I’m talking about Charles Sanders Peirce’s article Deduction, Induction and Hypothesis. He describes a kind of thinking he calls “abduction”, which is the drawing of plausible conclusions that could not be arrived at based solely on what is observed. It’s an operation that makes human thinking different to AI’s
To try to explain it briefly before moving onto more interesting points—think of all the information you receive through sensory organs. This is all “external world information”. That’s the information AI has at its disposal. As the theory goes, humans in contrast are able to use something more than external world information. There are certain conclusions we draw that cannot be derived purely from what we take in through sense perception. That extra information exists in the human mind innately.
I’d like to be able to present an easy-to-grasp demonstration proving humans perform abduction and machines do not. I won’t succeed, but it will be interesting to consider why it is difficult to find clear examples.
To illustrate, I take Peirce’s example: “Fossils are found; say, remains like those of fishes, but far in the interior of the country. To explain the phenomenon, we suppose the sea once washed over this land.”
The process of arriving at the conclusion via the methods of AI might be described as follows. In the computer’s dataset, far more often than not, fish live only in water, not on land. In the dataset, floods sometimes occur. Given the statistical frequencies, the conclusion is “the sea once washed over this land”.
The human proposes the same explanation, and some people claim the human and computer arrive at their conclusions via the same methods.
But—that the human and the machine produce similar conclusions in many instances does not mean they perform the same operation to arrive at that result.
(AI is, after all, copying us, so perhaps it’s to be expected that its conclusions on occasion match ours.)
I don’t find the examples Peirce gives in his article very conclusive, and I’ve quoted the one I thought best (though I find his argument interesting).
When trying to conceive examples of abduction, all I can think of are grand ones e.g. Newton’s positing gravity. It’s by no means certain that that new idea, “gravity”, could be arrived at by reproducing high-frequency combinations of items (e.g. words) in past data. But such examples are too exceptional and too removed from daily life.
An example closer to mundane experiences would be easier to grasp. But everyday occurrences are easy to grasp perhaps because they are things of which there are many past examples, many past pairings of “phenomena” and “result”. And therefore they are precisely the sorts of things AI can somewhat reliably handle. (It depends on repeated examples.)
To formulate a rule expressing the methods of human ingenuity would, like all general rules, require many similar, repeated occurrences, from which the common elements can be abstracted and thus the rule discerned.
But the clearest proof of human ingenuity is perhaps only in the rare instances—precisely the instances of which there are few or no repeated examples. And those instances are so different in substance that it is hard to find similarities across them. What features can be found common to Newton’s innovations and Dante’s?
I certainly can discern none that could fit into a clear formulation, like in a physics equation or a rule of logic or computing.
We need repeat occurrences to form the basis of a reliable theory, but since human ingenuity may lie in things that are never repeated, we may be unable to form a theory that proves or describes human ingenuity.
And if our ingenuity may be clearest in grand examples, that does not mean that methods of thinking in those instances are not the same methods we use in many mundane cases.
We might speculate that those more mundane cases are what makes a human interesting and human-seeming, while interactions with computers in contrast feel bland.
Perhaps right now as a society we are conducting the experiment that will produce the evidence needed to more easily differentiate human minds from machines’. As we deploy AI more and more in our world, we see ever more what conclusions it is capable of drawing.
I’m talking about Charles Sanders Peirce’s article Deduction, Induction and Hypothesis. He describes a kind of thinking he calls “abduction”, which is the drawing of plausible conclusions that could not be arrived at based solely on what is observed. It’s an operation that makes human thinking different to AI’s
To try to explain it briefly before moving onto more interesting points—think of all the information you receive through sensory organs. This is all “external world information”. That’s the information AI has at its disposal. As the theory goes, humans in contrast are able to use something more than external world information. There are certain conclusions we draw that cannot be derived purely from what we take in through sense perception. That extra information exists in the human mind innately.
I’d like to be able to present an easy-to-grasp demonstration proving humans perform abduction and machines do not. I won’t succeed, but it will be interesting to consider why it is difficult to find clear examples.
To illustrate, I take Peirce’s example: “Fossils are found; say, remains like those of fishes, but far in the interior of the country. To explain the phenomenon, we suppose the sea once washed over this land.”
The process of arriving at the conclusion via the methods of AI might be described as follows. In the computer’s dataset, far more often than not, fish live only in water, not on land. In the dataset, floods sometimes occur. Given the statistical frequencies, the conclusion is “the sea once washed over this land”.
The human proposes the same explanation, and some people claim the human and computer arrive at their conclusions via the same methods.
But—that the human and the machine produce similar conclusions in many instances does not mean they perform the same operation to arrive at that result.
(AI is, after all, copying us, so perhaps it’s to be expected that its conclusions on occasion match ours.)
I don’t find the examples Peirce gives in his article very conclusive, and I’ve quoted the one I thought best (though I find his argument interesting).
When trying to conceive examples of abduction, all I can think of are grand ones e.g. Newton’s positing gravity. It’s by no means certain that that new idea, “gravity”, could be arrived at by reproducing high-frequency combinations of items (e.g. words) in past data. But such examples are too exceptional and too removed from daily life.
An example closer to mundane experiences would be easier to grasp. But everyday occurrences are easy to grasp perhaps because they are things of which there are many past examples, many past pairings of “phenomena” and “result”. And therefore they are precisely the sorts of things AI can somewhat reliably handle. (It depends on repeated examples.)
To formulate a rule expressing the methods of human ingenuity would, like all general rules, require many similar, repeated occurrences, from which the common elements can be abstracted and thus the rule discerned.
But the clearest proof of human ingenuity is perhaps only in the rare instances—precisely the instances of which there are few or no repeated examples. And those instances are so different in substance that it is hard to find similarities across them. What features can be found common to Newton’s innovations and Dante’s?
I certainly can discern none that could fit into a clear formulation, like in a physics equation or a rule of logic or computing.
We need repeat occurrences to form the basis of a reliable theory, but since human ingenuity may lie in things that are never repeated, we may be unable to form a theory that proves or describes human ingenuity.
And if our ingenuity may be clearest in grand examples, that does not mean that methods of thinking in those instances are not the same methods we use in many mundane cases.
We might speculate that those more mundane cases are what makes a human interesting and human-seeming, while interactions with computers in contrast feel bland.
Perhaps right now as a society we are conducting the experiment that will produce the evidence needed to more easily differentiate human minds from machines’. As we deploy AI more and more in our world, we see ever more what conclusions it is capable of drawing.
.
.
.
Take a look at my other articles:
On why AI cannot innovate
On Ilya Sutskever’s failure to answer the Q: How will AGI be invented?
On Eliezer Yudkowsky’s unsubstantiated arguments
On what words mean to computers
.
Twitter/X: x.com/OscarMDavies