Hallucinations can be experienced by humans, so why should we be surprised that a machine that crafts language can experience them as well?
For those of you who have gone to our program, you’ll understand that there are risks involved in using Assistive Intelligence. These risks should be considered and understood prior to deploying any applications, especially the ones that are facing customers in the public domain.
Accuracy of data is essential to every business, and is a bare minimum requirement for every accounting department, marketing team, and executive in any organization. Ultimately, when we seek information from software today, we generally trust it to be correct.
Unfortunately, Large Language Models sometimes act like that overconfident analyst in your project, who just sounds very convincing, but actually cannot back up the proposed solutions. The issue is that, as we start integrating Assistive Intelligence into your analysis and functions—particularly those functions that, as described before, require accurate data—there is a risk that the data provided to those teams may, in the simplest terms, be made up.
The disability of Large Language Models to make things up, is often referred to as hallucinations. Children make things up, adults sometimes make things up too, and most stories are technically made up. Large Language Models are trained on language, and amongst all that language, they don’t always understand data or its relations to reality.
This risk of making things up, can be managed when building Assistive Intelligence tools. Just like we can introduce controls for bias, we can introduce self inquisitive loops in the software that we build using Large Language Models.
You don’t need to become an expert from a technology perspective, but as a leader, you have to understand the fundamental concepts that form the building of Assistive Intelligence tools. Better data and their awareness of what that data means is always beneficial, and this is the gap that we want to close with our education programs.
Firstly, we have to recognise that we are part of the laity, then ask for help, and finally be one of the ones that are no longer “non-technical”. You don’t need to become a technical guru, but a fundamental understanding of how these tools are built will go a long way in helping you empower your teams to create the new AI-driven operating model.
