk, published an argument this week that this publication has been circling for months without stating plainly: AI cameras are very good at detecting visible unsafe behaviour, and detection is not the same thing as understanding. The argument. A camera can register a missing helmet, an unguarded step into a restricted area, a mooring party standing in a bight. What it cannot register is why — whether the officer was undermanned, whether the procedure was unworkable, whether raising the problem last month produced silence. Byelyavtsev's sharpest line is the one to keep: "A camera does not observe competence. It observes actions treated as signs of competence." He invokes Goodhart's law — when a measure becomes a target, it ceases to be a good measure — and applies it directly: a fleet reporting 96% PPE compliance may have improved its safety culture, or may have taught its crews where the cameras point. Nothing in the data distinguishes the two. The systems he names are the ones we have covered. Captain's Eye reached its 100th installation across Eastern Pacific Shipping's 300-plus vessel fleet. ShipIn's FleetVision BASE is subsidised by NorthStandard for member vessels. Pacific Basin has it on more than 100 bulkers. Orca AI illustrates the piece. He also reaches for a precedent from mining, where NIOSH estimated conventional proximity-detection systems could have prevented around 82% of studied fatalities — a genuine result, and a useful reminder that these systems do prevent real harm. Why we are running this, and why it is uncomfortable. Check our archive. We reported Ultranav scaling FleetVision to 420 vessels in March, NorthStandard becoming the first P&I club to fully fund a deployment in April, ShipIn's $52m raise on a base of 1,300 vessels and 92 owners on 9 August, and Stealth Maritime rolling cameras across 50-plus tankers after 85% of officers said they would welcome them on 16 August. We have been one of this technology's more consistent chroniclers. In August we noted the company had more ships and fewer published outcomes than a comparable study. That observation still stands, and a serving master has now made the case better than we did. Why this matters for ship managers: you are probably being sold, or already running, a system whose value is expressed as a compliance percentage. That percentage is a proxy. Proxies are not worthless — a falling PPE violation count is worth having — but a proxy becomes actively harmful the moment it is used to rank masters, set bonuses or answer a charterer. What to do now: before your next review, ask three questions of your own data. Has the violation rate fallen because behaviour changed, or because people learned the camera's field of view? Does anyone review flagged events with the crew, or only about them? And is a single number from this system feeding any appraisal, TMSA response or commercial claim? If the answer to the last one is yes, you have converted a measure into a target, and Goodhart is already at work. The honest caveat: this is a signed opinion column, not a study. Byelyavtsev presents no fleet data of his own, and none of the vendors named are alleged to have made false claims. The mining figure comes from a different industry and a different technology. And the counter-argument is real: detecting an unsafe act before it becomes an injury has value even if it tells you nothing about culture. The fair conclusion is narrower than either side would like — these systems measure something worth measuring, and it is not what the word "culture" implies. [Splash247 · AiatSea, 16 August]

This story is part of the Maritime AI Digest of 30 August 2026.