One way to make this clear to yourself is to observe how much more difficult it is to "define bread" than it is to answer "is this bread?"
~Scott Pilgrim
There is a lot of context that is needed to get to a positive identification (maybe the word you meant) of bread, but that is true of many definitions present in dictionaries, etc. today.
Since there are some inputs x where O(x0) = False, some where O(x1) = True, and the laws of physics are continuous(yes, even in quantum mechanics), Buridan's Principle implies that you are incapable of deciding the breadness of arbitrary input in bounded time.
But the set of inputs that an oracle implicitly defines, could change if the oracle changes. And you could change your mind or die tomorrow.
So you would need a very large number of definitions of bread, indexed by (time, person). Any one of them could be a valid definition - it's theoretically possible to make you look at 1000 pictures of bread so your brain is encouraged to make a bread-detector neuron, and then scan your brain and calculate its response on any input - but you don't know which one is correct to use for any purpose.
i.e. If I want to start a bakery, should I use your current bread-oracle to define "marketable bread", your bread-oracle as of 5 years ago, should I take a statistical ensemble of brain scans from millions of people, or should I use my own?
It seems like just having a function that returns true on some inputs and false on others doesn't tell you much, whereas traditional mathematical definitions have strict relations to other things.
but they do, the definition of many words changed over time some to even start to mean the opposite of what they initially did.
[1] I don't think this is a very useful definition of bread.
It's a practical issue that affects CPU design, which is also mentioned in the paper under the name "Arbiter Problem".
Notably, the public reacted similarly then as HN does now, rejecting the notion that meaning is only constructed and, furthermore, hopelessly solipsistic.