My personal intuition is that I'm able to retain more detail about things that I've casually queried ChatGPT about, compared with search. The sense I have is that the effect is similar to reading a paper book, compared to Kindle.
Was having a version of this discussion this evening--take as given that we all hold flawed models of the world. Does AI help us achieve particular aims more or less effectively than some other tool (search, talking to "the best available human", guessing, etc)? Pragmatic in the William James sense.
Yeah, exactly! And I was just adding a footnote that it really isn't even about propositional belief in a narrow sense, but an understanding of the structure of the question, the discourse, and the evidence. The belief isn't even the target. The bad searches are bad because they hide a wealth of evidence accumulated since the 1930s, the good searches show the initial evidence from the 1930s and the significant rebuttals since then, which is what makes them good.
Of course looking at "technology in use" is important - it's the way (as STS scholars remind us) to avoid simple tech determinism. Still, design counts: traditional search uncovers original sources and Larry Page's engine was inspired by the practices of academic citation. LLMs oth r not modelled on those citational practices (even though they appear to do that by inserting ex post facto citations). This is a key historical and design difference and it shouldn't b glossed over when examining the use of generative AI in academic contexts.
I don't think I'm glossing that over. But ultimately we should see some element of that reflected in the understanding (broadly) of the user using it.
Now, the problem is that we can't test things perfectly, so absolutely the method of production should inform some intuitions about what we should see with understanding at scale. It's similar to nutrition, where the research environment is so confounded that it makes sense to ask questions like "what did we evolve to eat?" to develop some starting points on what a working theory should look like.
But they are ultimately intuitions and we should work to push on them. As we've seen over the past 30 years, these things are not static. Academic citation formed a basis of Google, which subsequently shaped SEO, which shaped Google. The first Google bombs show up in the mid-2000s, exploitation of data voids starts as a marketing technique then moves to propaganda, paywalls drop in the 1990s creating rich traceable networks of citation. They develop search quality rating. Blogosphere comes up, cheap results mess with Google, they go the EAT model, things get better for a bit. Paywalls come back up in late 2010s, and now EAT leads to stuff you can't read. They go to EEAT to broaden it out, and try to catch the better free stuff. The whole thing is just very dynamic, so while design matters I just want to note that the environment it exists shifts constantly and that initial design of Google is just a sliver of what Google search is now. This is basic STS too but these things don't resolve to simple definitions of mechanism (whether "stochastic parrot" or "bibliometrics") once the dynamism of the full system comes into play.
Thanks for that considered reply. And the reminder that if academic citation practices played a role in the original design of Google search it's [d]evolved since then. Moreover the opacity of the search and of llms also pose challenges in our ability to evaluate them. Still, isn't it easier to evaluate the provenance of a truth claim in a web page that is surfaced by search (however biased that search may be) then in synthetic text that is generated by an LLM? We can look at the pages it cites. Or the pages that cite it. We can divine its context or in Shah and Benders language we can "sitúate" it in ways that are not as readily available when evaluating synthetic text. The provenance of information is something that needs to be conserved. And while search could do it better than it does currently it still does it better than LLMs do. That at least is what I think Bender and Shah are getting at in Situating Search. Are you persuaded by what they argue there?
A similar issue happens with both current and older AI. If you ask ChatGPT for the benefits of drug X or the side effects of drug X, that’s all it gives you—one-sided information. I was discussing this with someone a few months ago, and I told them to reframe their prompt.
For example, instead of asking “What are the benefits of drug X?” or “What are the side effects of drug X?”, try “What are the impacts of drug X?” That tends to give a more balanced answer.
And to the bigger question of what’s better: AI or search? Well, research* has shown that AI is less dependent on a person’s skill or education, it’s faster, and it tends to be more satisfying to use.
My personal intuition is that I'm able to retain more detail about things that I've casually queried ChatGPT about, compared with search. The sense I have is that the effect is similar to reading a paper book, compared to Kindle.
Was having a version of this discussion this evening--take as given that we all hold flawed models of the world. Does AI help us achieve particular aims more or less effectively than some other tool (search, talking to "the best available human", guessing, etc)? Pragmatic in the William James sense.
Yeah, exactly! And I was just adding a footnote that it really isn't even about propositional belief in a narrow sense, but an understanding of the structure of the question, the discourse, and the evidence. The belief isn't even the target. The bad searches are bad because they hide a wealth of evidence accumulated since the 1930s, the good searches show the initial evidence from the 1930s and the significant rebuttals since then, which is what makes them good.
Of course looking at "technology in use" is important - it's the way (as STS scholars remind us) to avoid simple tech determinism. Still, design counts: traditional search uncovers original sources and Larry Page's engine was inspired by the practices of academic citation. LLMs oth r not modelled on those citational practices (even though they appear to do that by inserting ex post facto citations). This is a key historical and design difference and it shouldn't b glossed over when examining the use of generative AI in academic contexts.
I don't think I'm glossing that over. But ultimately we should see some element of that reflected in the understanding (broadly) of the user using it.
Now, the problem is that we can't test things perfectly, so absolutely the method of production should inform some intuitions about what we should see with understanding at scale. It's similar to nutrition, where the research environment is so confounded that it makes sense to ask questions like "what did we evolve to eat?" to develop some starting points on what a working theory should look like.
But they are ultimately intuitions and we should work to push on them. As we've seen over the past 30 years, these things are not static. Academic citation formed a basis of Google, which subsequently shaped SEO, which shaped Google. The first Google bombs show up in the mid-2000s, exploitation of data voids starts as a marketing technique then moves to propaganda, paywalls drop in the 1990s creating rich traceable networks of citation. They develop search quality rating. Blogosphere comes up, cheap results mess with Google, they go the EAT model, things get better for a bit. Paywalls come back up in late 2010s, and now EAT leads to stuff you can't read. They go to EEAT to broaden it out, and try to catch the better free stuff. The whole thing is just very dynamic, so while design matters I just want to note that the environment it exists shifts constantly and that initial design of Google is just a sliver of what Google search is now. This is basic STS too but these things don't resolve to simple definitions of mechanism (whether "stochastic parrot" or "bibliometrics") once the dynamism of the full system comes into play.
Thanks for that considered reply. And the reminder that if academic citation practices played a role in the original design of Google search it's [d]evolved since then. Moreover the opacity of the search and of llms also pose challenges in our ability to evaluate them. Still, isn't it easier to evaluate the provenance of a truth claim in a web page that is surfaced by search (however biased that search may be) then in synthetic text that is generated by an LLM? We can look at the pages it cites. Or the pages that cite it. We can divine its context or in Shah and Benders language we can "sitúate" it in ways that are not as readily available when evaluating synthetic text. The provenance of information is something that needs to be conserved. And while search could do it better than it does currently it still does it better than LLMs do. That at least is what I think Bender and Shah are getting at in Situating Search. Are you persuaded by what they argue there?
* "it shouldn't b glossed over when examining the differences between generative AI and search in academic contexts."
A similar issue happens with both current and older AI. If you ask ChatGPT for the benefits of drug X or the side effects of drug X, that’s all it gives you—one-sided information. I was discussing this with someone a few months ago, and I told them to reframe their prompt.
For example, instead of asking “What are the benefits of drug X?” or “What are the side effects of drug X?”, try “What are the impacts of drug X?” That tends to give a more balanced answer.
And to the bigger question of what’s better: AI or search? Well, research* has shown that AI is less dependent on a person’s skill or education, it’s faster, and it tends to be more satisfying to use.
* https://jakobnielsenphd.substack.com/p/seo-is-dead
And
https://jakobnielsenphd.substack.com/p/search-vs-ai-whats-faster