Normative and descriptive Judgements

Remember a few weeks ago, when we posted the beginning of a discussion on how to teach logic and critical thinking, or rather how we don’t teach it anymore?  Those are related thoughts have remained prominent for me since writing that, but it was another forum of intellectualism entirely that prompted this related post, and I think the insight took the paper’s authors by surprise, too.  They set out to study the decision-making process for machine learning/artificial intelligence applications, so that they could be leveraged in such settings as content moderation, and incidentally stumbled upon a new understanding of how our human minds make decisions.  You can read their paper here: “Judging facts, judging norms: Training machine learning models to judge humans requires a modified approach to labeling data.”  The implications of AI/ML decision-making aside – perhaps we’ll do a separate post on that one day – today’s focus is on normative and descriptive decision-making in humans.

That we are learning about our own mental processes from studying and contemplating computers’ is fascinating in itself.  It forces us to probe more objectively the processes and algorithms we implement biologically which might otherwise remain hidden in the background.  In this case, the paper’s authors discovered that our decision-making is significantly more normative than descriptive under most circumstances, even when it would seem that a descriptive technique would better fit.  I would not go so far as to call this a logical fallacy, but it is a characteristic of how human reason functions that highlights a flaw in a rationalist approach to knowledge (watch for a post soon regarding epistemology and the ideas of rationalism, empiricism, and skepticism).

Descriptive decision-making consists of what we typically imagine as an orderly, logical approach to deriving a conclusion.  Presented with a decision to make, data on the subject, and a set of criteria, a descriptive approach will evaluate the data on the subject to see how it interacts with the criteria, and derive a judgement based on compliance with or deviation from those criteria.  This is how a computer model will tend to evaluate a subject, and it is how many of us probably believe we make most of our decisions, especially those about which we stop to think.  After all, compared to this ordered approach, the normative approach seems like putting the cart before the horse.

Instead of examining the data and coming to a conclusion, as in descriptive decision-making, a normative approach will first examine the possible conclusions, the criteria involved, and then what the available data are.  This might seem a minor difference, but it turns out that, in the normative approach, humans are inclined to come up with a conclusion and then examine the data through the lens of their preconceived conclusion, and thence formulate reasons and justifications why the data and criteria support their conclusion.

Examining legal decisions is the perfect forum, and a prime demonstration of the importance and vast implications of these findings, to understand what this means in practice.  Consider how the US Supreme Court renders judgements and decisions.  They vote on an outcome after which the responsibility for writing the opinion which will accompany that judgement is assigned.  Only after a judgement is made is the supporting reasoning established.  While this is less explicitly the case in lower court settings, research suggests that the same kind of normative decision-making occurs in jury trials and other forums, both judicial and otherwise.

This is not a referendum on whether normative decision-making is superior or inferior to descriptive decision-making.  That is a deeper question, and even the paper which remarked upon the discovery noted that it is unlikely to be desirable to transition to descriptive decision-making in traditionally normative contexts without significant philosophical debate.  However, if AI/ML systems and models take a descriptive approach, we should not be surprised that they come to different conclusions from their human counterparts who are likely to take a normative approach.  This informs how such systems are operate, how and where they should be implemented, and even how people will react to them.  But before you begin to think that normative decision-making is a flawed method compared to descriptive, consider that descriptive decision-making is just as prone to bias and achieving a “desired” conclusion – it’s merely a difference of where that bias is implemented.

What is valuable, though, is an understanding of both of these techniques, and what our predilections are for implementing them.  We can account for the flaws in both systems of reasoning if we know which one we are using, but if we think that we’re using a descriptive approach when we’re really using a normative approach, that will represent a blind spot in our thinking and a potential hole in our logic.  It would be like creating a decision matrix to evaluate options, but tailoring the criteria so that the option you know you want to win will come out on top at the end of the analysis process.  It’s another reminder that we are never as objective as we think we are, and that being a true “critical thinker” requires constant vigilance.