When someone claims one AI model is more honest than another, what does that mean? What about when they claim that one model hallucinates less? Or perhaps that one model scores better on harmfulness or helpfulness? Or that it is less sycophantic? Or even that one model performs better?

Constructs are critical building blocks that knowledge claims necessarily depend on. Poor conceptualizations can thus have an insidious impact on the credibility and reliability of scientific or any other type of knowledge claims. In fact, poor conceptual foundations undermine every choice an AI researcher or practitioner makes about the design, development, evaluation, deployment, or use of any AI artifact. And yet, the impact of building on top of conceptually shaky ground is widely overlooked in AI work.

Here is a challenge: pick almost any construct you believe to be currently central to AI work and it won’t take you long to realize it has many competing definitions. Sometimes even conflicting definitions. And if you look more closely for clearly, explicitly articulated definitions for that construct you will rarely find one. This is particularly salient for constructs related to the properties (or if you want capabilities) of AI systems—e.g., what these systems are intended or believed to do, their impacts—e.g., the consequences of deploying and using these systems, and any stated north-star goal for AI work—e.g., how the AI community understands and assesses progress. 

This is often thus not so much about whether we agree or not about those definitions; often, we are not even told what those definitions are. But if one does not really know what it means for an AI system to have a certain property or capability, how can this property or capability be reliably measured or computationally represented?1 And if we do not know what it is—and thus we cannot possibly know whether the measurements or representations capture anything meaningful about it—then how trustworthy can claims about an AI system having that said property or capability be? One should at least be very skeptical.

It also does not matter whether one believes or not that an AI system can even have that property or capability—or what someone’s epistemic commitments might be.2 To vet a knowledge claim, at a minimum, we need to know which specific conceptualization of a construct the claim is making assertions about.

Of course, none of this is new and we have a lot to learn from other disciplines about the importance of conceptual clarity. As someone who thinks quite often about evaluation and evaluation practices—as foundational to understanding the impacts of AI systems—I just think this is so critically important, yet still so startlingly neglected.

While a lot has been written on this topic, for more on poor conceptualizations and their consequences, I recommend starting with these short papers (from psychology and marketing research):