AI and workplace safety: spotting risk without surveilling people

You have done the hard part on safety. Procedures are written, protective equipment handed out, training up to date, toolbox talks held. And yet the accident rate no longer falls, near-misses keep coming, the same situations return. Then comes the seductive promise: a smart camera that spots, in real time, everyone not wearing their helmet. Before installing that eye, it is worth seeing where the path leads, because there is another one, more effective and far less risky.

Why safety plateaus

Once the basic measures are in place, what remains no longer shows to the naked eye. The risk does not disappear, it dilutes. A near-miss is reported, filed, then forgotten. Another, described in different words, lands in a different folder. Nobody is tasked with reading a thousand reports together, and yet that is where the signal hides.

It is exactly the mechanism found elsewhere, in recurring quality non-conformities or in unplanned shutdowns whose signal was already there: the fact was present, diluted in the volume, invisible as long as nobody looked at the whole. Safety rarely plateaus for lack of rigour. It plateaus because the useful information is scattered.

What AI does best, and most discreetly

The soundest use of AI in safety is also the least intrusive: making the reports you already have speak. A language model reads your near-miss reports, your incident analyses and your write-ups, all in free text, and groups those that describe the same situation, even when worded differently.

Where you saw a thousand scattered reports, it surfaces the recurring precursors: the same gesture, the same station, the same time of day that precedes incidents. It does not tell you what to do. It shows you where to look, at a scale no safety officer has time to cover by hand. It is prevention intelligence, drawn from your own data, without a single camera.

The surveillance trap

Then comes the camera that detects protective equipment and behaviour. Technically, it works. Legally and humanly, it is a minefield.

On the legal side, filming employees continuously to assess them runs into the GDPR and labour law: the system must be proportionate, informing people and consulting the works council are mandatory, and permanent, systematic surveillance of a workstation is regularly ruled unlawful. The European AI regulation, moreover, classes worker-monitoring systems among high-risk uses, subject to heavy obligations. This is not a compliance detail, it is a red light.

On the human side, the cost is worse still. A team that knows it is filmed and graded stops trusting. Near-miss reports, that precious fuel, dry up, because reporting an incident becomes a personal risk. People work around the system, they hide rather than report. The camera installed to improve safety ends up draining the very source of information that made it progress.

If you use vision, change the target

Image analysis does have a legitimate place in safety, provided you reverse its logic. You do not watch the person, you watch the situation.

Detecting that a human enters the danger zone of a machine during its cycle, and triggering a stop, protects without identifying anyone. Processing happens as close to the sensor as possible, without recording or storing a face, and the alert serves immediate prevention, not the building of a file. The difference fits in one sentence: you watch a hazard, not a worker. This kind of system is built with the operators and the works council, not against them, because those who hold the job know where the real risks are.

The human decides, and the data stays with you

The principle running through the whole subject is that of any responsible use of AI in industry, stricter still here, because the data touches people. AI flags a risk pattern; humans investigate, understand and decide. Never an automatic sanction drawn from a system. A risk report is a hypothesis to look into, not a verdict on someone.

And because this data speaks of your employees and your installations, it has no business in an outside service. Processing on site, locally, without the data leaving the company, is not one precaution among others: it is the only defensible approach, for the same reasons that hold for any sensitive data handed to an AI.

Where to start

The first step needs neither camera nor budget: it is your near-miss and incident register, as it stands. Group it by meaning, locally, and check whether a precursor you suspected stands out. If it does, you hold a concrete line of action, obtained without surveilling anyone. That is the right order: exploit first what your teams have already reported to you, before even thinking of installing any kind of eye.

The essentials

AI has a real place in workplace safety, but not the one first offered to it. Its strongest and healthiest contribution is to surface the signals diluted in your reports, not to film your teams. If vision becomes necessary, let it target the situation and not the person, anonymously and to prevent. The decision that involves an individual stays human, and the data stays with you. On those conditions AI makes safety better, instead of making it suspect.