I've started reading Michael Polanyi's Personal Knowledge. In the second chapter Polanyi states the following "fact" regarding scientific observations:

"The derivation of data and checking of data that bridge the gap between our instrument readings and the magnitudes figuring in our formulae can never be fully automatic. For any correlation between a measured number introduced into an exact theory and the corresponding instrument readings, rests on an estimate of observational errors which cannot be definitively prescribed by rule."

Explaining with:

"This indeterminacy is due in the first place to the statistical fluctuations of observational errors... In consequence of such random errors we can only proceed from the probable values of initial data to probable values of predicted magnitudes, and since no strict relationship exists between these two sets of figures, the process remains to this extent indeterminate. Apart from these fluctuations we have always the possibility of systematic errors. Even the most strictly mechanized procedure leaves something to personal skill in the exercise of which an individual bias may enter."

(Personal Knowledge, 1962, p19; my own emphasis)

In the chapter he gives for example Astronomy, but it seems quite obvious Polanyi regards any scientific observation as something that cannot be automated to its full extent and will always need a human brain at works.

I wonder, is this statement true? More precisely, was it true in the time of publication (1958), and if so has there been any development since then?

[I thought of asking this in a scientific SE such as Phys.SE, but this seems on the border of phi-sci and because I know the community here better I figured I'd get more unbiased answers here.]

  • 4
    No. Quine's Two Dogmas, where he already argued that observations are theory laden and there is no clean "language of observation", came out back in 1951, Polanyi just applies it more specifically. Of course, in established theories, where meter readings have non-controversial theoretical interpretations, recording them can be and is widely automated. Even then, people may have to decide what is or is not a glitch, an outlier, etc.
    – Conifold
    Apr 16, 2020 at 8:00
  • @Conifold nice! Two Dogmas is right there on the reading list :) Apr 16, 2020 at 8:27
  • 1
    You may also find Einstein's thoughts on the subject interesting, from a 1920s conversation reported by Heisenberg:"Whether you can observe a thing or not depends on the theory which you use. It is the theory which decides what can be observed... Observation means that we construct some connection between a phenomenon and our realization of the phenomenon.", see Do theories come from observations or do they determine what is observed? And unlike Quine, Einstein was a firm realist.
    – Conifold
    Apr 16, 2020 at 22:38
  • @Conifold and indeed Polanyi attempts to prove this point exactly using Einstein, quoting a question he asked Einstein regarding the effect Michelson's experiment had on relativity theory (it had no effect, Einstein claims to have thought the theory when he was quite young). I also recall reading somewhere that Newton once said that theory is always what guided observation, never the other way around. Can't remember the exact quote, but it followed the same theme. Apr 17, 2020 at 8:25
  • Maybe information theory might provide another line of though to go for. This passage strikes me as circling around the difference between data and information
    – Philip Klöcking
    Apr 18, 2020 at 14:31

1 Answer 1


An often under-appreciated point: observation shows us events, but it does not show us processes or forces. An observation describes the state of an object at a given moment; a series of observations shows us stepwise changes in the momentary state of an object. But an observation is inherently a snapshot of reality in which motion and direction and connection may be implied, but are not shown. I've always felt this was the root of the uncertainty principle in quantum mechanics. If we (say) take a picture of a quarterback throwing a pass, we can sharpen the exposure to make the quarterback and football clear and precise, but then we can't tell whether the QB is actually throwing the ball or just posing statically. If we loosen the exposure we see the QBs body and arm start to blur, indicating motion, but only at the cost of precision and clarity. This can be a little confusing because the 'moment' of an observational snapshot is actually a short elapse of time compressed together as though it were a single instant: i.e., two observations might partially overlap if the time between them is small compared to the elapse that is compressed into the 'moment', leading to odd data issues. But all we get from an observation are these singular, momentary points.

Of course, what we want to talk about is the processes, systems, structures, forces, or what-you-will that these momentary snapshots are snapshots of. If we pull out that QB photo, we don't usually want to talk about the details of the photo itself, like the color of the jersey or the placement of the QB's feet. We want to be able to say something like "That picture shows Trent throwing the winning touchdown in the 2014 inter-varsity championships", something that embeds that moment in a continuous ongoing process that has relevance and meaning beyond what the photo can show. And so we build theories, which are narratives about what's happening under the hood such that we see the momentary observations that we see. This is fairly obvious with the Theory of Evolution, where what we have is an assortment of fossils and biological samples that show correspondences (our momentary observations from across millennia, in rough temporal order), and where we assert that these correspondences reflect an ongoing, continuous process of evolution from one form to the next. If it's less obvious with physics, that's only because we have a harder time seeing the 'narrative' aspect of mathematics (though anyone who thinks about how Newton derived the calculus will probably see it).

We can build machines that make observations. That is the purpose of most scientific instruments: to expand the limited human perceptual field to observe things we could not otherwise see. We can even use machines to automate those observations, so that we can go off and have coffee or see a movie while it chugs away. But (short of some quasi-mythological AI) a machine cannot reflect on the underlying processes that produce those observations. Say we had a movie of a quarterback throwing a pass. A machine could certainly record observations from each frame; it could probably (with modern software) pick out individual players, and possibly track the ball from frame to frame as a distinct object. But could a machine take those observations and theorize that there was some kind of contest involved that spans these individual frames: a contest with underlying rules and goals and structures, that causes the frames to happen in the way they do? Those things cannot be observed from the frames themselves, and there is no way to program a non-intelligent machine to 'automatically' draw them out of the combined observations.

  • I'm not sure about the observation as events part, but nonetheless I'll address the main point: couldn't a machine "deduce" that this type of ball-throwing motion picture is a football game by comparing and analyzing millions of types of ball-related activities? Furthermore, if it had the date, time and location it could even suggest that this was the same contest we ourselves deduce it was. And this isn't some imaginary AI, this is a very doable AI project. This doesn't seem to me the place where we won't be able to automate observational deduction. Apr 18, 2020 at 17:39
  • @YechiamWeiss: An artificial intelligence arguably could do that, but how does a non-intelligent machine 'deduce' anything? I have one set of images showing a QB throwing a pass; another set of images showing a pitcher throwing a pitch; a third set showing an olympic javelin thrower throwing a javelin; a fourth, a marine throwing a hand grenade. How would a non-intelligent machine 'deduce' these were different 'games' with different rules? And let's not invoke 'AI' as magical reasoning to explain away the conceptual difficulty here. Apr 18, 2020 at 19:28
  • When I'm talking about "AI", I mean machine learning, specifically deep learning techniques that are widely used and explored today. Currently image analyzing is one of its strengths. And the way it works is by receiving a huge data set (millions of pictures) that are related to different types of ball-games, with a classification of each image (data1=[image of pitcher throwing a pitcher, game=baseball]). After analyzing said data set it can (using some black-box "reasoning" we're unaware of) suggest with great accuracy the next image you'll give it. It really isn't magic. Apr 19, 2020 at 6:25
  • That's a great attempt. +1 Obviously the question goes to the heart of what human-level intelligence really is, a topic, by no means answered well by anyone. Why are you sure that observations are showing us the state of the object and not the state of the representation of the object?
    – J D
    Sep 15, 2020 at 19:01
  • I'm currently getting comfortable accepting Searle's rejection. online.sfsu.edu/kbach/Searle.html
    – J D
    Sep 15, 2020 at 19:06

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .