All of the knowledge and none of the experience – the limits of AI

If you have read my last three posts, you might think that I am 1) an incredible fan of AI in Construction and its opportunities and 2) overly optimistic about AI in its current state. I will absolutely admit to the first, but not to the second. AI, particularly large language models (LLMs), are still proving their worth. There continue to be limitations for AI tools that we must consider when taking a strategic approach to their adoption in construction.

Before launching into what I see as the primary challenges for AI tools, we need to distinguish between pre-trained machine-learning (ML) models and LLMs. The way they are designed makes them optimal for specific uses, and less-than-optimal for others. ML models are trained on a specific set of your data to produce consistent outcomes. In areas like manufacturing and quality control, ML models are a great choice, as they not only identify defects, they can also become predictive with enough data trained by the model. They create fixed rules. This design is also ML model’s greatest limitation. They do not behave well in situations where the rules change and complexity is very high. ML models are terrific at chess, but not great at situations (like job sites) where conditions are much less predictable and rules change.

LLMs, on the other hand, are trained differently. Their training isn’t on the specific application at hand (a robotic arm welding two pieces of steel together and determining if the weld meets a quality standard), but on the variety of texts in the electronic world, incorporating millions of different perspectives and stories. In that way, the LLMs “understand” the randomness of the world and can provide these perspectives for you. My third source of value for AI is this ability to provide perspective as a “virtual thought partner.”

But this source of knowledge is also the LLM’s greatest weakness. An LLM has often been compared to a well-read intern, with a lot of knowledge, but very little experience. I’d suggest that they are more like the marines in Aliens. In the movie, Ripley is riding in the jump ship with a team of space marines on the way down to the colony that was overrun by aliens and asks “How many drops is this for you, Lieutenant?” He answers “Thiry eight… simulated.” ?I would suggest that LLMs are all knowledge and no experience. What lets us make good decisions in the field is the combination of what we know and what we have done. That’s why we can’t let LLMs make decisions on how to update a schedule. It knows thousands of ways to update a schedule but doesn’t know which one is right for this situation. In fact, LLMs often provide different results, including wildly inaccurate ones. It’s not a bug, but a feature of the tool. That’s why an LLM can give you a variety of perspectives but isn’t capable of choosing the right one for this moment.

Finally, there is a common challenge for both LLMs and ML models – quantity of data. For ML models, you need a substantial amount of data to ensure that you haven’t over trained the model (and certainly to move from descriptive to predictive analysis). New innovations in model design are reducing the quantity required, but it remains an issue (it was a much larger issue for me when I started working with these models in 2018). LLMs have two issues related to data. The first affects the specific advice it can give based on your situation. ML models “know” two sets of things – what the base model was trained on (lots and lots of data on the public Internet) – and the prompt, which includes documents and information used to provide specific context as well as the on-going conversation about a particular topic. This context limits how many specifications, scope documents, or plans I can provide for a specific project and requires some “chunking” of analysis of information (like specifications) into smaller groups. Again, this is a technical issue that will likely resolve itself in the future. But there is a larger, potentially more challenging problem with LLM data. As data created by LLMs make their way onto the public Internet, the same LLMs will be training on their own output. If you have interacted with LLMs in a very long conversation, you will find that it can get less and less accurate or fixate on one specific area (we had an LLM decided that all projects must be healthcare projects in one conversation, even though we explicitly told it that it was for education). LLMs being trained on LLM data could spiral in interesting, random, or (potentially) wildly inaccurate ways.

I offer these limitations as food for thought as you build SOPs and guardrails for the use of AI in your company. There are substantial benefits for using AI, but we need to acknowledge its shortcomings, as well. The AI can provide you knowledge, but only the humans can add their experience to discern the right way forward.