Consider any phenomenon, X, which is ‘unique’ in the sense that an intelligent agent (e.g. a person, an algorithm) is not familiar with the structure, performance, functions, or any other characteristics of the phenomenon in any meaningful way.
How would we analyse such a thing?
There are three methods, each with its own advantages and disadvantages:
The most correct approach to reasoning is by deduction: starting from general principles and narrowing down to a specific explanation of X. However, if X is sufficiently novel, there may not exist such general principles which effectively produce narrower conclusions regarding the characteristics of X. Even if such conclusions were to be made probabilistically or conditionally, there is no guarantee that the intelligent agent could correctly determine what probabilities or conditions would be relevant or accurate.
Second, the most instinctive approach to reasoning is by analogy: appealing to experience and inferring conclusions based on similar known phenomena. This however presents many, if not more, of the risks of the deductive approach above. If X is sufficiently novel, no analogy may be useful at all in producing insight regarding X’s nature. Furthermore, since analogy precludes the possibility of probability and conditionality in conclusions, the only way it may provide results is by ‘casting the net’ wider and wider in the search space of similar known phenomena to find something of use. However, as noted above, this process may yield incredibly dissimilar phenomena if X is sufficiently novel.
The third approach is by experiment, which will reveal the nature of X more directly but may also be far more costly than the first two approaches and may yield negative repercussions to the investigating agent.
Therefore, the most appropriate method of analysing unique phenomena should be (1) deductive where at all possible, with (2) analogy or (3) experiment following depending on the potential for catastrophic loss for the intelligent agent.