From what I've seen over the last few years, most people will look at an LLM-based tool demo, where, e.g., a chat interface lets you search for "All PDF invoices from last month over 5MB" and they will <i>never wonder whether they've been shown all of the desired results</i>. That is, while my immediate thoughts are always "how did they achieve this?" or "which folders are they searching?" or "how can I be sure that my invoices are actually in those folders?" or "how does it even identify an invoice?", most people will just delight in the list of results. People don't ask themselves whether something is correct. They never have! Their scam radars are completely broken. They've always implicitly trusted the eminently and provably fallible heap of shambolic software we've always produced. That they are now being brainwashed into used <i>non-deterministic</i> shambolic software for absolutely everything increases the fallout of this fallibility. The tool they've used has established a fact in their world. Any documents that should have matched the search but did not, <i>no longer exist.</i> Because an infallible tool---in their minds---has established a <i>fact</i>, they no longer seek further. Those documents are gone. The information not within the LLM-based tool's purview <i>effectively no longer exists.</i> This isn't new, of course. To a lesser degree, this was the plague instilled on us by the oracle of search results decades ago. This psychological tendency was already being manipulated but, until the advent of LLM-based tools, all of this admittedly also terrible software was <i>deterministic</i>, meaning that, if it had been configured correctly, it was always going to deliver the expected results, the same way, every time. If you didn't receive the results that you'd expected, you could "debug" to determine which initial condition had not been satisfied: e.g., the folder where you put invoices wasn't in the search folders, or you weren't actually putting all of the invoices in one of those folders, or you were producing invoices that didn't match the pattern that you were using to identify invoices (e.g., files starting with "invoice_"), or whatever. These solutions may all sound technical, but they're also <i>solvable</i>. Fix the input and the output will be fixed, every single time. With LLMs, all inputs are filtered through a non-deterministic filter, with what were formerly hard, technical limits now expressed as hopeful instructions that are taken as suggestions. The results <i>might</i> include <i>most</i> of the files you were looking for---which is kind of amazing for a natural-language search---but they simultaneously eliminate any documents from the user's universe that should have matched the search but did not because LLMs are "sometimes wrong" . People will only notice when the shit hits the fan, at which time it will be no-one's fault because everyone did what they were told to do in the way that they'd been told to do it. <img src="{att_link}this_document_exists,_while_the_others_do_not.webp" href="{att_link}this_document_exists,_while_the_others_do_not.webp" align="none" caption="This document exists, while the others do not" scale="35%">