alt Hacker NewsThe point is that the AI inferencing is equivalent to a person reading half a dozen separate papers, comprhending the basic concepts of each, relating them together into a mental model of the topic, and then writing an essay that summarizes the basic points. The person isn't plagiarizing anything here, but engaging in research, understanding, and synthesis of various sources of information.
The person absolutely does have the advantage of having empirical awareness and the ability to test their conclusions against external reality. But lots of people do engage in "research" and build mental models of various topics with little or no empirical context, and rely mainly on digesting calcified knowledge from other people.
I'm afraid, the essence is that is not. Re-sequencing content is not the same as synthesis and therefore not the same as a person processing information and communicating their own conception of this. There's a vital difference.
(We can even observe this in the resulting text: we immediately grasp the level of competence of the author, just by the way they take their path trough and at the matter. With LLMs, well, there's this even temperature, ready-made feeling, regulated by probability thresholds and RLHF sanctioned phrasing, also known as "slop" – even rhythmic intensifications, like "not this, not that, but…", which is actually a figure for a synthetic construct, don't help –, since the text isn't the trace or product of an actual organized thought – or, at least, an attempt at an organized thought.)
PS: "empirical a priori judgement" was meant as translation of synthetisches Urteil a priori (Kant). I.e., our ability to mentally prove concepts like congruency, which are not a priori, but can be inferred without regression to empirical knowledge. Typically, this requires both our inner sense (time, sequence, etc.) and outer senses (space, configuration, etc.)