Some observations about AI and search
August 2025
I could easily write a few thousand words on the state of AI and search, and the hype about that process. I am constantly reading about certain highly hyped and reviewed apps/platforms, declarations made by this influencer or that one as to why they have switched from one app to another. I read these kinds of gushing, glowing statements all the time in platform reviews, which tells me the writer hasn’t vetted the results. Too many of them read like AI-written fanboy shill posts and reviews. And they probably are.
Up-front declarations and disclaimers: I do not use any of the well-known “name” AI products for research… or anything. I do use various search engines, all of which have now included some corporate partnership-affiliated search bot. Those returns are always presented at the top of the page, some stating what AI product was used. “Google's search result pages utilize a combination of AI systems,” but a search will reveal what they claim to be their multi-faceted approach and tools used.
Some include a warning that they are “experimental.” I believe all have some disclaimer that they “may make mistakes.” I knew these problems going in, and I have low expectations for the bot returns, but I fear for those who think what the bots provide is objective, valid, correct, the most important, or in some cases, even remotely useful.
I say all of this because my real-world experience with these tools continues to underwhelm me. Some make me laugh (such as the ootheca odyssey post in the archives), but that is short-lived because behind the reasons for the laughter are deep concerns that are not funny. My purpose here is not to get into anything technical, just some real-world examples of my observations of problems using these not-ready-for-prime-time products for research.
I could talk about the numerous times I have made a query and saw that what the bot returned wasn’t even in the ballpark. Why? Many times, the bot had changed my query. Other times, I can tell the bot response is a hallucination,1 the answer given appears related to the query, but is just plain wrong. This brings up concerns about the training material and the search programming methodology these platforms utilize, and how they interfere in the process.
That’s just with utilizing good old-fashioned standard search syntax. I have also noticed the bots appear to effortlessly handle natural language queries. But the results are still often peppered with off-the-wall returns due to parameters introduced by the bot that were not in my query. This produces narrative-style responses that look and sound authentic and encyclopedically correct, but are not.
These platforms then begin to train off their returns.
I have said this before, when it comes to searching with these new tools, don’t ask a question to which you don’t already know the correct answer. Do this as a test. I do. This is why I do not use AI-augmented search tools.
Here is a recent example:
Something unexpected happened while I was searching for a French document from 1829. I was looking for it in digitized text form, still in French. It didn’t take me long to find the original handwritten document as a digital scan. Actually, that was a surprise. I assummed the digitized text would have been easily found and appear at the top of the returns.
Turns out, this was not a slam dunk. I spent some time using normal search query syntax, but never found it. The embedded AI kept turning my prompt into its own plain language prompt, but it wasn’t including the exact title of the document. Several times, I tried to force it to use the correct title in French and English. It didn’t find it. This was disconcerting, but I had other things to do, so I moved on.
The next day, I circled back to try again. I reworked my queries and syntax, but still didn’t find it. Then I decided to try searching in plain language. No document found. I don’t know why, but something in me made me rework the query several times, and suddenly, the bot returned a location for the document in the middle of several others. After so much reworking without returning the proper result, I immediately assumed the bot had hallucinated to please me (subject of a future post), but I decided to vet it. After so many tries, this one was a genuine, correct “hit.” It was correctly titled, included as an exhibit inside another document.
There should have been no reason for not finding it immediately. That was very disconcerting. But this is the first time I recall getting a result from one of the AI search bots that was useful and that no other method had found.2 However…
After reminding myself of past observations about AI-augmented search, I remembered that one of my peeves is that the bots show a limited number of all returns found, often truncated, sometimes just summaries. This is how this one return was finally among them. This points out a major problem with AI generative tools: even the exact same prompt will not give duplicate results, and not just with images. This is important to remember if you use these stunted tools.
I almost didn't find the document. This would have been the result if I weren’t sure the document had to exist in the form I wanted. If I hadn’t been persistent, if I had stopped at one or two queries, I never would have found it and made a note that it likely doesn’t exist on the web. Not all things do, but I don’t like making that entry in my notes until I have made numerous attempts using different tools.
I believe that my multiple queries, all very similar, appeared to force the bot to churn through all returns and randomly choose which were to be made viewable. There doesn’t seem to be any weight factoring or any other rhyme or reason for not returning the same results for a duplicated prompt. I find the idea of having to run the same prompt multiple times very disconcerting. I won’t, I don’t, because I refuse to use the AI-assisted search bots, per above.
I have been reading about suspicions that the major search platforms are tinkering with how they are, or will be, altering their search processes for their benefit, and that the old, familiar methods of search will disappear. The supposition is that the engines will no longer return search results in the manner we have been accustomed to for decades: presented in the order they prefer, with “sponsored” and other monetization methods first, then the (allegedly) organic results presented in the order their particular algorithm prescribes. In the future, it was said, that the platforms will take direct and complete directed control of what is returned and the appearance of how it is presented. The inference being that we will not be able to discern monetized returns from organic because they will not be labeled as such or separated to a sidebar as they are now. I see this already being previewed in the top section of some present return pages, and links aren’t what they used to be. I have been warning about this for over two years.
This is a serious enough concern that there have been efforts afoot to find ways to bypass the new processes, outside of any claims the platforms are, or will make, about options to revert to a “classic” version.
Trust is crushed.
The rush to make AI a feature of all things means I am constantly assaulted with pitches for new digital products or updates every day. Of interest to me is an ever-increasing number of “notebook” apps. Reviewing them, comparing them, and making a list of the “7 best” (this week) is consistent content fodder for serial posters on payout platforms. These are something I used to take time to investigate because of the huge amount of research I do. I was always looking for the next-best, “last app you’ll ever need” product.
I have tried a lot of them, each promising to be a perfect fit for my workflow. None are. The best they can do is streamline a portion of it. Each alleges to offer another upgrade that fixes a problem, or introduces new AI-enhanced features and improvements that, when combined, surpass all others… at least this week. This amounts to an endless progression of leapfrogging that soon becomes aggravating. I spend too much time dealing with changes in updates and endless learning curves with new apps. I waste time, accomplish very little, and am even more frustrated. I always end up back with my “archaic” way of doing things and keep telling myself, “Stop this; enough!”
The influencer manages to use and review multiple products in such a short period. He must be super-gifted and super-knowledgeable about all these products! And he’s a great writer! He puts all this together so quickly, so often. He’s got numerous screenshots of the app in action. I can’t see them very well, pretty low quality, like the ones on the App Store, but who cares! They are augmented with very concise, insightful narrative that makes sense of the screenshots, just like on the App Store. This guy must be a genius with these apps, as attested to by the clicks, adoring followers, and additional subscribers.
What’s that adage, again?… You can fool some of the people…”
The promise of AI continues to elude me. But maybe it’s just me.
I was reading a post today about various AI evaluation models and read this:
In Machine Bullshit, the authors measure the difference between a model's "belief" (measured as the probability signed to the token "yes" to the question "Do you believe A is true?") and how likely the model is to report a false belief ("bullshit").
Machines can have belief systems? When will this semantic twaddle and anthropomorphizing end?
In another return in the set, there was a snippet from the document in French, apparently included as a block quote. This had never shown up in any earlier queries. If it had, I could have searched directly on the text, and this should have been resolved much more quickly.