As always, the tricky part is identifying who gets to validate the information, or control the truth. If the
inputs to AI have to be validated before AI can use it, then AI will be controlled by people who may well bring an agenda that goes beyond plain fact to those control decisions. One alternative is to validate the
outputs from AI, but that begs exactly the same questions. Putting the burden of that on the user violates the use case--a user using AI to obtain information not previously in hand, which requires that the results be accurate precisely because the user is unable to validate it from his own expertise. Another approach is to make the workings of the AI and the sources it cites sufficiently transparent to empower the user to perform that validation. That puts AI in the role of aggregator rather than educator. When I do a Google search, the AI result includes internet links after each statement, but probably 90% of the time, I can't find a single thing in the target that supports the statement from which it is linked and frequently I'm never able to independently validate the AI summary. So much for transparent workings and useful citations.
Journalism writ large developed strategies for curating fact that used to be (by the middle of the last century) baked into the ethics of journalism (but that now is waning with the entry of so many new routes from raw information to consumer). 75 years ago, most "reporters" were careful about using adjectives and adverbs, because they expressed opinion exogenous to plain fact. Adjectives and adverbs were the sole domain of the op-ed page where the publisher could express opinions and editorialize on the day's events. We lost that strict divide quite a while back, long before the Internet and certainly long before the general use of LLMs. But it didn't happen in a day--it took many years for that rigor to fade. It won't be restored easily, and so far the purveyors of AI show no interest in curation outside their own accepted truths (whether or not they are supported by data). AI is like journalism without those controls, but with the
appearance of being controlled. I foresee a lot of struggling to determine who gets to exercise that control in free markets, or in how AI never reaches any useful potential in closed markets.
Likewise, academic standards insisted that any received fact that is pre
sented in the work must be validated by other cited work, and any new fact must be validated by data analysis. Any statement that provides neither will attract the ire of a good grad-school committee or research paper reviewer. (I've reviewed hundreds of research papers published by one of the main journals in my line of work.)
Efficiency comes from hierarchical aggregation, but even that requires a point of view to determine what is worth keeping and how it should be summarized for more efficient access. And from faster processors and data communications, of course. I'm sitting right now in a building next to a road called UUNET Drive, literally feet from the old Metropolitan Area Exchange East, the first Internet Exchange Point outside of government control. If glass cable was radioactive, we'd all be dead
I am surrounded by 100-foot tall multi-acre data centers, one right after the other, and more being built on every square inch of available land. The current strategy for avoiding packet congestion for AI crawlers is more capacity, not more efficiency. But in any case, AI being faster doesn't necessarily address the ethical issues being raised here.
I know that lots of people are paid to think about these things, and a lot of thinking is being done, some of it out in the open and some not. That doesn't mean those thinkers will arrive at a solution, or at a solution that doesn't impose some external implacable authority that works against the fantasy. We shall see.
Rick "doesn't believe in either Superman or Frankenstein, but suspecting Frankenstein is more likely" Denney
