Artificial intelligence – clearing the fog

Potential benefits

The natural reflex of some is to look for the possible negative impacts of technological change, but AI can bring positive benefits too. These include:

• building a more thorough discovery technique into project preparation
• providing more thorough and repeatable discipline in data analysis
• using human factor analytics to inform incident investigations and reports
• gauging the impact of training by tracking language use in documentation and changes in behaviour patterns.

Many practitioners conclude that using AI systems to analyse data could provide three potential advantages. Firstly, its use on analysing company data could provide insight into more effective risk controls or even to justify that existing ones are proportionate and effective.

Secondly, huge amounts of data exist, AI can be programmed to find correlations in minutes that a human might labour for weeks to achieve. Used well, correlations found by AI can better support business decision-making. Without it, business leaders may have to make key health and safety investment decisions without robust evidence, using ‘gut feel’ to help justification.

Thirdly, it provides the prospect of perhaps aggregating data at an industry level to extract more accurate insights, but this presupposes that different businesses capture and store equivalent data. Data in every business is accumulating and growing exponentially. The more you have, the more it potentially confuses. Data can be stored in many places, in more than one format, or utilise different business and technical language. AI can overcome this hurdle.

White space

Traditionally, most safety and health analyses focus on incident reports – on the records of when things have gone wrong. But what about the opportunity to analyse data collected when things go right? AI algorithms can be written to examine the ‘white space’ this data creates. It might include looking across audits, inspections, risk assessment records, safety observations, specification documents, etc.

This is a much bigger data set than incident reports. AI can look across all this data – in words, number sets or both – and find correlations. It can be programmed to look for patterns, frequency, clusters of data or text. For example, one large UK supermarket used it to test and prove the hypothesis that there was a correlation between incident rates and management sickness absence. Looking at a few stores indicated that this may be the case, but analysing data from thousands of stores over several years provided confirmation.

Finding an interpretation

There remains a significant limitation, though. A machine cannot find facts. Its correlations are only an expression of signal strength – blowing away the fog of confusion and negating the need for intuitive navigation. But AI cannot interpret what the output might actually mean, meaning there is still a role for the human in this!