Why AI is not the right choice for Psychosocial Risk compliance documents

Many of us have noticed and responded to the need to learn to use AI to keep up with our competitors and meet expectations for productivity in our businesses. Most AI that we use for work are LLMs, Large Language Models. They feel so personable, and they can learn our preferences, share jokes and they write what feels like thoughtful and intelligent responses.

“Feels like.” That is the key, because what LLMs actually are are language prediction tools that are trained on groups of people having conversations and giving feedback until the average person would rate the responses as “empathic”, “funny” or “academic”. Having marked many first-year assignments at UniSQ, I can tell you that what a new student thinks of as “academic” is usually far from how a good academic writer uses language. AI “training” has a number of stages, and in the first stage it is trained on large data sets including millions of academic source documents. The LLMs learn to produce writing that is actually academic in tone, but when they go through Reinforcement Learning from Human Feedback (RLHF) these biases for what “academic” “feels like” or “looks like” to the average human user, they are trained to distort academic outputs to something that reads like a very poor caricature to an actual academic.

Why do you care about this when you are in business, I hear you ask?

If you ask an LLM to help you understand psychosocial risk compliance, the response will be, at least in part, wrong. Unless you are already an expert, you likely won’t be able to identify those mistakes. The AI is designed to sound like it knows what it is talking about, making you believe it has answered correctly when it has not. It will even “hallucinate” information, or make up a reasonable sounding answer, when it can’t find an answer, rather than saying it doesn’t know. This is not considered a flaw by the developers; it is a design decision. AI are not entities or agentic under legislation: they are not people and they won’t be the ones in trouble if the mess up, as PCBUs and decision-makers, you will be the liable party in both criminal and civil contexts.

I have benefited greatly from learning how to use AI while setting up my business. I write my own blogs and do most of my own technical writing. I get editing ideas from my AI then I refine the edits further. I will ask the AI to generate reflective points about topics when I already have a list of ideas (undisclosed to the AI, so as not to bias the response) then I contrast, compare and interrogate the AI about its sources, logic and intent if we disagree. This highlights my knowledge gaps, but it is based on my skills, knowledge and training as a foundation to monitor the AI outputs. Side note: AI is also very helpful if you don’t know how to do something specific in a software package, particularly when learning how to use it.

There are AI that use visiospacial processing and physics to handle safety features, like those in cars, and they create a “halo effect”, where users associate all AI models with being “safer” than humans. What you need to know is that LLMs and the AI in cars and such are as different as slugs and horses in the animal kingdom. They function completely differently and have different safety parameters. AI lacks the understanding of human behaviour and psychosocial dynamics in the workplace to actually help you with psychosocial risk compliance. It will find the first likely hits in a search then apply averaged information from its training to give you an answer you won’t assume is fake. This will not pass audit, but worse, it could lead to life-changing or life-ending harms in your workers, and I know you don’t want that to happen.

There are some great resources available to help you through the compliance process. If your business can’t afford a large and complex compliance process, we can still discuss a more affordable process that will suit your business needs that doesn’t leave you relying on tools that are not appropriate for this task. Reach out and we can book in a time to chat about how I can help you with this.

Taz Clifford
Principal
IDEA HR Consulting