A scientific hypothesis must be falsifiable. Otherwise, it becomes a claim of faith, and ceases to be scientific. When Einstein proposed the General Relativity in 1915, it made certain testable predictions. A 1919 solar eclipse presented a scenario in which the predictions could be tested. Observations were made, and the hypothesis passed the test.
One should naturally be skeptical of hypotheses that are untestable, and thus unfalsifiable. However, unfalsifiability has certain flavors that are hard to dismiss.
An interesting case occurred in this discussion about AI where Nate Soares argued that a sufficiently intelligent AI, realizing that revealing its dangerous intentions would prompt its human supervisor to turn it off, would learn to mask its motives and wait for an opportune moment, and Andrew McAfee countered by saying this was an unfalsifiable hypothesis. Soares rebutted by saying it was falsifiable: if an intelligent system that could invent new technologies emerged and humanity was not destroyed, the hypothesis would have been falsified.
Here is the part left as implication: if such an AI was developed and humanity was destroyed by it, the hypothesis would have been proven to be true. See the problem?
I will call this the usefulness threshold problem of the falsifiability criterion. By the time the hypothesis above is proven true, it is too late. Hypotheses subject to this problem are almost as useless as unfalsifiable hypotheses.
The “Shifty General Problem” brought up by Soares shortly afterward points to this problem. If you suspect your general would betray you, you can’t wait for your general to betray you to prove your hypothesis to be correct.
But I have a much larger problem with the falsifiability criterion. When we move from a higher information state to a lower information state, it is difficult to hypothesize about the higher information state.
Illustration 1: Our solar system is located in a relatively sparse region of one of the arms of the Milky Way. This is why stars are visible in the night sky. Imagine a civilization on a planet located much closer to the center of the galaxy, where the presence of numerous stars makes the night sky too bright to view distant stars. If an astronomer in this civilization were to hypothesize the presence of these distant stars, it would be dismissed as unfalsifiable in the absence of an inter-galactic travel.
Illustration 2: If it is indeed the case that objects in the universe are diverging at an accelerating speed, less and less of the universe will be visible further out in the future. There will be a time when it will be physically impossible for light from another galaxy to reach our galaxy. Hypotheses about the existence of other galaxies would be unfalsifiable.
In both cases, the hypotheses would be true, but it would not be possible to prove them.
I wonder what important facts we are missing because of the information state we operate within. I try not to be immediately dismissive of an idea simply because of its unfalsifiability. The falsifiability criterion cannot be disregarded though: the floodgates to utter nonsense would open if we did.