Position of collapse in von-Neumann chain I have often seen the following claim (for instance on Wikipedia: https://en.wikipedia.org/wiki/Von_Neumann%E2%80%93Wigner_interpretation):

In his 1932 book The Mathematical Foundations of Quantum Mechanics, John von Neumann argued that the mathematics of quantum mechanics allows for the collapse of the wave function to be placed at any position in the causal chain from the measurement device to the "subjective perception" of the human observer. 

But I simply don't understand how this can be true. To me this reads as if it would have no experimental consequence where the point of collapse (between measurement device and observer). But how can this be? For instance, for a given quantum experiment we could imagine two hypothesis:


*

*It is really the measurement device that causes the collapse (due to some proprety of the way it measures), completely independent on any human observer. The human observer just plays the role of observing whatever resulting value was recorded due to the collapse.

*The human observer plays a crucial role in causing the collapse (through consciousness or in some other way) so before the human gets involved, the measurement device has now entered into a superposition itself based on the possible states of what it measured. Only when the human observes the measurement device does the collapse take place.


To me it seems it must be possible, for any given experiment, to change the experiment to find out if one or the other case is true.
For instance, if case 2 is true, the whole experimental setup (including measurement device) retains the superposition of states until the human observes it. This should make it possible to elucidate the well-known quantum effects ("quantum weirdness") even after the measurement by the measuring device. For instance if the experiment being measured involved a binary state of potentially negative amplitudes then the measurement device would now also be in a binary state with potentially negative amplitudes. If the measurement device is made to emit, say, a photon with different polarization based on the binary value the device observed, then the device will (still assuming case 2) be emitting photons also in superposition with potentially negative amplitudes. These photons can then be used in a second experimental setup that would show the effects of destructive interference among particles (exploting the negative amplitudes). The human then observes the effect of the second experiment: If the destructive interference is seen, this confirms case 2 because it can be concluded that the measurement device did not collapse the wave-function. On the other hand, if the interference effects are not observed, it means that the measurement device itself actually did collapse the wave function even before the human looked at the measurement, thus causing the photons emitted of being in a state with only positive amplitudes (0,1) or (1,0) meaning that destructive interference was not impossible, human observer or not.
Since there are obvious experimental differences it follows that this cannot just be an interpretational isuse. What am I missing?
 A: If you actually tried this experiment you would not observe any interference effects because the measurement device is a macroscopic system which undergoes decoherence. So according to the criterion in your question you would say the device collapses the wavefunction. Decoherence is generally how physicists think about "collapse" now.
This doesn't resolve all the differences in interpretation because we could say even after decoherence the measurement device together with the degrees of freedom in the environment are still in a coherent superposition. Also even though we can't observe interference effects due to the measurement device itself, it is still in a superposition in the sense that we still don't know which state we will actually observe. There is more discussion in the paper in the link above.
A: The following is taken from Kurt Jacobs 2014. Quantum Measurement Theory:
How do we decide when it is appropriate to pretend that a measurement has occurred, and select one of the orthogonal subspaces as the outcome? We explain this in two stages. First the following process acts like a measurement, although it cannot guarantee that the measurement will not be undone. Consider a system in the state $\lvert \psi \rangle = \Sigma_n c_n \lvert n \rangle $, and allow the system to interact with a second system so that the joint state become $$w=\Sigma_n c_n \lvert n \rangle \otimes \lvert n \rangle_{sys2} $$
If we now trace out the second system, the state of the first system become the mixture $$\rho = \Sigma_n \lvert c_n \rvert ^2\lvert n \rangle \langle n \rvert $$
The interaction between the two systems, followed by the loss of access to the second system, turns the initial superposition into a mixture. Since a mixture of states is equivalent to the situation in which the system is really in one of the possible states $\lvert n \rangle$, but we do not know which one, the interaction produces an effective collapse of the wavefunction, but without selecting one of the possible outcomes. 
To be able to safely assume that a measurement has been performed, we need to be sure that it is not going to be undone. In the above scenario a second interaction could undo the correlation, and return the system to its initial superposition. It turns out that the effectively irreversible behaviour of large quantum systems provides a physical process, in fact the only known process, that allows us to safely apply a measurement and pretend that a collapse has taken place.  
When a microscopic quantum system is correlated with the macroscopically distinguishable states of a macroscopic system, then because of the unavoidable interaction of the macroscopic system with the environment, we can assume that a measurement of these macroscopic states, and thus of the microscopic system, has taken place.
A: No, you would not see the interference either way. Even if the collapse only happens when the observer observes the measurement device, the measurement device would not be in a superposition state prior to the observation by the observer. In fact, it would not be in any state at all! It would be part of an entangled state of the object + the measurement device. For example, in a typical case of measuring a spin$-1/2$ particle, the state of the combined system of measurement device + the particle would be
\begin{align}
\vert\mathrm{sees }\uparrow\rangle\otimes\vert\uparrow\rangle+\vert\mathrm{sees }\downarrow\rangle\vert\downarrow\rangle
\end{align}
if the particle being measured was in a superposition of $\vert \uparrow\rangle$ and $\vert\downarrow\rangle$. Here, the measurement device does not have a quantum state of its own, i.e., its density matrix doesn't have the property $\rho^2=\rho$. Rather, its density matrix is a maximal entropy density matrix.
A: Consider a two-slit experiment, with a detector in one slit. The standard thing to do is to assume that the detector (a device) collapses the wave-function, which is why you don't see the interference pattern on the screen. This is your option-1.
Now, consider the same experiment but do NOT collapse the wave-function at the device, but rather at the person looking at the screen. This is your option-2. The wave-function of the device and experiment would now be a superposition state. When the person looks at the screen, he will collapse the state with a probability corresponding to this superposition state, which will produce probabilities just like the single-slit state. Thus, he will see no interference pattern.
It is always possible to consider that reality is not fixed, the state is still in a superposition, and the measurement only occurs afterwards. The "cost" to this is that the superpositon will be quite convoluted. Consider a run of the two-slit experiments, with millions of electrons hitting the screen - but the person only looking at it when the experiment is done. The state of the system before the person looks at the screen is a grotesque superposition of infinitely many non-interference patterns, each with a corresponding data-record in the detector. Even including, say, a superposition of a paper-printout of the detector's recording.
