AI, Algorithm: The Union Will Read The Code
A New South Wales law, a French precedent, and what both mean for Western Australia
There is a sentence in a New South Wales statute that nobody in Perth has read, and it will change what a manager is expected to know.
The Work Health and Safety Amendment (Digital Work Systems) Act 2026 received assent on 18 February. Most Australian commentary treated it as a gig economy matter, filed it alongside delivery riders and warehouse pickers, and moved on to the federal announcements, which are more photogenic. The Act defines a digital work system as an algorithm, artificial intelligence, automation or an online platform. It inserts a new section 21A into the Work Health and Safety Act 2011, requiring a business to ensure that the allocation of work by such a system does not put the health and safety of workers at risk. And it gives work health and safety entry permit holders, who are union officials issued a permit by the Industrial Relations Commission, the power to require an employer to provide reasonable assistance to access and inspect a digital work system relevant to a suspected contravention.
Read that last clause again, slowly.
The regulator is careful about what is actually new here, and the distinction matters. SafeWork NSW states that entry permit holders already had a right to inspect any work system, plant, substance or structure relevant to a suspected contravention, and that a digital work system was always included in that. What changes is the obligation on the employer to help; to open the thing; to assist a union official in getting inside the system that allocates shifts, sets targets, ranks performance or decides who gets the difficult run on a Tuesday.
The wording of the definition repays attention, because a word went missing on the way through Parliament. An earlier version of the Bill included software in the list, and its removal is read by practitioners as putting email and messaging systems outside the definition. What remains covers an algorithm, artificial intelligence, automation or an online platform, which is broad enough to catch a rostering engine and narrow enough to leave the inbox alone.
Assistance is the whole story. A right to inspect a system you cannot open is a right to look at a screen. A right to require assistance is a right to an explanation, and an explanation has an author.
The Act also lists what a business must turn its mind to. Section 21A requires consideration of whether the allocation of work by the system produces excessive or unreasonable workloads, excessive or unreasonable performance metrics, excessive or unreasonable monitoring or surveillance, or unlawful discriminatory practices and decision-making. Rostering software, productivity dashboards and automated task assignment all sit inside that description, whatever the vendor calls them.
None of it has started. The Act commences on a day or days to be appointed by proclamation, and no date has been announced. Meanwhile the entry power carries a leash written into the statute: the delegate's ability to demand assistance is expressly subject to guidelines the regulator must issue and publish under a new section 118A. SafeWork NSW opened its consultation page in late May and has set out a staged process: targeted consultation with stakeholders, further targeted engagement, then a draft published for four to six weeks of public comment on the New South Wales Have Your Say platform. As I write, no draft has appeared. Anyone can register to be notified when it does.
So the statute creates a right and hands the drawing of its boundaries to an administrative document currently being written in a room most employers have not found. The people who prepare during the quiet period are never the ones who complain afterwards about the noise.
The compliance timetable can look after itself. What holds my attention is the figure who walks through the door.
Picture the union official of the twentieth century. Boots, a torch, a clipboard, a guard removed from a machine, a witness statement, a photograph of a cable across a walkway. The hazard was physical and so was the inspection. Everything that could hurt a worker could be seen by a person standing in a shed, and the skill of the delegate was the skill of looking properly.
Now picture the same official in front of a rostering engine. The hazard is a weighting. A parameter that penalises a driver for a slow delivery without recording that the lift was broken. A performance score decaying over fourteen days rather than thirty, so that a week of sick leave follows a person into the next quarter. Or take a fatigue model treating a twelve hour shift and two six hour shifts as equivalent, because the total is the same. None of this is visible from the floor. All of it allocates work, and allocation is precisely what the new duty covers.
The delegate who wants to do that job needs a different competence: to read documentation, and then to notice what the model does not measure. Some unions will build that capability. Some will buy it. Australian workplace relations has not yet imagined what an expert union looks like, one that arrives with a data analyst rather than a torch, and I suspect a good many boards have not imagined it either. They will meet it before they picture it.
There is a second consequence, and it should trouble a board more than the first.
Australian safety law ranks controls by hierarchy, and elimination sits at the top: remove the hazard, so far as is reasonably practicable, before attempting to manage it. Substitution and engineering controls follow. Training and procedure sit near the bottom, because the law treats anything relying on human compliance as fragile. Read section 21A through that hierarchy and a possibility opens that nobody drafting the Act would defend in public. The duty concerns the allocation of work by a digital system to a worker. Automate the driver out of the vehicle and there is no worker being allocated the run, no weighting left for a delegate to inspect. The risk has been eliminated in the most literal sense the statute allows, and the hierarchy will call that best practice.
Western Australia has already tested the manoeuvre without describing it that way. The Pilbara operates the largest autonomous haulage fleets in the world. Whatever mix of reasons produced those decisions, one of them requires no elaboration in front of a regulator: an empty cab cannot be crushed. The hazard and the occupation left the site together, because in that instance they were the same thing.
I should be careful not to overstate the mechanism. Reasonable practicability weighs the cost of a control against the degree of risk, and no regulator would accept the automation of a national fleet as a proportionate response to a psychosocial complaint. No Australian decision has yet treated automation as an answer to section 21A, and I am describing a pressure rather than a precedent. It works at the margin, on decisions already close to viable, where a compliance obligation becomes the last argument in a business case that was nearly written. Nor does automation empty the duty; it relocates it. The operator watching eleven autonomous trucks from a control room at three in the morning is a worker whose task is allocated by a digital system, squarely inside section 21A. The duty follows the survivors.
The philosophical problem underneath is older, and it has been sitting there for a century without anyone needing to look at it.
We built the methods of workplace safety to protect people. Root cause analysis and the hierarchy of controls, the whole apparatus of industrial engineering: all of it was designed to find the source of harm and remove it rather than manage its symptoms, and all of it was built by people who assumed, without ever writing the assumption down, that the analyst and the protected were on the same side. The method never contained the instruction. The people applying it did.
Hand the same method to a system with no stake in the outcome and it completes the reasoning we always stopped short of. Ask what causes variability in an operation and the answer is the human. Ask what remains after every engineering control has been applied and the answer is still the human, now expressed as the largest residual term in the equation. The model does that faithfully, at speed, arriving at the conclusion a century of practitioners declined to reach because they knew the people in the shed by name.
So we now have a methodology conceived to protect humans, executed by a machine, which identifies humans as the residual hazard. How do you prevent that?
The instinct is to ask the machine to be more ethical, which mistakes where the choice happens. Every risk assessment answers the question it was given, and the question is set before any model runs. Ask what puts workers at risk and you receive one list. Ask how to minimise residual risk and you receive another, on which the workers appear as the problem. No system chooses between those two questions. A person does, in a meeting, usually in a sentence nobody records.
Prevention, then, is a matter of writing the constraint back into the brief. The engineers of the last century carried it implicitly; ours have to state it. The purpose of a safety analysis is to make work survivable for the people doing it, which means the objective has to include the continued existence of the work, and the assessment has to be scoped as a question about conditions rather than about the elimination of variables. A manager who commissions an analysis without specifying that has already chosen, and has chosen the version where the answer is known in advance.
Two things follow, and I offer them as the beginning of an argument. The question put to the system should be recorded and retained alongside its output, since a recommendation cannot be reviewed by anyone who does not know what was asked. And the people affected should be present when the question is framed rather than consulted on the conclusion, because consultation after the framing is theatre with better documentation.
France has arrived at part of this from a different direction, which is why the comparison is useful rather than decorative.
There is no French AI workplace safety statute. What exists is an old provision of the Code du travail, article L2312-8, requiring the employer to consult the social and economic committee (the CSE, the elected staff body that replaced the works council in 2017) before any important project introducing new technology capable of affecting working conditions. Employers argued for a while that installing a new HR tool was a technical upgrade requiring information rather than consultation. The courts have been dismantling that argument, case by case.
On 2 September 2025 the tribunal judiciaire in Paris held that installing a generative AI platform was a new technology triggering the consultation obligation. On 15 July 2025 the tribunal judiciaire in Créteil ordered two subsidiaries of a press group to stop using an AI tool deployed without prior consultation. Then, on 29 January 2026, the tribunal judiciaire in Nanterre suspended the rollout of two HR management packages at a company of around two thousand employees, with a penalty of five hundred euros for each day of delay. The employer had told the committee about the change. The judge found that telling was not consulting, and looked at what the software actually did: extended the user base from two business units to the entire workforce, embedded decision support algorithms in evaluation and assignment, and used HR data in a new way to recommend training paths. A version upgrade, the company said. A judge in Nanterre could see otherwise.
What matters for our purposes is the scope of the French obligation. Consultation attaches to the project rather than to the tool, so a reorganisation that removes roles falls inside it on its own account. A French employer cannot use automation as an exit from the conversation, because automation is the conversation. From 2 August 2026 the European AI Act adds another layer, treating AI used in recruitment, evaluation and task allocation as high risk, with documentation and human oversight obligations attached.
The limits deserve stating on both sides, since the temptation is to romanticise whichever system one is not living under. The French CSE has no veto. It can issue a reasoned unfavourable opinion and, where the legal conditions are met, appoint an independent expert, and courts have refused that where the project was slight; a Paris decision of 10 February 2026 declined an expertise for an optional tool inside an existing licence over a short period. The Australian entry permit holder is bounded too, by guidelines not yet written, with existing penalties for misuse of entry powers still in force. Nobody is handing source code to a delegate on a Monday morning. An employer who declines to open the system ends up in a dispute about right of entry, which the Act routes to the Industrial Relations Commission, free to deal with it in any manner it thinks fit. Australian bargaining practitioners will recognise the shape of that: the boundary of the new right will be settled case by case, in the tribunal, by the parties who show up prepared.
Yet the asymmetry is the finding. Australia has built a duty about the machine and left the decision to install it largely unexamined. France examines the decision and has comparatively little to say about the machine. Each country has legislated the half of the problem its own tradition made visible, and each is now discovering the other half by accident.
Which brings me back to the manager, because this is where the two arguments meet.
For twenty years the promise of digital management was that it removed the awkwardness from authority. No favouritism in the roster, no argument about the target, because the system produced both. A generation learned to say that the tool had decided. New South Wales has legislated to withdraw that sentence from circulation, at least in one state, at least once the proclamation comes and someone with a permit asks. Good managers will not miss it; the ones I admire never used it, knowing that the moment you cannot explain why someone got the worse shift, you have stopped leading them and begun administering them.
The harder discipline is the one nobody will inspect. Once proclaimed, a permit holder entering a workplace on a suspected contravention will be able to require assistance to get inside a rostering engine. No permit holder has a right of entry to the question a manager asked before the analysis began, and that question, set quietly and never minuted, is where the workforce is actually decided.
New South Wales legislated its own obsolescence in the same breath. A new section 276D obliges the Minister, once satisfied that Safe Work Australia's model laws cover substantially the same subject matter, to review whether these provisions remain appropriate and whether the national model offers workers a higher standard. The state has written down, in statute, that it expects to be overtaken. Nothing in any of this binds an employer here, where we work under the Work Health and Safety Act 2020 (WA), and no other jurisdiction has legislated the same duty. What the review clause does is name the route by which the question would arrive: through the model laws, drafted nationally, adopted here as they have been before.
The guidelines will be published, the date will be proclaimed, and somewhere in New South Wales a delegate will ask a company to open a system.
Picture that room. Someone from operations, someone from legal, a laptop turned around on the table, and a delegate reading a screen that nobody in the building has looked at properly for three years. Whoever framed the question that shaped what is on that screen has long since moved on, probably promoted. That absence is the thing I would want explained, and nobody in the room will think to ask about it.
21A Duties of persons conducting businesses or undertakings involving digital work systems
(1) A person conducting a business or undertaking must ensure, so far as is reasonably practicable, that the health and safety of a worker is not put at risk from the allocation of work by a digital work system used by the business or undertaking.
(2) A person conducting a business or undertaking must consider whether the allocation of work by or using a digital work system creates or results in any of the following risks: (a) excessive or unreasonable workloads for workers at work in the business or undertaking, (b) the use of excessive or unreasonable metrics to assess and track the performance of workers at work in the business or undertaking, (c) excessive or unreasonable monitoring or surveillance of workers at work in the business or undertaking, (d) unlawful discriminatory practices or decision-making in the conduct of the business or undertaking.
Work Health and Safety Amendment (Digital Work Systems) Bill 2026 - Parliament of NSW