Generative AI is the most important development in IT and in software/hardware in my lifetime. It is game-changing. Just about every aspect of every business is going to have to be rethought in the context of generative AI’s impact on work.
One burning question every company faces is: What happens when machines can do jobs you never considered they could do — faster and perhaps even better in most use cases? In my field, IT, for instance, I can envision a day when a generative AI model will be as good at resolving IT issues as most systems engineers. Perhaps there’s even a point at which the large language model (LLM) becomes better than your best systems engineer. (No offense to all the brilliant systems engineers out there!)
There is something standing in the way of this trajectory, however: Trust. Indeed, it will take time to build confidence that generative AI models will serve up accurate, relevant answers to solid natural language prompts.
Two reasons for lack of trust in AI
Two pressures in particular are weighing on this crucial trust issue:
1. Uncertainty about what’s happening “under the hood.”
This year, I think there will be challenges related to trusting the models because we do not fully understand how they work. Don’t get me wrong — we computer scientists, data scientists and data engineers understand how these LLMs work on a mathematical level, but we don’t really understand how the model works internally. This uncertainty is especially true when the data used to train the models is partially sourced from the “Wild West” of the unvetted internet. While we love many of the responses we get from models such as ChatGPT, it is nevertheless a bit scary to contemplate how the model arrived at those answers.
Also, generative AI models still “hallucinate.” Accordingly, smart enterprises taking advantage of LLMs still rely on humans to validate responses before acting on the information (and if they don’t, they should). My colleagues and I laugh about this infamous example: “Some species of dinosaurs even developed primitive forms of art, such as engravings on stones.” I’d love to unearth those tablets!
These wildly off-kilter answers are not new with AI models; they even predate November 30, 2022, when the floodgates of generative AI as we now know it broke open. One parallel in my own experience has to do with using voice assistants such as Siri, which (or should I say “who”?) doesn’t always understand my Midwestern accent. I am certain my Boston-based colleagues can relate! Needless to say, I don’t always trust Siri.
The difference, though, is that output from Siri generally does not affect my decision-making in any significant way — except perhaps when I can’t find the best “bey-gull” when I’m in New York because Siri doesn’t get me.
In short, the greater the significance of the AI outputs, the greater trust matters. We, therefore, have a long road ahead in mission-critical industries (e.g., healthcare, manufacturing, energy) before AI models can be fundamentally trusted.
That brings me to the second challenge.
2. The depth, breadth, history and quality of data.
The other pressure that calls up the trust issue is data. Not all gen-AI models are (or will be) trained on good data. But that is changing fast. As my company’s CEO, Dave Keil, recently explained in this Fast Company article, “The breakthrough technology before us is not necessarily the foundational models built on complex neural networks. Instead, the real ‘disruptive innovation’ of generative AI is the data itself.”
Why? Because when generative AI models are fed the right data, they can achieve greater and greater accuracy, breaking down this second obstacle across industries where data has been coursing for years. So, in industries such as IT, the trust gap is narrowing by the minute, as companies like mine have been sitting on mounds of expert-level data for decades.
Having a lot of data is one thing; however, understanding context for what is right and wrong —to generate data insights — is another. As McKinsey analysts explain, not only do LLMs sometimes hallucinate, but “… the underlying reasoning or sources for a response are not always provided. This means companies should be careful of integrating generative AI without human oversight in applications where errors can cause harm or where explainability is needed.” That’s what is called the “human-in-the-loop.”
If we combine depth, breadth and history of data (to feed the model for greater accuracy) and human expertise (to provide context and explainability), companies will start to garner greater trust in AI models. I expect that trust to grow exponentially if the data and insights are robust. Admittedly, I am a bit biased about this equation because the company I founded more than 25 years ago was built on data fidelity. Indeed, these are exciting times!
Great data will separate the winners and losers
Across every industry, as companies everywhere digitize their traditional products and/or go digital to enhance the customer experience, those who sit on the most data will gain the competitive edge. Without question, that will be the difference between winning and losing among companies leveraging AI and machine learning. It will take both huge amounts of data and accurate context for that data to ensure that industry-specific AI models are continuously trained properly, in turn becoming more and more accurate over time.
That said, I think it is imperative for business-centric AI models, including generative AI ones, to harness cleaner, more relevant data. Otherwise, we’ll all be left thinking that dinosaurs were the da Vincis of the Mesozoic Era. Or, worse, that you can’t find a good bey-gull in New York.