In a discussion with Werner Heisenberg, Albert Einstein once remarked that only theory determines what is observable. According to this view, there is no theory-free observation. Some thinkers in the Vienna Circle, i.e. the movement of logical empiricism, had tried to build the foundation of science on a language of theory-free “protocol sentences”. This idea was criticized by others who advanced the view that any such statement describing an observation already depends on a theory.
If the idea that observations are always theory-dependent is true, scientists are in a similar situation as scholars trying to understand some piece of culture, for example a text: their understanding develops in what is called a hermeneutic circle. You start with some preconception (a first version of a theory). You make observations based on that preconception (like a scholar trying to understand single statements or sections of a text). Based on those observations, you can then revise the theory and then reinterpret the observations in the light of the revised theory. The language “in terms of which” you describe observations is revised as the theory is revised. The understanding of the entities under investigation (e.g. a part of nature or a product of a culture) and the theories co-develop.
As the knowledge about phenomena is evolving, and the concepts and thus the language used is evolving with it, the phenomena as described in the observational statements may change. Things that appeared to be instances of the same may now be seen as instances of different concepts (division or bifurcation). In some cases, it might be possible to view different phenomena as instances of one underlying structure (unification). The development of the knowledge can be described as proceeding by the division and unification of analytical spaces (a conception I have taken from the work of Kurt Ammon). Analytical spaces are defined as consisting of some knowledge together with the objects the knowledge refers to.
If science proceeds in the mode of a hermeneutic circle, the difference between science and scholarship appears to be gradual. Science consists of some analytical spaces that are very large and relatively stable. In the sciences, large amounts of objects or observational data can be assimilated into a relatively small number of analytic spaces. Scholars, on the other hand, are facing phenomena that are much more complex. Their knowledge consists of relatively small, special and less stable analytical spaces.
If observations at any time depend on theories, i.e. on the state of knowledge and the associated language, then the world of phenomena cannot be taken as a stable given. What you experience depends on what you know. Some of that knowledge might be hard-wired into the structure of your perceptive apparatus, but some may change. But even if parts of our perceptive system have a fixed, “hard-wired” structure, science does not depend on this structure. If part of the structure has some plasticity and can be changed through learning processes, i.e. if part of it is, in a way, artificial, then we are not bound to leave the border of the observer always in the same place. We can redefine that border. For example, we can put devices like microscopes or measuring devices in front of our eyes. We can add artificial systems to process the information generated by those devices.
By redefining the border and the structure of the observer, we can change what is observable. For example, a human being might never be able to observe a top quark directly, but a system that consists of human beings, a particle accelerator, a network of computers etc. is able to make such an observation. The structure of this system can be viewed as the embodiment of a theory. That theory might be wrong. For example, what appeared as an observation of neutrinos moving faster than light in one experiment then turned out to be caused by a technical problem of the device.
Another example: we might seem not to be able to observe the experiences of other people. However, we can gain knowledge about them by using a technique like interviews. We simply ask the people. We are then using a theory about how the experiences of the interviewed people are mapped into their utterances. That theory might be wrong or partially faulty (e.g. we might assume that people are not lying when actually they sometimes do) but we can still say that we are observing the experiences of other people (based on a theory). In a similar way, a historian might be able to infer knowledge about past events by applying some assumptions to some sources. He might change the assumptions, e.g. by performing a critical assessment of the sources. The historian’s results can be viewed as (theory-dependent) observations.
If all observations are theory-dependent, there is no reason to limit ourselves to a certain “horizon” of phenomena. What is beyond that horizon can be made observable by using a theory. For example, we might assume that an “external world” actually exists, that it has certain properties etc. and then interpret our observations based on such a theory. We can never eliminate all metaphysics because we cannot eliminate every bit of theory, but there is no fixed horizon between phenomena and noumena (to use Kant’s terminology) because the cognitive system and the attached devices and systems of information processing (including our own brains and the cognitive structures “running” on them) have no fixed structure. The “horizon” can be moved by changing the theory and the structure of the observing system. The observing system may consist of people (and their knowledge that, in turn, is changeable), measurement devices and experimental set-ups, including things like telescopes or particle accelerators, computers, libraries, scientific journals, interview recorders, black boards etc.
There is no fixed a-priori structure of this observing system. There is always the possibility of it containing faults, and there is no theory-independent birds-eye view from which we can decide what is correct or not correct. It is not possible to establish truth once and for all. We should, therefore, regard knowledge always as preliminary, containing anomalies like errors, gaps, inconsistencies, vagueness, misconceptions, inefficiencies etc. A critical assessment of assumptions, theories and hypotheses is therefore always necessary and there is not complete and explicit formal theory or algorithm to do so. This also means that philosophy is not going to go away.
(The picture is from http://commons.wikimedia.org/wiki/File:CMS_Higgs-event.jpg. It is showing a snapshot from a simulation of the decay of a Higgs boson in the CMS particle detector at the CERN research facility. The simulation, running on a big computer, is an embodiment of a theory. The results can be compared with observations from experiments, so the simulation can be used to produce predictions. A match can be viewed as evidence for the presence of the Higgs particle, i.e. as an observation of such a particle. The (sub-)programs that are part of the simulation and the programming language used to implement them can be viewed as an extension of language used to express the theory.)