Unsafe Act vs Unsafe Condition: The Difference

Almost every safety observation card, inspection form and incident report in circulation asks the person filling it in to choose between two boxes: unsafe act or unsafe condition. It looks like a trivial administrative question. It is not. Whoever ticks that box is, in practice, deciding whether the organisation will spend money on a physical change or have a conversation with an individual. That is one of the most consequential decisions in an operational safety system, and it is usually made in three seconds by someone who was told the categories were obvious.

The message up front: the distinction between an unsafe act and an unsafe condition is real and worth understanding, because the two call for genuinely different responses. But the boundary between them is far softer than the two boxes suggest, it collapses almost always in the same direction, and the direction it collapses in puts the cause at the sharp end where the work is done rather than at the desk where the conditions were created. Learn the definitions, then learn where they break.

This article is a general explanation of two widely used categories. How any specific organisation defines, classifies and responds to what it observes is a matter for its own safety management system, its legal jurisdiction and its own competent people, and this is not a substitute for either.

1. Where the two categories come from and what they are for

The act and condition split is not a standard. No standards body issues it, no regulation I am aware of defines it, and it has no edition year. It is a piece of vernacular safety practice that has been in general use long enough that most people encounter it as if it were a technical definition. It survives because it is easy to teach, easy to put on a form, and it maps onto something that feels true: some of what makes a workplace dangerous is out there in the plant, and some of it is in what people do.

That intuition is worth keeping. The categories are most defensible when they are treated as a way of describing what was observed, rather than as an explanation of why it happened. A missing guard and a bypassed interlock are different kinds of observation. Treating them as the same thing would be unhelpful. The trouble starts when the description is promoted to a cause.

It also helps to keep this separate from two neighbouring distinctions that people muddle into it. The act and condition split is not the same as the difference between a hazard and a risk, which is covered in the hazard versus risk explainer, and it is not a method for finding hazards in the first place, which belongs to hazard identification methods. This article stays on one question only: what each of the two categories is, how to tell them apart, and what the split is and is not good for.

2. What an unsafe condition is, with examples

An unsafe condition is a physical state of the workplace, the equipment, the materials or the environment that could cause harm. The test is that you could photograph it and it would still be there after everyone went home. Nobody has to be present for the condition to exist.

Examples, weighted toward maintenance and facilities work because that is where this site's readers live:

• Guarding and enclosure. A missing or damaged machine guard, an access panel left off a drive coupling, a pump coupling guard that was never refitted after the last overhaul, an open floor penetration with no cover.
• Access equipment. A ladder with a bent stile or a missing foot, a mobile tower with a wheel that will not lock, a fixed ladder whose cage fixings are corroded, a plant room reached only by standing on a pipe.
• Egress and housekeeping. A fire exit blocked by pallets, a corridor narrowed by stored spares, a stairwell used as a store, an exit door that has been chained.
• Electrical. An exposed conductor, a damaged flexible lead, a panel door that will not latch, missing blanking plates in a distribution board, unlabelled isolators.
• Mechanical and fluid. A leaking valve gland, a failed pressure gauge, a corroded pipe support, a compressed air line with a damaged coupling, a hot surface with no lagging.
• Environment. Poor lighting in a switch room, standing water on a plant room floor, inadequate ventilation in a battery room, extreme heat in a rooftop chiller yard, noise from an untreated fan.
• Structure and fabric. A damaged handrail, a loose grating, a ceiling tile displaced over a walkway, a roof access hatch with no edge protection nearby.

Note what these have in common. Every one of them is closed out by changing a physical thing: fit the guard, replace the ladder, clear the exit, repair the lead, fix the leak, add the light. The corrective action is a job, and a job has a cost, a part and a date. That is precisely why the condition category tends to be the honest one.

3. What an unsafe act is, with examples

An unsafe act is something a person does, or fails to do, that departs from a safe way of working. It requires a person to be present, and it exists only while the behaviour is happening. You cannot photograph an unsafe act after everyone has gone home, and that asymmetry matters more than it first appears.

• Bypassing protection. Defeating an interlock so a machine can be run with a door open, jamming a guard switch, overriding a trip to finish a commissioning check, propping open a fire door.
• Working without isolation. Starting work on equipment that has not been locked out, working on a live panel that could have been dead, breaking into a line that has not been drained or depressurised. Isolation discipline is its own subject, covered in the lockout tagout guide.
• Not using an available control. Standing on the top rung, not clipping on where an anchor exists, not wearing hearing protection that is in the van, not using the extraction that is already fitted.
• Improvised access or lifting. Climbing on a valve or a pipe instead of using a platform, using a pallet on forks as a work platform, lifting by hand what should be lifted mechanically, using a sling of the wrong rating because it was the one to hand.
• Working outside the method. Doing a task in a different order than the method statement, skipping a step, carrying out work the permit does not cover, extending scope without going back for authorisation.
• Shortcuts under time pressure. Not fitting temporary barriers for a quick job, not testing for dead because the panel was tested an hour ago, leaving a tool in a running machine rather than stopping it.
• Omissions of assurance. Not checking a pressure test before applying it, not confirming an isolation with the operator, not inspecting access equipment before use.

Read that list again and something should already be uncomfortable. Nearly every item on it invites the question "why would a competent person do that?" and nearly every plausible answer is a condition, a design or a system, not a character flaw. That is the whole argument of this article, and we will get to it properly in a moment.

4. The two side by side

The clean contrast, before we complicate it:

The last four rows are the ones worth sitting with. A programme that finds mostly acts is a programme whose corrective actions are cheap, individually targeted, and will not hold. A programme that finds conditions is a programme that spends money and stays fixed.

5. Why the distinction is genuinely useful when used properly

I do not want to knock the categories down before giving them their due, because used properly they earn their place for one solid reason: the two call for different kinds of response, and confusing them produces the wrong response.

A condition is fixed by changing the physical thing. There is nothing to discuss about a blocked fire exit beyond who clears it and where the stored material is going instead. Treating a blocked exit as a behavioural issue, briefing the workforce about egress awareness while the pallets stay where they are, is an absurd outcome, and yet it happens, because a briefing is faster than solving a storage problem.

An act, handled well, is addressed by understanding why the behaviour made sense to a competent person at the time and then changing whatever made it sensible. That is a different investigation and a different set of possible fixes. Treating a behaviour as a physical defect gets you nowhere either, because there is no part number for it.

The second real benefit is coverage. An organisation that only ever records conditions is looking at the plant and not the work. An organisation that only records acts is looking at people and not the plant. Recording both, deliberately, is how a proactive programme sees problems while they are still cheap. That is the same logic that makes near miss reporting valuable, and the near miss article owns that ground in detail.

6. What each observation actually tells you

A more useful way to read a set of observations is to ask what each one reveals, rather than which box it fell into. Illustrative examples:

Every row in the lower half is classified as an act and explained by a condition or a system. That pattern is not a coincidence, and it is not a rhetorical trick. It is the normal shape of maintenance work.

7. Where the distinction collapses, and why always in the same direction

Here is the substance. The categories look symmetrical. They are not.

Most unsafe acts are made likely by a condition, a design or a system. Someone climbs on a valve because there is no access platform. Someone bypasses an interlock because the machine genuinely cannot be commissioned with it engaged. Someone works without isolation because isolation takes an hour, the job takes five minutes, and the schedule allowed five minutes. In all three cases, "unsafe act" is a factually accurate classification and a useless one. It records what happened and hides why it happened, which is the opposite of what a safety record is for.

Conditions are created by acts and decisions too, just earlier and usually further up the organisation. A guard is missing because somebody did not refit it and somebody else signed off the job. A ladder is damaged because the inspection regime is not running. An exit is blocked because a storage shortfall was raised three times and never resolved. So a condition is very often an act whose author has left the room. That observation is fatal to the comfortable idea that conditions are impersonal facts of the plant while acts are personal failings. Both have authors. The difference is which author is standing in front of you when the form is filled in.

The classification is therefore a choice, and it determines the fix. This deserves to be said plainly: whoever fills in that category field is effectively choosing whether the organisation spends capital or has a word with somebody. Categorise as condition and you generate a job, a cost and a durable change. Categorise as act and you generate a briefing, a signature and nothing that survives the next schedule squeeze. That is a budget and strategy decision disguised as a tick box, and it is almost always made by whoever is nearest, fastest and least empowered to commit money.

8. The same event classified both ways

This is the centrepiece. The events below are hypothetical and illustrative, deliberately ordinary, and each is written twice: once as an unsafe act and once as an unsafe condition or system finding. Nothing about the event changes. Only the classification changes, and with it everything that follows.

Look down the right hand column. In every pair, the act classification produces an action aimed at a person and the condition classification produces an action aimed at the work. The act actions are all cheap, all fast, and all of them leave the situation exactly as it was, which is why the same finding reappears. The condition actions cost money and hold. If your corrective action log is full of briefings and reminders, that is not evidence of a behavioural workforce. It is evidence of a classification habit.

9. The statistic you should stop repeating

Search this topic and you will very quickly meet a confident claim that some specific proportion of all accidents is caused by unsafe acts, usually attributed to an early twentieth century researcher and quoted to the percentage point. You will find it on training slides, in inductions, and in vendor material. I am deliberately not repeating the figure, because I would be repeating a number I cannot stand behind, and neither can the people quoting it. There are three separate problems with it.

First, the provenance. The figures come from old work whose method and whose definitions of "act" and "condition" are contested, and they are quoted today far outside any context in which they were gathered, across industries, technologies and regulatory environments that did not exist when the work was done. A number with that history is not a fact about your plant.

Second, and worse, it is circular. The claim is generated by the very classification scheme it appears to validate. If an investigator is asked to sort causes into "act" and "condition", and a person was present at essentially every event worth investigating, then acts will be found at essentially every event. The scheme guarantees the result. You could run the same exercise with categories of "decision" and "hardware" and produce an equally impressive proportion of decisions. A ratio produced by a classification cannot be evidence for that classification.

Third, and this is the part that matters operationally, a statistic of that shape functions to locate blame at the sharp end. Whatever its authors intended, a headline claim that the overwhelming majority of accidents are caused by what workers do has a predictable institutional use. It justifies investment in behaviour and observation over investment in plant and design. It makes the people nearest the hazard the natural object of intervention. And it gives anyone resisting an engineering spend a citation to reach for. That is why the number persists long after the method behind it stopped being defensible: it is useful to somebody.

For what a disciplined causal analysis actually looks like, the after the event work belongs to the incident investigation guide. It is worth noting that a published international standard for root cause analysis does exist: IEC 62740:2015 "Root cause analysis (RCA)", adopted in Europe as EN 62740:2015, which sets out principles and process steps and describes a number of recognised techniques. It is a voluntary international standard and law nowhere by itself. The relevant point for this article is that it deals with after the event analysis and explicitly excludes assigning responsibility or liability. The distinction between explaining an event and allocating blame is built into the standard, which is exactly the discipline the act and condition split tends to lose.

10. Why the hierarchy of controls sits directly on top of this

This is the connection that turns the whole discussion from a taxonomy argument into a practical one. The hierarchy of controls ranks control measures by how much they depend on people behaving as intended, from elimination at the top down to personal protective equipment at the bottom.

The hierarchy is a principle rather than a standalone standard. It is required by ISO 45001:2018 as amended by Amd 1:2024, clause 8.1.2, which is a voluntary and certifiable international standard rather than law in any jurisdiction, and whose tiers run from elimination, through substitution, engineering controls and reorganisation of work, administrative controls including training, to adequate personal protective equipment. It is also required by ANSI/ASSP Z10.0-2019, section 8.4, a United States voluntary consensus standard. It is described by NIOSH on free public pages, and NIOSH is a research agency with no regulatory power. Worth stating accurately because it is misquoted constantly: US OSHA does not codify the hierarchy of controls in any regulation. Some US general industry standards express a preference for engineering and administrative measures over protective equipment, but the term itself is not defined in the regulations. Nobody should write "the hierarchy of controls standard" or "as defined by OSHA". The hierarchy of controls guide works through the tiers properly.

Now put the two ideas together. When you classify something as an unsafe act, the corrective actions available to you are almost all in the bottom two tiers: a procedure, a briefing, a training refresh, a protective equipment reminder. When you classify the same event as a condition or a design finding, the top three tiers open up: eliminate the need to go up there, substitute the task, engineer the access, redesign the isolation.

That is also why the act classification feels so productive and delivers so little. Administrative and behavioural controls are the weakest tier precisely because they depend on people performing correctly under pressure, every time, indefinitely. A corrective action log dominated by them is a log of controls designed to decay.

11. Behavioural observation programmes, treated fairly

It would be easy to read all of the above as an argument against watching how work is done. It is not, and I want to be careful here, because behavioural observation gets dismissed too readily by people who have never had to close the gap between a written method and what actually happens on a Tuesday afternoon.

Watching work as it is really performed, with the people performing it, is genuinely valuable and is often the only way to see that gap. Procedures are written by people who are not doing the task, in conditions that do not exist, with access and tooling that may be theoretical. The only way to find out that the method statement assumes a platform that was removed in a refurbishment three years ago is to go and watch somebody try to follow it. Done with that intent, observation is one of the most informative activities available to a safety function, and it surfaces design problems that no desk review will ever find.

The failure modes are specific, and they are not inherent to the idea:

• Counting acts as a performance metric. Once a number of observations, or a proportion of safe behaviours, becomes a target, the number is what gets managed. Observers learn what result is welcome and the data stops being information.
• Observation experienced as surveillance. If the person being observed believes the output may be used against them, they will work differently while observed, which destroys the only thing the programme was for. An observation programme that changes the behaviour it is trying to see has no data.
• Observing only the sharp end. Programmes overwhelmingly observe the people with tools in their hands and almost never observe planning meetings, scheduling decisions, permit authorisation queues, procurement of access equipment or design reviews, all of which shape the behaviour being recorded.
• Volume without engineering change. A programme that produces a very large body of observations and no physical or design changes has become an administrative activity. The observations were collected, categorised, reported and filed, and the plant is unchanged.
• Findings that only ever run one way. If your observers only ever find unsafe acts, that is a fact about the programme, not about the workplace. It means the form, the training or the incentives have made conditions invisible to the people doing the looking.

None of those are reasons to stop observing. They are reasons to design the programme so that a conversation about why the work is done that way is a legitimate output, and so that "the equipment makes this impossible" is a finding the form can actually hold. A good observer spends more time asking than ticking, and the best output of an observation is frequently an engineering request.

12. How to record these so the record is worth having

The practical advice here is short and it is mostly about resisting the form.

• Describe what was seen and the circumstances, before assigning anything. "Technician on the valve body reaching the actuator, no platform available, job scheduled for twenty minutes" is a usable record. "Unsafe act: working at height improperly" is not, because everything that would let somebody fix it has been deleted.
• Capture what made the behaviour reasonable at the time. Not as an excuse, as data. Time available, access available, tooling available, what the alternative would have cost. This is the single highest value field on any observation form and it is usually absent.
• If a category is required, treat it as one field among several and not as the conclusion. Many systems demand the field. Fine. Just do not let it be the thing the corrective action is derived from.
• Avoid naming individuals in the record where the finding is about the work. If the record needs a name to be actionable, ask whether the finding is really about the person.
• Trend the circumstances, not the categories. Five separate observations of people improvising access in the same plant room is one engineering problem, and only the circumstances field will let you see that.

And the organisational point, which is worth being blunt about: a reporting system in which "unsafe act" is the commonest finding will generate corrective actions aimed at people, which is the weakest class of control there is. You can read the category distribution of a safety reporting system as a forecast of where its corrective actions will land, and therefore of how durable they will be. If you want the pattern to change, the categories are the wrong place to intervene. Change what the form asks and what the corrective action process is allowed to conclude. The corrective and preventive action article covers the downstream process, and if the engineering requests that come out of it are simply queueing, that is a backlog problem, which is its own subject in the maintenance backlog and downtime tracking guide. Where the record is made in a maintenance system rather than a safety one, the same discipline applies to cause coding, as set out in the failure codes structure guide.

13. The language problem

There is a narrower issue worth naming. In ordinary English, "unsafe act" carries blame. It reads as an accusation whatever the procedure says about no fault reporting, and people are reluctant to file a report that will be read as an accusation of a colleague or of themselves. That suppresses reporting, and suppressed reporting is a worse outcome than an imperfect category.

For that reason a number of organisations have moved to more neutral wording, describing what was observed about the work rather than labelling a person's behaviour. I think that is a reasonable change and I would support it. I would also be honest that renaming a field changes nothing on its own. If the response to a neutrally worded observation is still a briefing aimed at the individual, the workforce will correctly conclude that only the vocabulary moved, and reporting will not recover. The wording follows the response, not the other way round. If you are running a toolbox talk programme and every observation becomes a talk, the vocabulary is not your problem.

14. The maintenance case in particular

This audience deserves the specific version, because maintenance work concentrates every weakness in the act and condition split.

Maintenance is performed on equipment that was designed to run, not to be maintained. It happens in the spaces left over after the plant was laid out. It happens under time pressure, because the asset is down and somebody upstream is counting the minutes. It depends on access, isolation, spares and information that are frequently all four inadequate at once. The result is a completely predictable pattern: maintenance shortcuts are the downstream consequence of access constraints, isolation constraints, spares constraints and time constraints, and the shortcuts recur in the same places because the constraints are in the same places.

Two examples of the pattern, which every planner will recognise. Working at height in plant rooms produces improvised access again and again at exactly the items that were installed without maintenance access in mind, which is a design finding rather than a discipline finding, and the working at height article covers the controls. And isolation shortcuts cluster around assets whose isolation arrangements are disproportionate to the routine tasks performed on them, which is a plant modification opportunity sitting in plain sight in your observation data.

That framing changes how you read your own data. A cluster of behavioural findings in one plant room is an engineering survey waiting to be commissioned. A recurring isolation shortcut is a modification proposal. A repeated permit scope breach is a process throughput problem. None of those are visible if the observations were filed as acts and closed with talks. For readers coming to this from the wider safety function, the HSE overview sets out where this sits in the whole.

The idea to walk away with

An unsafe condition is a state of the workplace. An unsafe act is something a person does or fails to do. Both definitions are sound and the difference between them is worth knowing, because a physical state is fixed by changing the thing and a behaviour is addressed by understanding and changing what drove it.

What matters more is that the boundary between them is a choice, not a fact, and that the choice determines whether the organisation spends money or has a conversation. Most acts are made likely by a condition, a design or a system. Most conditions were created by an act or a decision made earlier by somebody who is no longer in the room. Once you see both of those, the two boxes stop looking like a taxonomy and start looking like a fork in the road, with the upper tiers of the hierarchy of controls down one branch and a briefing down the other.

Final thoughts

If I could change one thing about how this distinction is used in practice, it would not be the definitions and it would not be the vocabulary. It would be the order of the questions. Ask what was seen and what the circumstances were, then ask what made the behaviour reasonable, and only then, if the system insists, choose a category. Done in that order, the category is harmless, because the information needed to fix the real problem has already been captured. Done in the usual order, the category arrives first and quietly deletes everything that would have made the record useful.

And be suspicious of any programme whose findings run overwhelmingly one way. A workplace where nearly every problem turns out to be somebody's behaviour is not a workplace full of careless people. It is a workplace where nobody is looking at the plant, the design or the schedule, and where the corrective actions are cheap for a reason.

Related reading: Hierarchy of controls, Near miss meaning, examples and reporting, Incident investigation, Hazard identification methods, Hazard vs risk, Lockout tagout, Working at height.

Primary sources: ISO , IEC , NIOSH , US OSHA .