The Goldilocks Trilemma

You are Goldilocks, the fairy story character who enters a bear’s family house and tries the porridge in each of their bowls to decide which one to eat (possibly challenging but bear with me). A Goldilocks Dilemma it you will. You must make a decision based upon 2 variables – the states of the porridge and your taste preference. And you do. If you plotted a graph of this process, it might look something like this:

So far so straightforward.

Having quickly sorted out your breakfast, you start your day job as a Stakeholder in AI enhanced HAZOP tools. You are new to the field (isn’t everyone), but you sense that not engaging with use optimisation is not an option (probably sensible). This is a rather more sophisticated challenge as not only is there an additional variable (creating your trilemma), but the key one is pivotal in saving future lives.

AI enhanced HAZOP tools have arrived. Kairostech’s AI HAZOP Manager was launched by Vysus in 2023 and has already been adopted by several companies. Several other tools are available. The trilemma variables are:

  1. Tool Autonomy (AIA)
  2. Tool Anthropomorphy (output humanising – AIH)
  3. MAH scenario detection

As an experienced HAZOP chair (100+) and an amateur interest in industrial psychology, I decided to be the putative Goldilocks and have a subjective stab at the trilemma.

First, I generated a few bounding tenets:

  • MAHD reference for conventional, pre-AI HAZOPs would be 100
  • 0% AI Autonomy would represent this state
  • AIA and AIH values would go from 0% to 100% in 25% intervals

This led to the following evolution:

Several conversations over a 6-month period with Alfredo Haubold, Vysus Product Manage with responsibility for Kairostech HAZOP Assistant, helped me to create a subjective scoring:

I even managed to generate a 3D model (please excuse the non-mapped colour cod-ing!):

My (extremely) draft efforts indicate that the best combinations may be:

  1. AI HAZOP tools generate models which serves as contributor to human participants discussion and are humanised to communicate in a courteous, reinforcing and roleplaying way
  2. AI HAZOP tools generate models which serves as prompt for human participants discussion and are humanised to communicate in a courteous and reinforcing way

So far so bespokely qualitative.

As to how to bring some universal qualitativeness to the trilemma, your ideas are as good as mine. However, major issues await:

  • How do you measure MAH scenario detection for either conventional or AI enhanced HAZOPs?
  • How do you test the impact on degrees of AI output humanising on the behaviour and effectiveness of HAZOP teams?
  • How much AI autonomy do you risk in the early stages of adoption?

Goldilock’s Trilemma. Challenging. But potentially so much more satisfying and meaningful than just deciding which bowl of porridge to have for breakfast.