At first, I thought this question would lead me directly into a discussion about evidence, prediction, or scientific methods. But before I even got there, I stumbled over a word I had used in the Indonesian version. And that word was “benar.”

I realised that from the beginning, I had been using the word “benar” to talk about the value of information or a model.

The English version did not make me nearly as uneasy. There, I used true, which is grammatically an adjective related to truth. As I understand it, truth usually refers to the concept or state of being true, while true is used to assign a truth-value to a statement or proposition. The relationship between the two, and what actually makes a proposition true, is of course a much larger philosophical question than what I intend to discuss here.

In Indonesian, the word “benar” feels much more ambiguous. It can be used to talk about the truth of a statement, but it can also be used in contexts involving moral judgement, behaviour, or something considered appropriate or normatively right.

After looking into it, I found that there are many other words with slightly different nuances: tepat, betul, sahih, valid, akurat, absah, and others.

Unfortunately, when I was writing yesterday, “benar” was simply the most intuitive shortcut that came to mind, and I was too lazy to revise it. 😑

Rather than merely revising it, I decided to continue yesterday’s note from its final question.

Model A in yesterday’s thought experiment states:

“The tendency for X increases as variable Y increases.”

Model A might be very simple. But if the relationship between X and Y can actually be observed and empirically tested, its simplicity does not by itself make its claim less true than that of a more complex model.

And by less true, I do not mean that truth has a scale like 50% or 80%. I only mean that the simplicity of Model A is not, by itself, a reason to consider its claim lower in truth-value than the claim made by a more complex model.

In other words, complexity and truth are not the same scale.

I am also not trying to say that truth is always something that can be observed or empirically tested. That would be a much larger claim than what I am trying to discuss here. Within the context of my thought experiment, I am simply assuming that Model A makes a factual claim about the relationship between X and Y, and therefore that relationship can be questioned empirically.

Model B, on the other hand, contains dozens of variables and interactions that sound genius, nerdy, elegante. 😭

Model B explains:

“The tendency for X emerges from 24 psychological factors caused by 50 social factors, consisting of 30 economic and political factors, 20 cultural factors influenced by several ecological parameters such as climate affecting temperature, biodiversity affecting noise from insects and birds in the environment, air pollution, noise pollution, and visual pollution from surrounding human activity.”

But if most of those variables lack strong empirical grounding, cannot be tested, or do not make the model fit observations any better, that complexity gives us no additional reason to trust it.

So, in the context of empirical claims like the one used in this thought experiment, one reason to trust a model more is when it can explain or predict something that actually occurs, when its assumptions can be tested, and when its results remain consistent with the available evidence.

But this does not mean that empirical testability is a universal requirement for everything that can be considered true. My thought experiment has a limit because I chose an example that involves a claim about an empirical phenomenon from the beginning.

There are many other kinds of questions that do not work in exactly the same way.

For example, when we talk about numbers, we do not usually conduct experiments on nature to find out whether the number 2 actually “exists” somewhere, because it cannot be treated in exactly the same way as the claim that “X increases when Y increases”. The former leads us into questions about concepts, definitions, logic, or mathematical and metaphysical structures, rather than empirical observation alone.

This made me realise that perhaps I was too quick to apply one method of evaluation to every kind of claim.

The way we obtain reasons to trust something may depend on the kind of claim being made.

And if that is the case, the question “what makes a model worth trusting?” turns out to be much larger than I initially thought.

Then I started wondering: what does simple actually mean? What is complexity? Is a more complex explanation more true?

I also don’t know whether there is any measure of complexity that the entire human population would agree on. But intuitively, I feel that an answer or explanation becomes more complex when it involves many variables and the interactions between them at once, to the point where we need to spend more cognitive resources processing it.

If simplicity is the opposite of complexity, then a simple model might be a model that discards variables considered irrelevant to answering a particular question. Such a model might even be intuitively understandable to the person asking the question.

But I still don’t quite get it. So I tried a thought experiment.

Two Models

Imagine there are two models. Model A and Model B are trying to answer the same question.

Model A explains:

“The tendency for X increases as variable Y increases.”

Meanwhile, Model B explains:

“The tendency for X emerges from 24 psychological factors caused by 50 social factors, consisting of 30 economic and political factors, 20 cultural factors influenced by several ecological parameters such as climate affecting temperature, biodiversity affecting noise from insects and birds in the environment, air pollution, noise pollution, and visual pollution from surrounding human activity.”

Model B sounds much cooler. 😭

To me, B is clearly the more complex explanation. But its complexity doesn’t tell us whether the model is true or not.

Model B has something like an aesthetic of explanation. It sounds scientific because it contains many variables, many causal relationships, and many details. From there, I can easily make the following leap:

But this thought experiment creates another problem for my assumption: that a complex explanation might be harder to understand intuitively.

A model can be structurally complex yet written so elegantly that it is easy to understand. Conversely, a model that is actually simple can be explained in isekai language until the reader feels like they are reading a spellbook for summoning a demon king.

So, “hard to understand” is not a reliable indicator of a model’s complexity.

And yes, at this point I am judging my own explanation of Model B to be elegant. 😌

Are All Those Variables Necessary?

Model B looks more complex.

But now the question becomes:

Are all 24 psychological factors actually necessary?

What if 17 of them have no replicable effect?

What if some of them don’t have a consistent definition?

What if 5 of the factors are actually consequences of Y rather than independent causes?

What if some of the variables merely make the model look more complete without improving its ability to explain or predict anything?

Conversely, what if the very simple Model A actually predicts the data more accurately?

And what if Model A is indeed too simple, causing it to fail when conditions change, while Model B can predict phenomena that Model A cannot explain?

At this point, I started to realise that the question of complexity cannot really be separated from the question of what a model is for.

A model does not have to include everything that might influence a phenomenon, because it is built to answer a particular question. Therefore, variables that are not necessary for that question might actually be better left out.

But leaving variables out can also become a problem if those variables turn out to be important for explaining the phenomenon or cause the model’s predictions to fail when the context changes.

So what gives us a reason to trust a model