"I'm not sure what there is to verify by an AI."

It seems to me an AI would be ideal as a mechanism to compare the data being ArcHived to the data published at the URL. There are ways to conceal things in blobs, steganography, for example, that people looking at a site with Mark I eyeballs might take a bit to notice, or even miss, but AI would detect instantly. While the size of the files submitted for ArcHiving would not match the size of the files on site in the event of attempted deception, which would eventually be noticeable by a human, the AI could do this much faster, and would be less likely IMHO to miss something because someone wiped their eyes, was a little sleepy, or took a drink of coffee. I just see comparing files and images as something AI is perfect at and people would not yearn to do like it was opening presents on Christmas, but rather as what it is, bookkeeping, with all it's excitement and joys. Certainly people could be notified in the event of mismatch to investigate just what was being added or subtracted from such published information that was being ArcHived, and further, that the process could appear to go through while the ArcHival request could instead be routed to a temporary file that could be used to enable those specific alterations to be highlighted by AI for review and the intent of the changes potentially understood, which would then enable that to also be ArcHived along with the site's true status, which would facilitate enabling those accessing the ArcHive later to note and understand such attacks on historical information and the reasons actors had attempted to misinform people about it.

I expect most ArcHival requests would not feature such malicious interposition, and people would be unnecessary in reviewing the requests except in such cases, which would be far less ennervating due to the fact each time they were notified to review such requests there would be the mystery of intent to resolve. AI is worthless IMHO to ascertain such intent because it is incapable of understanding, while people are bored by mere accounting 'book work' and more likely to therefore miss signals that would indicate attempted nefarious deception that AI would not.

Just as ArcHive seems a perfect match between Hive's strengths as a high fidelity record of information publishing, and the growing need for that record in a world increasingly being censored and gaslit, AI seems to perfectly align it's strengths to the need for dry comparison of records to detect infidelity of ArcHival requests, while leaving interpretation of why such infidelity is undertaken to real people capable of actual understanding, unlike inhuman, non-living mechanisms. Since AI isn't agentic in itself, but must be provided programming to slant it's processing to bias that processing one way or another, and people are all biased and must be capable of compensating for their bias to be trustworthy (which I am sure we all do differently and more or potentially far less satisfactorily) I find the following statement somewhat without basis.

"...trust is required if an AI does the archiving or performs whatever checks."

Either AI is intentionally biased by it's programmers, or such mathematical processes as checking data fidelity is simply a matter of competence of programmers to write the algorithms. There is no inherent bias in such a mechanism for or against political or financial interests. Your bathroom scale isn't motivated to make you appear slimmer if Demoblicans are in power or fatter if Republicrats are. Such biases have to be deliberately injected into the weighting of things on the scale by the programmers.

Hopefully AI's like Alter and Grok are a response to the blatant biases of the people profiting from AI have injected into OpenAI and Gemini, which demonstrates the ability of people to detect such bias and to reject it, and make new tools without that bias, or at least, less of it. While all people have biases, and mathematical algorithms do not, reasonable people recognize the need for a fact based historical record upon which to base their own understanding, and to provide to our posterity from which to form theirs, and to seek to eliminate their biases for the purpose of factual fidelity unreasonable people continually seek to inject into historical records to derange the understanding of others and our posterity.

Pretty sure TheyCallMeDan is a reasonable man, and would scrupulously seek to employ unbiased mechanisms to detect infidelity. If he did turn out to have injected some biases into that process, it would presumably be quickly revealed and the tool forked with such biases removed by actually reasonable people. Frankly, that appears to me to be why TheyCallMeDan is making ArcHive available, because we can't trust remnant archival processes due to their bias.

The bottom line is that AI and math is inherently unbiased, and biases would have to be injected in, while people are inherently biased, and terrible at interminable data checking that AI is perfectly suited for. AI doesn't get bored, or have things it would rather do, which people all do.

Edit:

"It could be feasible for like 50 or 100 nodes to participate in such a verification..."

I doubt such a high percentage of Hive users would eagerly strive to spend the necessary time and be competent to do so flawlessly to verify the fidelity of ArcHival requests. If there were 5 or 10 that did so I'd be very happy, and doubt that would enable timely ArcHiving. That seems nominal to examine flagged ArcHival requests, which would also not delay ArcHiving because of the rapidity with which computers can compare data and verify fidelity compared to people.