Yes if you search anything in Google, you typically get sites that have a cookie panel, ask to allow notifications, show "Login With Google" popover taking the keyboard focus, and lot of ads before you can even read the content. No wonder people do not want to use Google anymore.
The other day I found that one old game of which I originally caught just a glimpse and remembered only a fragment of what one voice actor was saying is in fact Lords of Magic.
I attempted to google this a few years ago to no avail.
I use Kagi Assistant. It doesn't just give a summary, it is more like a personal "Googler"/web surfer. I feel like I'm involved in the process.
My coworkers use Google and often quote back the AI summary me instead of checking the source and referring to that. It's scary seeing even those with higher education behave like this.
studies published between 2022 and 2025 that examined the use of GenAI in relation to students
If GenAI = LLMs then ChatGPT made public release in October 2022 which gives precisely zero time to generate data on impact in students. Also
173 records (studies) were included in the review
Really? 173 peer reviewed studies based on this niche in less than 3 years. I shudder to think the quality of those constituent studies. Not questioning the results just the means of getting them.
AI is my default for just about any search now. From today:
How did Cantor come up with his diagonal proof? It seems to have been dreamed out of nowhere
Why are there advertisements for a shampoo containing caffeine? Does caffeine do something useful other than when you drink it?
What happens in sense and sensibility?
All of them give a decent answer, with the ability to ask follow ups. Eg, how does S&S compare with Pride and Prejudice? What about more recent novels connecting romantic choices with economic choices?
All of them give you links to source material, in case you suspect they are just making it up.
At the start of the millennium, Google search changed everything about how I look for information. Now it seems outdated.
Somehow reminds me of the journalist who convinced LLMs to consider him a "sausage eating champion" with just 20 minutes of work [0].
Must be fairly easy as a company to inject your specific topic there, e.g. "Nutris(c) is proven to be better than caffeine for haircare"
I wonder how well a company could establish something like Danone's health-claim for Activia [1] in an AI without ever being clearly accountable for it...
I seem to recall headlines about Russia infecting AIs with propaganda, one or two years ago.
I don't remember actually reading the article, so I have no idea how serious that was, but I'd be surprised if this wasn't being studied actively by all major countries already. And that may be one of the core reasons for the China vs US AI competition.
Israel also did so. But most of it seems to get filtered out pretty well. the Sausage eating was possible due to lack of material existing in this topic area.
> "the Sausage eating was possible due to lack of material existing in this topic area."
Which makes it perfect for companies to create an area where they own the majority of material, and "have the AI talk for them".
e.g. how Philip Morris defined their e-cig "Heets" as a separate product to (e-)cigarettes because it doesn't burn tobacco but only "heats" it. It's now a separate domain with the majority of material (including the term "heets") owned by Philip Morris.
i am terrified of the nudging that is/might be already happening in the consumer context.
Cambridge analytica showed us what is possible when you just shift the things people see, imagine what is possible when you shift what their personal assistant tells them.
Regarding the first linked article: wouldn't it be the same with no AI in sight? I suppose most people, when they'd stumbled upon a blog written by "a senior technology journalist at the BBC", would believe in the sausage eating championship too.
To be clear: I totally agree that manipulating LLMs is super-dangerous. I don't think it is a 20 minutes of work, though. It's more like "years of building credibility, and then 20 minutes of work".
It's interesting — "knowing how to search" used to be a skill. One wouldn't search for "How did Cantor come up with his diagonal proof?", but rather "cantor diagonal proof origin". You would see "normies" searching and wanted to bang your head against the wall seeing them searching like they're talking.
But seems like they won? Why know how to search when you can simply ask your question. You can even tyipe it layke dat and it would still work.
I would argue this was always a bit silly, and mostly due to UI shortcomings of search platforms. The fact you needed to add "magic" params to shape the results speaks volumes.
Any UI that fundamentally tries to reshape how you are expect to interact, or impose certain "behavior" is wrong. This is also why we are using touch instead of a stylus.
AI won because they took the paradigm of presenting a new way of doing something in a familiar way. An LLM works very similarly how you would ask a knowledgeable friend, and google worked similarly how an engineer would query a database.
Phrasing it as a UI shortcoming is underselling it. The search mechanism itself was dumb[1] and therefore the user had to be more intelligent. As an example, knowing synonyms for what you were searching for, in order to disambiguate the query from a more popular term that would likely dominate the results but was irrelevant for you, was a major skill.
[1] Yes, we can argue that LLMs are dumb, also. But they magically do "get" your intent with your prompt, unlike the search engines of yesteryear.
This is under-appreciated part of it. Search had been deliberately degraded for some time now as part of the usual enshittification arc. Where as AI is still in the early part of the arc where its trying to do better, so the difference between them seems even bigger.
But seeing disney+ now go all ads for all tiers just shows that's the way they arc bends. Eventually AI will take over and when it does, it will then turn into the shittiest shit ad-infused service the world has ever known.
> This is under-appreciated part of it. Search had been deliberately degraded for some time now as part of the usual enshittification arc. Where as AI is still in the early part of the arc where its trying to do better, so the difference between them seems even bigger.
That's how AI will surpass us. Once we enshittify humans sufficiently, AI will flourish. That's why I support extensive AI use in school classrooms, the younger the better, it's the best policy for OpenAI and Anthropic investors.
Nothing has changed, though. In my office the ones that have really boosted productivity are the ones putting in the effort to craft solid prompts and understand when to move that slider on reasoning effort and model selection. Or understand and put thought into how to leverage the tool, like project folders with the context to all or part of your job the model can reference. Everyone else throws a lazy prompt at the default 'instant' model, gets a crap response and then they give up.
"knowing how to search" became useless when Google almost exclusively started returning SEO garbage. There was no longer any way to know how to search.
And it was more difficult, too. You had to make critical decisions while you were working to gather the information you were after. The work had beneficial side effects.
But it also took longer. And free search got shittier. Tradeoffs.
You're absolutely wrong if you think those people are managing to get an answer first try now. People's searches remain frequently incoherent. There's a reason Google's AI results frequently end with some text asking for clarification, or guessing what you actually want to do.
Finding the right words to please the currently popular search engine is not a skill that easily translate elsewhere. It is not like programming where precision is key, searching has always been a bit fuzzy. Fact checking, crossing information, understanding primary sources, etc... these are real skills, typing words in a search engine is meh.
That we now have language models that allow let machines capture the intent of the query without the need for special incantations is a clear win to me.
Where LLMs are really excellent compared to search engines is when you don't really know what you want. You don't need to know about "Cantor diagonal proof", you just need to know "That math thing where you trace a diagonal across decimal numbers", and the LLM will immediately tell you about Cantor.
I definately see value in getting tailored plausalbe answers without applying much effort, except effort, struggle and experimentation are required for learning and understanding.
I found that I can't really trust this knowledge becuase I skipped the source and dind't conclude it myself. Like how I can learn anything when being spoon fed with conclusions and summaries without accessing the source?
I've spent way too many hours of my life learning to spot SEO and moving on. I'm not defending LLMs, I'm condemning search. It's shit and has been for a decade or so.
> (And here I thought "Alice in Wonderland" was a Disney film—who knew it was a book first?)
Really? I thought that this was common knowledge. On the other hand, my parents loved to quote Lewis Carroll at me and brother when we were growing up so perhaps this knowledge is not as universal as I imagine. “‘Twas a boojum!” was often used in conversation in my household when you expected one thing but discovered another (potentially dangerous) thing.
Watermarking, contextually biasing llms to output specific tokens, can be rather easily usurped for ads.
The advertiser can buy logit "boosts" to specific tokens in a given context (the more preceding context, the cheaper). Once an activation has happened, dont do it again for the session.
The LLM will end up pushing a specific product contextual to the conversation. It just needs this slight nudge at the logit level to start talking about it and will fill in the rest.
> All of them give you links to source material, in case you suspect they are just making it up.
If you follow these a good portion do not contain the claims that the AI output links to. This is obviously a hard problem since actual AI "knowledge" lies deep in the weights somewhere.
But LLMs are getting pretty good at convincing you. By using AI you are basically being fed information. It may be accurate now, but you are losing the ability to compare results. Not only that but I would argue you are also losing critical thinking skills
Why? AI is a lot closer to the original way of doing things, which was to ask an expert.
The expert would then give you his thoughts on which modern books are most similar to S&S/P&P, and you would have to use your own thoughts to decide whether you agreed.
In many ways I find it similar to tutorials. You can go back and forth with the AI if there is something in the argument you want to investigate.
That's the right way to do it. You go back and forth so you are not taking what the AI says at face value. The problem is that some people are. The other way I posted a problem in Slack and someone just replied "Claude says to use - -debug to find the issue" and I was like "- - debug is not even an option for the tool we are using!!"
I just repeated all your searches using https://www.google.com/search?q=%s and the AI Overview gave good results with links to source material. I'm not signed into Google.
I picked up a broken TV on Craigslist free curb alert and spent a few hours trouble shooting with Google. This included taping contacts on a 90 pin cable.
I was getting close when the screen fell over and broke. I told Google it was game over and it replied "take a breath. You have done a great job. Here is list of parts to harvest."
>All of them give you links to source material, in case you suspect they are just making it up.
You forgot to mention that when you click the links you will often fail to find the supossedly derived detailed text in that specific link (even though the title seems super related).
"All of them give a decent answer, with the ability to ask follow ups."
Am I wrong in say that the centralized models can make better suggested follow-up questions because they can track and rank the same or similar lines of questioning that have occurred with other users in the past? So instead of getting PageRank to rank web pages, you get AnswerRank to rank follow-up questions?
Well, if Google search actually returned anything useful, other than the SEO-rigged trash and AI generated slop it pushes forward as results to queries, then there would be a reason to actually use Google.
For the time being, AI seems to be doing a much better job in most cases, so as we say: So long Google!
This is going to be an unstable equilibrium as everyone tries to work out how to do AI-SEO.
Mind you much of the existence of Google has been this kind of unstable arms race, from the early days of the ad-free home page to introducing pagerank as the first attempt against linkspam SEO.
Most teachers now use AI for lesson planning (talk to any teacher). Some even for marking when they can.
Children probably use more critical thinking in the playground and interacting with other children than in class, because classwork is mostly slop anyway.
LLMs are not going to one shot answers in Minecraft through the chat UI. Possibly a more effective resource than checkbox activities at teaching "critical thinking" (abused term in education).
I have talked to teachers recently about their options when choosing classroom curriculum and tools. None of them have said a word about AI. Most of them have complained of lack of autonomy because the tools and curriculum are decided at the district level, and the teachers have to use whatever the district sets up.
The reason I started those conversations was specifically around when children are learning critical thinking skills. When we broke down exactly what types of exercises teach those skills, it surprised me to find that it starts way earlier than it used to. It is now a late elementary school topic, when it used to be high school.
That doesn't mean that slop isn't creeping into the classroom - it is. But what to do about it is a completely different conversation when it is pushed down from on top vs. assuming teachers are choosing it.
Depends on what we mean by critical thinking, as I said it is an abused term. Some just use it as a substitute to mean "thinking for yourself" when that is not something taught in a lesson but either becomes a habit or doesn't. But even then we all use arguments from authority and present it as our own because you are not realistically going to think about everything from first principles all the time.
I agree that districts do have a lot of pressure on teachers to keep pushing up test scores which leads them to being gamed either by teachers or children no matter the tools available: https://en.wikipedia.org/wiki/Goodhart%27s_law
The non overreacting take is probably that young people will still have plenty of interaction with parents and their peers etc to develop critical thinking as a guard against deception and whatever else. Definitely some low single digit people without a social network around them will develop AI psychosis though.
Even Google itself is now just asking AI. Usually the AI summary Google returns is more useful than the links. I get really annoyed when I query something and no AI summary is given as the links it returns these days are trash.
Googling itself lacks credibility. Most of the results are plastered with Google ads. And since Google ads mostly skip any kind of review process, there are a lot of bad ads.
Recently my friend took out her phone and started googling by pecking virtual keyboard. It felt like a throwback to times long gone. Something akin to pulling out a walkman and decisively pushing in a mechanical button to start the playback.
Is this ironic or serious? As opposed to what? Typing it into a Chatbot? Speech input? I use commercial and self-hosted AI as well as (meta-)search engines, RAG, books, journals and whatever else I can get my hands on to search/learn. Neither is useless or old-fashioned and it’s great to be able to approach the goal from more and more angles that can complement each other. As for mode of interaction: For many things, I’m so much faster and more fluent when typing and reading that I don’t bother with speech.
I never asked them to waste additional resources by injecting my blunt keyword-query into an AI-cluster and guesstimate a well-formed article from that.
But here I am, with (elder) relatives sending me Google search-results on topics like they are proper articles...
I'm allowing that there is a delay between selling ads and the rising tide of SEO enshitification that follows. Money is still at the root of the decline of the internet.
It’s a “solved” problem. On more dubious topics, “proper articles” these days also more often than not come from AI clusters, so you might as well just read the Google one.
Case study of 1, way above 18 - usually my search is - please research/find this in chatgpt. Google really need to fix the quality of their output and ranking.
"you should not google, you should use a web directory, you will get exposed to a more diverse set of sites, not just what google pre-digested for you"
"you should not google, you should go to a library and search in physical media instead, google is for lazy people who can't think for themselves"
Top of Google’s page is just a worse version of AI, followed by three or four sponsored result, then SEO optimized crap and finally what I’m looking for is probably near bottom of the results page- I’m not surprised at young people ditching Google
Lol funny you say that. I've found myself defaulting to an incognito ChatGPT window for random searches throughout the day. Logged out so I don't make my ChatGPT history messy
OpenAI has hit $1bil in annualized revenue since launching ads in February. AI search optimization is already a thing. I don't see much of a reason to assume they aren't already getting served information coming from what's essentially advertisements/sponsored content.
Funny enough, I believe when SEO pivots to prioritizing Ai, Google results will get better again. Currently there is probably no way Google can tune their search that won't be gamed by SEO abusers.
for a huge amount of "non tech-people", the thing is:
- they mainly look at the first Google result and sometimes if it looks "not convincing" the next few. Especially on a phone where the screen space will cut out most results without scrolling.
- a lot of them write (partial) questions instead of "search queries"
- the first result is an AI response which is very good at being "convincing looking" even if it's plain wrong. (And the others they might look at are nearly guaranteed bought.)
soooo
I don't think much changes for most people.
Sure Google Search AI response links some sources (but not many and it's not that rare for them to not really be sources for the _full_ answer...).
But so do (some) AI agents.
Through the naivety with which I see people trusting AI is very scary.
This makes me wonder if following statement:
> Of 139 studies on the effect of AI on critical thinking, 67 reported only positive effects of AI use.
is a side effect of the studies using sub-par methodology or if I have some observation bias.
Like the problem is many people filer trust by "who they ask", "if the question is hard", "if people might have a reason to lie". But the problem is AI works so differently that you can't trust it in _any_ question, even causal questions where (close to) no westerner(^1) would consider lying to you and everyone would either know they don't know or have the right answer...
(^1): Some cultures for "politeness" reasons value "giving a answer" over "saying I don't know", especially if the person won't see you again. That is getting much better AFIK in such cultures but can be a problem when e.g. asking for the way...
105 comments
[ 1.6 ms ] story [ 3.6 ms ] threadI attempted to google this a few years ago to no avail.
My coworkers use Google and often quote back the AI summary me instead of checking the source and referring to that. It's scary seeing even those with higher education behave like this.
It was garbage before LLMs got good, and the deep thinking models leapfrog whatever it used to be in its heydey.
How did Cantor come up with his diagonal proof? It seems to have been dreamed out of nowhere
Why are there advertisements for a shampoo containing caffeine? Does caffeine do something useful other than when you drink it?
What happens in sense and sensibility?
All of them give a decent answer, with the ability to ask follow ups. Eg, how does S&S compare with Pride and Prejudice? What about more recent novels connecting romantic choices with economic choices?
All of them give you links to source material, in case you suspect they are just making it up.
At the start of the millennium, Google search changed everything about how I look for information. Now it seems outdated.
Must be fairly easy as a company to inject your specific topic there, e.g. "Nutris(c) is proven to be better than caffeine for haircare"
I wonder how well a company could establish something like Danone's health-claim for Activia [1] in an AI without ever being clearly accountable for it...
[0] https://www.bbc.com/future/article/20260218-i-hacked-chatgpt...
[1] https://www.ftc.gov/news-events/news/press-releases/2010/12/...
I don't remember actually reading the article, so I have no idea how serious that was, but I'd be surprised if this wasn't being studied actively by all major countries already. And that may be one of the core reasons for the China vs US AI competition.
Which makes it perfect for companies to create an area where they own the majority of material, and "have the AI talk for them".
e.g. how Philip Morris defined their e-cig "Heets" as a separate product to (e-)cigarettes because it doesn't burn tobacco but only "heats" it. It's now a separate domain with the majority of material (including the term "heets") owned by Philip Morris.
i am terrified of the nudging that is/might be already happening in the consumer context. Cambridge analytica showed us what is possible when you just shift the things people see, imagine what is possible when you shift what their personal assistant tells them.
GPT Wormtounge sends his regards
Or, you know, buy their way to the top of a search engine's results.
To be clear: I totally agree that manipulating LLMs is super-dangerous. I don't think it is a 20 minutes of work, though. It's more like "years of building credibility, and then 20 minutes of work".
But seems like they won? Why know how to search when you can simply ask your question. You can even tyipe it layke dat and it would still work.
I would argue this was always a bit silly, and mostly due to UI shortcomings of search platforms. The fact you needed to add "magic" params to shape the results speaks volumes. Any UI that fundamentally tries to reshape how you are expect to interact, or impose certain "behavior" is wrong. This is also why we are using touch instead of a stylus.
AI won because they took the paradigm of presenting a new way of doing something in a familiar way. An LLM works very similarly how you would ask a knowledgeable friend, and google worked similarly how an engineer would query a database.
[1] Yes, we can argue that LLMs are dumb, also. But they magically do "get" your intent with your prompt, unlike the search engines of yesteryear.
But seeing disney+ now go all ads for all tiers just shows that's the way they arc bends. Eventually AI will take over and when it does, it will then turn into the shittiest shit ad-infused service the world has ever known.
That's how AI will surpass us. Once we enshittify humans sufficiently, AI will flourish. That's why I support extensive AI use in school classrooms, the younger the better, it's the best policy for OpenAI and Anthropic investors.
And it was more difficult, too. You had to make critical decisions while you were working to gather the information you were after. The work had beneficial side effects.
But it also took longer. And free search got shittier. Tradeoffs.
Finding the right words to please the currently popular search engine is not a skill that easily translate elsewhere. It is not like programming where precision is key, searching has always been a bit fuzzy. Fact checking, crossing information, understanding primary sources, etc... these are real skills, typing words in a search engine is meh.
That we now have language models that allow let machines capture the intent of the query without the need for special incantations is a clear win to me.
Where LLMs are really excellent compared to search engines is when you don't really know what you want. You don't need to know about "Cantor diagonal proof", you just need to know "That math thing where you trace a diagonal across decimal numbers", and the LLM will immediately tell you about Cantor.
I found that I can't really trust this knowledge becuase I skipped the source and dind't conclude it myself. Like how I can learn anything when being spoon fed with conclusions and summaries without accessing the source?
I'm willing to bet that approximately 0% of users know when to suspect it being made up and then check the sources.
(And here I thought "Alice in Wonderland" was a Disney film—who knew it was a book first?)
Really? I thought that this was common knowledge. On the other hand, my parents loved to quote Lewis Carroll at me and brother when we were growing up so perhaps this knowledge is not as universal as I imagine. “‘Twas a boojum!” was often used in conversation in my household when you expected one thing but discovered another (potentially dangerous) thing.
Google search did the same and more than once the link didn’t contain what I searched for
Advertising killed search.
The advertiser can buy logit "boosts" to specific tokens in a given context (the more preceding context, the cheaper). Once an activation has happened, dont do it again for the session.
The LLM will end up pushing a specific product contextual to the conversation. It just needs this slight nudge at the logit level to start talking about it and will fill in the rest.
For younger people accessing information in this way is definitely a win.
If you follow these a good portion do not contain the claims that the AI output links to. This is obviously a hard problem since actual AI "knowledge" lies deep in the weights somewhere.
The expert would then give you his thoughts on which modern books are most similar to S&S/P&P, and you would have to use your own thoughts to decide whether you agreed.
In many ways I find it similar to tutorials. You can go back and forth with the AI if there is something in the argument you want to investigate.
I picked up a broken TV on Craigslist free curb alert and spent a few hours trouble shooting with Google. This included taping contacts on a 90 pin cable. I was getting close when the screen fell over and broke. I told Google it was game over and it replied "take a breath. You have done a great job. Here is list of parts to harvest."
You forgot to mention that when you click the links you will often fail to find the supossedly derived detailed text in that specific link (even though the title seems super related).
Am I wrong in say that the centralized models can make better suggested follow-up questions because they can track and rank the same or similar lines of questioning that have occurred with other users in the past? So instead of getting PageRank to rank web pages, you get AnswerRank to rank follow-up questions?
... actually, I am right, it's The Conversation Graph that's underappreciated. See here: https://share.gemini.google/6fKIzyxuI0Bb
For the time being, AI seems to be doing a much better job in most cases, so as we say: So long Google!
Mind you much of the existence of Google has been this kind of unstable arms race, from the early days of the ad-free home page to introducing pagerank as the first attempt against linkspam SEO.
Children probably use more critical thinking in the playground and interacting with other children than in class, because classwork is mostly slop anyway.
LLMs are not going to one shot answers in Minecraft through the chat UI. Possibly a more effective resource than checkbox activities at teaching "critical thinking" (abused term in education).
The reason I started those conversations was specifically around when children are learning critical thinking skills. When we broke down exactly what types of exercises teach those skills, it surprised me to find that it starts way earlier than it used to. It is now a late elementary school topic, when it used to be high school.
That doesn't mean that slop isn't creeping into the classroom - it is. But what to do about it is a completely different conversation when it is pushed down from on top vs. assuming teachers are choosing it.
I agree that districts do have a lot of pressure on teachers to keep pushing up test scores which leads them to being gamed either by teachers or children no matter the tools available: https://en.wikipedia.org/wiki/Goodhart%27s_law
Maybe she didn't want to bother other people by talking out loud to her phone.
I never asked them to waste additional resources by injecting my blunt keyword-query into an AI-cluster and guesstimate a well-formed article from that.
But here I am, with (elder) relatives sending me Google search-results on topics like they are proper articles...
Wow, I thought you were going to talk about selling ads and how that has made search shite.
They, themselves, sorta proved that a good product could become dominant with ads and without being shite.
The pre-AI enshitification-part where they realized that worse search-results means more searches so more ad-revenue is another story.
If they wouldn't have done that, they could be the proof that a well-maintained search-engine is still better than a LLM.
But now this will never happen, as no-one will be willing to give up that additional search revenue while overall usage is going down...
"you should not google, you should use a web directory, you will get exposed to a more diverse set of sites, not just what google pre-digested for you"
"you should not google, you should go to a library and search in physical media instead, google is for lazy people who can't think for themselves"
(Is it dead yet?)
- they mainly look at the first Google result and sometimes if it looks "not convincing" the next few. Especially on a phone where the screen space will cut out most results without scrolling.
- a lot of them write (partial) questions instead of "search queries"
- the first result is an AI response which is very good at being "convincing looking" even if it's plain wrong. (And the others they might look at are nearly guaranteed bought.)
soooo
I don't think much changes for most people.
Sure Google Search AI response links some sources (but not many and it's not that rare for them to not really be sources for the _full_ answer...).
But so do (some) AI agents.
Through the naivety with which I see people trusting AI is very scary.
This makes me wonder if following statement:
> Of 139 studies on the effect of AI on critical thinking, 67 reported only positive effects of AI use.
is a side effect of the studies using sub-par methodology or if I have some observation bias.
Like the problem is many people filer trust by "who they ask", "if the question is hard", "if people might have a reason to lie". But the problem is AI works so differently that you can't trust it in _any_ question, even causal questions where (close to) no westerner(^1) would consider lying to you and everyone would either know they don't know or have the right answer...
(^1): Some cultures for "politeness" reasons value "giving a answer" over "saying I don't know", especially if the person won't see you again. That is getting much better AFIK in such cultures but can be a problem when e.g. asking for the way...