Construction Management is the highest circulation construction-based publication serving the UK built environment.
Digital Construction
How to evidence your use of AI
Justin Stanton Editor, CM Digital
Image: BiancoBlue | Dreamstime.com
Construction professionals using AI should keep a detailed record of how they applied it to arrive at a decision. That was the advice from Ian Yeo in a presentation at Digital Construction North in Manchester yesterday (18 November).
Yeo, founder and CEO of PlanOps and former CEO of Operance, pointedly titled his presentation: AI without a record is a liability. He noted that construction professionals may need to provide evidence of how they used AI to regulators (the Health & Safety Executive, for example), an insurer or an opposing party in a dispute.
He set out the five key elements that professionals should record:
This is not a paywall. Registration allows us to enhance your experience across Construction Management and ensure we deliver you quality editorial content.
Registering also means you can manage your own CPDs, comments, newsletter sign-ups and privacy settings.
Request and inputs – instructions, source revisions, and the content actually supplied to the model.
Generation – what tool was used, what model was used and on what date.
Output – the original output and the final issued revision.
Review – the named, competent reviewer and the sources they used to check the output, their findings and any changes.
Approval and use – who approved it, when and for what purpose.
There’s no need to use this process for minor, simple uses of AI, Yeo explained. “I don’t think everybody should do this for everything,” he said. “If it’s just an email that needs to be polished, that’s fine: we don't need to record that. But if it’s a decision that’s based upon information from AI, and it’s to do with health and safety, with design, with compliance and it may have a real commercial effect, I think we should keep a record of that information.”
Furthermore, he recognised that recording and reviewing in this way manually will be laborious, and the process will need support from automated tools.
Not a tick-box exercise
He noted that the review of AI output should not be a tick-box exercise: “This needs to be a proper review. It’s not just a couple of seconds to look at that information. We need to understand what was provided at the start, what we’re looking to get out, and importantly, where the gaps are.
“Reviewing AI output is actually quite difficult, because we get polished documents. They look really good – it convinces us that it’s a good output. In my days on site, you could tell there was something wrong with some information you were provided with. It had a bad smell about it. With AI, we don’t get that. We need to do a proper review. It’s about going back to training and competency. People need to understand that and know what they’re looking for and see the tell-tale signs.”
As a takeaway, he provided this five-question test for the audience:
Can you see what the AI was asked, and what it was given?
Can you retrieve the exact output that was used?
Is the final version in the project record?
Is there a named, competent reviewer, with sources and intended use checked?
Would you be comfortable showing the whole trail to a regulator?