The Hidden Gaps in AI-Powered Safety Solutions

AI-powered solutions might not be capturing the most important hazards.

• By Carrie Taylor
• Apr 04, 2025

What the makers of ergonomics AI don’t want you to know

Let me start by disclosing that I lead a team of ergonomists, and in 35 years of practicing ergonomics I’ve seen a lot of technology come and go. Our team uses biomechanics software to evaluate the loads on a worker’s body, and we incorporate fatigue research to account for the effects of repetitive or prolonged tasks.

Our team prides itself on being early adopters of technology; every year, we review the available tech, and decide what to license for the upcoming year. When you search for ergonomics applications for AI (artificial intelligence), you’ll find numerous programs that use markerless motion capture and software to automate the ergonomics assessment process. We’ve been watching this technology, always hopeful that it will help to automate our process so we can do more, faster. Every ergonomist we have met would prefer to spend time solving problems, rather than studying them.

What Professional Ergonomists do

Bear with me as I review the process that we use, some of which is pretty tedious. First, to select tasks for analysis, we survey employees and supervisors, and screen the physical demands analysis for hazards. If a few tasks are flagged, we can focus on those. If the job involves a lot of demanding tasks, we know we’re going to have to look at all of them. For each task, we:

• Photograph or videotape the task
• Measure the height, reach, and sideways locations of the hands, relative to the feet
• Measure the efforts applied by the hands. These can be weights, pushes, pulls, pinches, or grips. Most efforts require us to take at least five readings. Some tasks require us to assess a “typical” effort and a “peak” effort, such as when parts are not fitting well, or they are cold, or there’s a section of carpet that a wheelchair has to cross.)
• Time the frequency and duration of the efforts applied. If the task has some variability, we need to take multiple measurements and calculate averages.
• Input all of this data into software, which tells us how hard that task is for the “limiting user”. (If this worker is safe, stronger workers will also be safe. The selection is based on the worker population, but we want to protect at least 75% of workers.) We currently use “Work(s)” software.
• If we’re concerned about cumulative demands, we have to repeat this for all efforts in the job, and combine them in a cumulative model.

What does the AI do?

Markerless motion capture is the currently “hot” technology in the safety world, often combined with machine learning programs to create tools for ergonomics assessment. It allows a user to analyse a short video, to calculate how long a worker spends with the arms and legs flexed forward.

The most common approach we’ve encountered is for the tool to compare exposure data  from a video with an assessment tool that was created in 1993 (RULA, authored by McAtemney and Corlett). The tool is based on the premise that repetitive awkward postures are associated with higher risk of injury, especially if combined with high force requirements. This model was meant as a screening tool – it categorizes risk on a 1-7 point scale, with 1-2 meaning “negligible risk, no action required”. It included two force thresholds: 2 kg and 10 kg. It’s rare to encounter a job with a low risk score.

If the job involves lifting, lowering, pushing forward while walking, pulling back while walking, or carrying, some AI tools can analyse the demands based on the 1991 analysis tool authored by Snook and Ciriello that showed how much weight people felt safe handling, under various posture and frequency conditions. The conditions in the study were all two-handed tasks, handling weights or forward pushing or backwards pulling, with no twisting, and good grips. Pushing and pulling always included walking.

Still other AI tools apply unidentified analysis tools, so the user must simply trust that the developers have found and applied appropriate tools.

AI tools create beautiful, simple, colour-coded reports that show a risk score.

What does AI currently not do?

None of the AI models that we have found are good at assessing loads on the body in three dimensions. This means that they almost always ignore or underestimate risk associated with sideways bending or reaching, twisting, one-handed loads, heavy gripping or pinching, and pushing in any direction other than forward or backwards. (Ask yourself how many jobs involve only forward/back/up/down movements.)

The “force” hazard assessment is rudimentary at best. Most of the programs allow you to “skip” the force data altogether, or to estimate the force. This feels unacceptable, because force is arguably the most important factor. The amount of force, the direction of force, and the type of grip used, are all critical in risk assessment.

Ask yourself this:

• If you were required to bend to pull an empty bottom file drawer toward your body, would you be at a high risk of injury? (Likely not.)
• What if the drawer was full of books and the pull force was 30 lbs?
• What if the drawer’s one small handle allowed only two fingertips to pull?
• What if there was not enough space to stand facing the drawer, and pull back toward your body? Would your risk be higher if you stood beside the drawer and pulled sideways, with two fingertips?

If you don’t understand the load, the direction of the load, and the way the operator interfaces with (grips) the load, you cannot properly assess risk.

I was curious, so I asked one of our ergonomists to run the variables above through our analysis tool (Work(s) Ergo). Work(s) uses a biomechanical/fatigue model to calculate a “demand/capacity ratio”. If this ratio is greater than one, the task is not acceptable. Our goal is for tasks to be acceptable for at least 75% of the worker population.

Here’s what we found:

Video analysis captures how repetitive a job is, if a full cycle is assessed. How often you repeat a task is important in risk assessment, but I’d still rather pull the empty drawer 25 times than pull the full drawer sideways with two fingers, once. And if you don’t account for the force, the currently available tools will report the repetitive empty drawer as much higher risk.

AI can’t do a great job at evaluating demands on the elbows, wrists, or hands. It’s tough to assess these joints by video, because they are often obstructed by clothing, the worker’s body, or other equipment. If your workplace includes a lot of hand-intensive work, markerless motion capture will be disappointing for you.

AI can only assess the risk for the worker who is actually shown in the video, rather than for a “small worker” or “tall worker”. You can’t ask, “Here’s what I recorded for a 5’10” worker – what would this look like for a 5’ 2” worker?” AI cannot predict the impact that a proposed change will have on injury risk, because it cannot assess a job without a video. You can’t say, “If I lower the bench by 6”, will the risk be reduced to acceptable levels?” This is important! A good ergonomics assessment should provide the client with confidence that the proposed change will mitigate the risk. The client should be able to understand the impact of reducing the drawer’s pull force, allowing better clearance to stand in front of the drawer, or optimizing the handle design, so the most effective solution can be selected.

Most of the AI tools are running analyses in a “black box”. Only the developers know exactly how the tool is assigning the score based on cumulative loading, or how it is factoring in the force direction and magnitude.

When the developers of these tools talk about “validation”, they focus on how well the software captures the actual movement patterns in the workplace. To be honest, it’s pretty remarkable that we can get decent accuracy here – when I was in university, we needed three cameras and reflective markers stuck onto people wearing bodysuits to capture video data. And we had to analyse the video one frame at a time, clicking on each marker with the mouse. The fact that we can measure arm and back positions using only one camera is impressive. But true validation occurs when we can say that the tool predicts the risk of strain/sprain injury, and none of the tools have accomplished that yet.

I’ve talked about my concerns with a lot of my colleagues. Ergonomists in a variety of private and public organizations have expressed concerns about the use of tools that have not been validated (no one really knows if a high score has been associated with a high incidence of injury), cannot be openly examined, and are being widely used without appropriate force measurement. We’re also concerned that organizations adopting this technology without the appropriate ergonomics expertise will get frustrated and throw up their hands. This will cast shade on ergonomics in general.

If you’re considering adopting new technology, make sure it will do what you need it to do, before you commit. For our part, we want AI to:

• Accurately and simply gather all the data that we need.
• Use an accepted, science-based model to account for exposure to repetitive or sustained, forceful, awkward body positions, including hands, wrists, elbows, shoulders, neck, and back, in 3 dimensions.
• Report risk as one value that allows us to prioritize multiple projects.
• Help us to evaluate potential solutions for high-risk tasks. The model should be sensitive enough to account for small changes in hand position, force, frequency, and task duration.