Types of judgement

Predictive judgement estimates what will happen. Evaluative judgement decides what matters and how much. Moral judgement decides what is owed to whom. Machines do the first. The claim that AI has made judgement scarce is really a claim about the other two.

Predictive judgement#

Who will default, which patient deteriorates, whether this deal closes, how long this will take. It has an answer that arrives eventually, which makes it the only one of the three that can be scored, and being scoreable is why it is the one machines have taken.

It is also the kind humans are weakest at. The clinical versus actuarial literature has sixty years of comparisons in which simple consistent rules hold their own against expert prediction, and the modal result is a tie rather than a rout. A profession that defined its expertise as prediction has been competing on the ground where its advantage was always thinnest.

Evaluative judgement#

Which outcome are we optimising. What counts as good enough. Which trade-off do we accept, between speed and thoroughness, coverage and depth, this quarter and the next three years. Whether this case is like the others.

This is the kind that gets left implicit, and the kind that machines appear to perform. A model ranking candidates is applying weights somebody chose, usually without recording that they chose them. A scoring system that flags a case as high risk has embedded an evaluation of what risk means. The evaluation happened; it happened earlier, elsewhere, and once, and is now applied a million times by people who cannot see it. That is the substance of the judgement premium, and the reason it is hard to price: the work is invisible even to the person doing it.

Moral judgement#

What is owed to whom, who carries the cost, and whether this should be done at all. It is distinguished from evaluation by consequence falling on somebody, so it cannot be delegated to a system that will not carry it.

A machine can produce the output of moral reasoning, fluently and at length. It holds no stake in the result, which means that what it produces is an argument rather than a judgement. This is the terrain the book calls override, described at the three terrains, the one where the cost of being wrong falls on a person rather than on a process.

Why separating them changes the argument#

The general claim, that AI has made human judgement scarce and valuable, is imprecise in a way that matters commercially. AI has made prediction abundant. Abundant prediction raises the value of the evaluation and the moral judgement that were always attached to it and were almost never separated out, named, documented or assessed.

That gives a sharper diagnosis than the general claim allows. An organisation that has automated prediction and left evaluation implicit has automated the visible half of its decisions and orphaned the other half. Nobody owns the weights. Nobody revisits them. The noise that shows up later is usually evaluative rather than predictive in origin.

It also sorts the training question. Prediction responds to calibration training, which is measurable because the forecasts resolve. Evaluation responds to review of the reasoning, because it has no outcome to score against. Moral judgement responds to forcing functions like the Principle Test, and to somebody being named. Three different problems, three different remedies, and a programme that treats them as one will train the easiest of the three and report success.

Where the boundaries blur#

The three are not clean categories. Deciding what counts as default is evaluative and shapes every prediction made afterwards. Deciding whose false negatives matter more is moral and arrives disguised as a threshold. The value of the distinction is in noticing when a decision presented as one type is doing the work of another, which is the usual way a moral question gets settled by a technical one.

What this does not establish#

The three-way division is an organising argument rather than a finding, and nothing here measures how often organisations conflate them. The separation of prediction from judgement is set out in the economics of AI literature and no first use is claimed for it; the distinction between factual and evaluative reasoning is much older still. What is added here is the third category, and the claim that evaluation rather than prediction is the part organisations have left undocumented, which is an argument from practice.

Related SuperSkills research#

• What is judgement
• Models of judgement
• Clinical versus actuarial judgement
• The three terrains