Let’s say your organization wants to roll out an artificial intelligence (AI) tool to help complete some tasks in your workplace.
But you’re concerned that introducing AI might also unintentionally increase risks to your workers’ health and safety. For example, some systems could have been trained on limited data about certain groups of people, and as a result, might produce different outcomes for different groups of workers. Others might have been developed without consideration for factors of the work environment which impact worker health, including the way work is performed, the layout of the workplace or the psychological conditions.
So, what can you do to make AI systems safer for workers?
To answer that question, Institute for Work & Health (IWH) researchers have developed a preliminary framework for how to integrate occupational health and safety (OHS) considerations along the life cycle of an AI system.
Across recommendations, three key insights were consistently highlighted. They included:
- establishing a mechanism for shared responsibility in the organization to oversee AI development and use, for example, an AI committee;
- ensuring transparent communication about why and how AI is being used to foster trust in the system; and
- promoting AI literacy so everyone can participate in discussions about AI.
“There’s a lot of discussion about the risks AI could pose to workers, but there isn’t much advice out there about how to integrate OHS principles into AI development and adoption. That’s even though these systems are being used so widely in workplaces,” says Dr. Arif Jetha, IWH associate scientific director and scientist. He’s one of the authors of the framework alongside Jared Bierbrier, IWH practicum student.
“We wanted to provide practical recommendations for how to make AI safer for workers to address questions we have frequently received from our stakeholders,” Jetha says.
To develop the recommendations, the authors drew on existing academic and non-academic literature and other practical resources as part of the Partnership on AI and Quality of work (PAIQ) project. The paper compiles the researchers’ synthesis of available evidence from these sources about how to make AI healthier and safer. It was published in the American Journal of Industrial Medicine (doi: 10.1002/ajim.70090).
Seven opportunities to make AI safer for workers
The framework provides recommendations at seven points along the lifecycle of an AI system. They each aim to help organizations prioritize workers’ safety when developing the technology.
1. Defining AI system’s purpose
Worker safety should be a key consideration right from the start, when thinking about what an AI system will do. If a factory worker will be working alongside an AI system, for example, will the technology speed up the pace of work to an unmanageable level for workers?
How to promote health and safety:
- Balance efficiency with safety considerations. Ensure the system aligns with the pace of human capabilities, so it doesn’t cause physical strain, mental fatigue or other unsafe outcomes.
- Conduct an impact assessment to see how the technology might interact with aspects of a job or impact workers, and whether the technology could increase the risk of injury.
- Engage in a co-design process with the workers who would be impacted, and with those who are aware of the health and safety legislation that might apply.
2. Acquiring data and preparing to train the AI system
The next phase is to acquire and provide data to train the AI system so it can learn and recognize patterns. Historically, some AI systems have been trained on limited data that excluded certain groups of people, which could result in biased outputs.
How to promote health and safety:
- Gather training data from the specific worksite where the tool will be used so it is most relevant.
- Include health and safety data like hazard exposure, worker characteristics (inclusive of all workers to avoid biasing the system) and incident or injury records.
3. Training the AI system
With the data acquired, developers would then set out to train the AI system on how to do its task. The system should learn to do its task without introducing risks to workers health and safety.
How to promote health and safety:
- Recruit OHS professionals and workers to help identify unsafe AI behaviours, biased outputs (for example, when some worker groups experience poorer outcomes) and other hazards.
- Adjust prompts and training tactics based on the issues identified.
4. Testing the AI system
Next is the feedback stage, where workers from across the organization should have the opportunity to report if the system functions well across the organization, or if it produces errors that disproportionately impact certain groups.
How to promote health and safety:
- Do a health and safety stress test to see how AI performs in safety-sensitive situations (like alerting a worker to a hazard). Test high-risk prompts to make sure it doesn’t generate incorrect or unsafe responses.
- Ensure occupational health and safety professionals are involved in testing.
- Consider having an AI oversight committee evaluate the system for safety and regulatory compliance.
5. Deploying the system in the workplace
When the system has passed initial tests, it is ready to be more fully integrated into the workplace. At this phase, workers should know why the AI system is being implemented, how it affects their safety, and how to voice any concerns. The goals should be to foster trust in the system and to make sure the system benefits all workers with minimal harm.
How to promote health and safety:
- Communicate transparently to workers about what the system is doing, how it makes its predictions and how it could complement their jobs. This is a key step to reduce fears associated with AI systems in the workplace.
- Ensure the AI system is flexible so it can incorporate worker feedback.
- Have OHS professionals conduct risk assessments and recommend controls such as administrative policies, engineering adjustments or personal protective equipment if needed.
- Train staff about AI and the particular system so they can identify issues.
6. Monitoring performance of the system
After it has been implemented, it’s important to monitor the system in case any new issues arise.
How to promote health and safety:
- Have OHS professionals, organization leaders, and/or an AI oversight committee watch for performance declines, new hazards or other health and safety concerns. This is especially crucial for systems that are continually learning and adapting as they operate.
- Review OHS incident reports to find out if AI systems may have contributed to them by failing to work as intended, providing incorrect or harmful outputs or operating beyond its limits.
7. Retiring an AI system
An organization may also choose to stop using an AI system. That could be because it isn’t improving productivity, it’s causing hazards for workers or there are other systems that might better fit the workplace. In this case, organizations will need to support workers through the transition while protecting data privacy and ensuring work can continue.
How to promote health and safety:
- Train workers for whatever might replace the AI’s function, whether that’s a new technology or a manual procedure.
- Archive securely or destroy sensitive data that has been collected by the system.
- Document how the system worked for compliance and reporting purposes, to help inform future AI systems and to help prevent workflow disruption.
“There currently exists few strategies that take an OHS lens to AI system design and deployment,” says Bierbrier. “Our framework provides practical strategies that can be adopted by OHS professionals, workers, organizational leaders and AI system developers across the life cycle of AI development, implementation and eventual retirement.” He notes, however, that the recommendations are based on a broad definition of AI, and recommendations for specific AI systems are still needed.
“This framework is just a first step for us, as this technology is changing fast,” say Jetha. “As part of the PAIQ project, we are hoping to build on these recommendations and provide evidence-based best practices for prioritizing worker safety in AI adoption.”
“Working across disciplinary boundaries to bridge OHS with computer science and engineering, organizational behaviour and human resources expertise will be necessary to make these systems safer and healthier,” he adds.