# Observations are not Sources of Knowledge It is often thought that through observation we can uncover knowledge about the world.This is false. Observation does not provide us with knowledge. It is not generative. In reality, knowledge is generated via *creative conjectures*. We start with a [Problem](Problem.md). We conjecture some solution. And then we compare that solution against reality. It is the problem solver, *us*, that is *generative*. In this light it is clear that observation is *discriminative*. It is used to *criticize* theories. But what about a data scientist simply going to explore some data, curious to see what they find? Surely in that case observation is generative? Where else could any knowledge come from if not the observations the scientist made? Yet even here, the role of observation is discriminative. Let me explain. The scientist sets off to look at this data with no preconceived biases or expectations of what the data may look like. As far as they are concerned their mind is a blank canvas, open to all possibilities. They have no specific problem they are trying to solve, they are just [curious](Problem.md###Curiosity) to see what they find. Every single sentence in the above paragraph is *false*. Let us start with the idea that the scientist has no problem they are trying to solve. They most definitely do have a problem (likely multiple problems) they are implicitly interested in solving. They are trying to satisfy their [curiosity](Problem.md###Curiosity). They believe their is something they may learn or find in this data, and investigating it will solve the problem of not knowing what it is. So, they do in fact have a problem. But what about their mind being a blank canvas, open to all possibilities of what the data show? Surely anything learned from this data must have come *from the data* - it must have been *generative*. Again, this is false. We can see this via extending our example. Imagine we now not only have the scientist, but a literary novelist as well who is looking at this data. The scientist is the "driver" of the computer, writing code to conduct the analysis, generating plots and print outs of measurements, while the novelist looks over their shoulder. Each has the exact same stream of electromagnetic waves entering their visual system from the computer. So the input they receive is identical. However, the *information* they receive is not[^1]. Each may see the exact same scatter plot, but because they have different background knowledge, the context within which they see this is different. They each come loaded with all sorts of implicit theories about the way the world works, the way data works, knowledge of mathematics, relationships between entities, and so on. It is against this background that any observations are viewed. But even now, the observations do not *add* to the background! They don't add additional layers of knowledge. Imagine that the scientist sees that some slice of data has interesting, unexpected properties. On a quick first pass it may seem that this slice conferred some new knowledge - but it is really the other way round! The scientist has some implicit theory of how the data would look (perhaps their implicit theory is that the slice would have been uninteresting), but the slice contradicted their theory. In Popperian terminology it [Falsified](falsification.md) the theory. That is *all* it did! Any subsequent ideas about this slice, why it may be interesting, how it relates to other slices or other parts of the problem - those are all *creative conjectures* that come from the *scientist* - not the slice! Notice that this particular slice of data likely would have had *no impact* whatsoever on the novelist. The same observation, given different background knowledge (implicit theories), can yield very different conclusions. What we can see is that [All Observations are Theory Laden](All%20Observations%20are%20Theory%20Laden.md), they are *contextual*. They do not exist "in a vacuum". Now, you may ask what the practical benefit of this distinction is - won't we arrive at the same result regardless? No! The benefit of this distinction is that it casts direct light on the fact that *your preexisting theories matter*. You cannot simply "observe your way to success". All observations will be in the context of your own theories (both explicit and implicit). The explicit theories tend to be rather harmless because they are so exposed. However, you should try to bring your implicit theories to the surface, making them explicit. The better these theories and the more clearly you can reason about the them, the more useful observation will be. --- Date: 20240806 Links to: [Fabric of Reality](Fabric%20of%20Reality.md) [Beginning of Infinity](Beginning%20of%20Infinity.md) pg 17 Tags: References: * [#7 On the derangements of science - by Arjun Khemani](https://news.criticalrationalism.org/p/on-the-derangements-of-science) [^1]: Information is subjective, see [What is Information](What%20is%20Information.pdf)