AI has no way to determine what is, and is not, copyright material that it scrapes from the net.

Even linking to regulations you will find the AI struggles with interpretation.

Things done well and accurately are done slowly as short cuts always generate omissions.

If you are pursuing this route make sure you have a very large indemnity cover for when those omissions affect peoples lives and livelihoods.

Hi Ardee

Your first post here, so welcome to the Forums.

In addition to important issues (including those of copyright) already pointed to by Roundtuit, trying to get AI to do analysis based on e.g. legislative requirements is difficult even if the AI focuses entirely on legislation AND is capable of working out the latest version of any specific piece of legislation is at any time.

Starting with the premise that you are going to focus on the UK as it gets MUCH more complicated if you venture beyond.......

Let's assume that the AI is reasonably clever and realises that it needs to access the legal requirements via legislation.gov.uk and can work out that Law A applies in England and Wales but not in Scotland and Northern Ireland and DOES look at the latest variant, it might still not have the most up to date version as there might be some recent amendment that the website has yet to capture. Perhaps the on the ball OSH professional is keeping tabs on what is on the horizon and may hence realise that SI 2024 No 5799 has been made and comes into force on 5 May 2025 and so amends e.g. the PUWER Regulations.

Which already leaves a dilemma - do you base your risk assessment on what the law says now, or what it will say on 5 May?

But occupational health and safety law in the UK is much more nuanced than e.g. what it says in OSHA Regulations in the US where almost all requires are very clear. "DO this." "DON'T DO that."

Instead a mix of a very few requirements which set "strict liability" duties, amongst others qualified by "so far as practicable", "so far as reasonably practicable" [or its twin ALARP], all intermixed with a few mentions of "reasonable", "suitable" and other words that dilute the demand.

How is your AI going to decide which legal judgment on the interpretation of e.g. "reasonable practicability" to default to and decide whether or not it needs to at all?

You can, of course, programme your AI to capture everything in every single publication that HSE and other UK enforcing authorities currently have in publication, but will AI be able to find all the stuff that is out of print that might still be relevant? 

AI might also capture all the HSE press releases about successful prosecutions and assume that everything that HSE asserts in those press releases is completely accurate and not recognise that if a defendant has pleaded guilty [not necessarily as the defendant actually thinks they are guilty at all, or thinks that they were not THAT guilty] what HSE might take from that might be far from a true picture of the circumstances - sometimes you have to have been in the Court to know what was actually said and recognise when the HSE press release is not only miles away but contradictory as to the facts AGREED in Court.

Hence, if you place too much reliance on AI to do things like risk assessment, as Roundtuit indicates, you need to make sure that your insurance cover is VERY robust.

best quote I have seen on AI...we were smart enought to invent AI, dumb enough to need it and now too stupid to know what to do with it..

there are aready AI legal models out there and some being used in trials in courts to reduce the time but there are still significant errors...however the human (a competent human) still needs to validate the output...I would say for this application you still have a long way to go to get as bad as some of the RAMS I have seen over the years...  sounds like you are looking afor a quick buck, there isn't as it takes thousands of hours to train and although there are 'free' models out there to start and it is predominatly US based behaviours...yes behaviours.  I have one trained on my process safety work for the last 40 years (personal use) and it still comes out with crap sometimes...but helpful to remind me what I did when I was young and stupid...

AI systems like GPChat are only as good as the information that they have access to. So if  you ask it about the legal issues surrounding risk assessment let’s say, it will be able to find loads of information  on the internet about it ad put together a NEBOSH style answer about the topic but if you ask it to create a RAMS for a particular piece ( a monkey with a typewriter can and are used to write generic RAMS) it will struggle to create anything useful.  So far as is reasonably practical not only includes the official factors such as cost and effort but also an organisation’s appetite for risk. How is AI  meant to judge that?

Is that really so much worse though than "Well, we save down a copy of a RAMS for something else and then change the title, and if we have time we do some search-and-replace of some of the words to make it fit the job"?

I am sure that apps are already available that use AI to support risk management. Even ChatGPT does a remarkably good job at summarising hazards and effective controls considering it is a general AI tool. I can only imagine a specialised app being even more useful at helping people to manage risks, noting that this is not the same as automating the whole process.

AI does get things wrong but one its strengths (illustrated by ChatGPT) is that it gives comprehensive but clear and coherent answers to problems. This makes it easy for people to identify errors. Compare that to human generated text related to any complex issue where it can often be very difficult to understand what the issues are and hence whether a reasonable approach has been followed.

The human will always be responsible. I don't understand why someone using AI will make any difference in a law court to their level of negligence.

Producing a risk assessment (the noun) could be done by AI with the right rules. It's ultimately just text.

Undertaking a risk assessment (the verb) is more difficult for all the reasons already laid out and why competent human judgement and nuance is necessary. 

In real terms, the act of assessing is more critical than the recording of that assessment. Starting with the end document and working backwards is the wrong way around. Although it is already practiced by companies the world over via existing RA model templates, so who would notice? ;)

Then of course you also need to consider the EU considers AI software to be included under its revised Product Liability Directive (EU 2024/2853) the internet having no respect for geographical borders anything available "on the market" falls in to scope.

Harvard Business Review  "AI Won’t Replace Humans — But Humans With AI Will Replace Humans Without AI." Seems like a reasonable way to view most AI applications.

https://hbr.org/2023/08/ai-wont-replace-humans-but-humans-with-ai-will-replace-humans-without-ai

Not sure if this helps but seems to be HSE view on AI and health and safety

https://www.hse.gov.uk/news/hse-ai.htm

Exactly Chris

“Doing” H&S is not just about remembering rules and regulations or regurgitating policies but it is also about applying your life experiences. If you ever go on a training course you want to share people’s experiences, especially those that drive their approach to H&S; the bad accidents, the cock ups the funny stories. AI, no matter how powerful has not  had a life; its “experience” is based on what it has been told usually picked up from the Internet, and we all know how balanced a world view that can lead to.