I’m glad you’re engaging with this (and I had a back and forth with Megan on X). I think the more important issue is revising the prose, which is what I spend time on my Substack. My recent piece on paragraphs (!) has over 100 likes, which is crazy for a piece on human v. LLM paragraphs! My piece last month on “not x but y” has over 300 likes. There is clearly a need for engagement with craft and how to tell the difference.
"Lots of working musicians lost their livelihoods to jukeboxes and radios replacing them in hotels, bars, and restaurants."
Worldwide there are more working musicians now than there were then, per capita.
Every single time that technological change has been predicted to be a net job killer, that change has proven to be a net job creator. Every single time. And the only reason people have to believe that now is different is because they live now and they need to believe that they live in a unique, special time. But there's nothing special about them and there's nothing special about now.
Yes... the point of the piece is even when industries end up "better off" than they were before, lots of real people are impacted in the short to medium run.
The transition costs aren't frivolous just because the end state passes a benefit cost analysis. The transition from agricultural to industrial employment took more than a century and involved enormous suffering even though it was obviously worth it.
It's good to ground ourselves in that even if we're optimistic about the new equilibrium we're headed towards. In part because it's callous to tell someone whose entire life is about to be upended that it'll be worth the cost and in part because there are policy levers we can use to smooth transitions for people who were caught unawares.
An apocryphal story I heard years ago hit me in a way the old Luddite debate never quite did. A western economist was being given a Potemkin Village-style tour of rural China during the Great Leap Forward. When his minder brought him to a site where a dam was being constructed, he observed the work for a while, then asked, “why are the workers using picks and shovels? I know you have earth-moving equipment.” “We do, of course,” he was told. “But the peasants need jobs.” The economist slapped his forehead and exclaimed, “Oh! I’m sorry! I misunderstood completely — I thought you wanted a dam built. If you wanted a jobs program, why didn’t you give the workers spoons?”
“Every single time” is surely false, if you’re really talking about jobs in an industry growing as the industry adds automation. Agriculture is the big clear example.
Even with musicians, it may be that global employment in music has grown per capita, but I suspect a lot of that has to do with emergence from poverty. In the United States, musicians are only about half the per capita employment of what the were a century and a half ago (though the absolute number has increased).
Not every good has enough elasticity for consumption to grow faster than productivity.
I’m a trade press reporter and I find AI a valuable tool for transcribing, summarizing, finding background information, analyzing data, and putting dense jargon in layman-friendly language. Of course you have to fact check, but that’s not so different from working with human sources who might have grudges, biases, or memory issues of their own.
Anti-AI Puritanism at this point just reads as willful ignorance. Something doesn’t need to be perfect to be useful.
I think there's two elements to the "jobs" impact.
One is that technology that improves productivity can drive cost down, therefore increasing demand, and thus preserving or actually increasing the number of jobs in an industry. I believe more people (globally) are employed in automobile manufacturing today than were in 1925 or 1955 by a substantial margin. Cars per worker is much greater today but total cars sold is much, much greater. I used this analogy with my undergraduates who are worried about computer coding jobs. I'm not sure that it will work out the same as for cars, but I can easily imagine a positive future scenario where people with modest coding ability and Claude Code can cheaply solve many problems that, currently, get no attention. For example, municipalities or small non-profits could actually have well organized IT systems, integrated with their websites etc.
The other is what we might call the "fixed audience" problem. This is something that will hurt journalism specifically. It is sort of a demand side analog of to cost-disease. Much as primary school or nursing is always going to require a large number of people per customer, _there is only so much news_ that ordinary citizens can, or should, _consume in one day_. It follows that if AI can do the jobs of reporters, or assist some of them in being more productive, _there won't be space for the rest of them to be read by the public_.
But aren't we kind of already there? Who has time to read _more news_?
Here's my positive scenario for journalism. It starts from my experience of using AI and how it is different from experience of using social media.
The first time I saw Twitter was in 2006. I was in grad school and an undergrad showed it to me after a lab meeting. I said "wait, so you any idiot can just force their garbage ideas into your feed, and you _want_ this?" Back then it was people tweeting pictures of their lunch. I started studying social media a few years later and have never had this original view overturned. Too many people talking too much sharing low quality information to try to get attention. Did trust in expertise and journalism decline only because of this? No. But there is some truth to McLuhan's intuition that the medium is the message. The medium of social media is "scream out trash" and our public sphere is now a littered park.
The first time I used AI, I had a very, very different experience. I had just seen my colleague David Rand present his research (with Tom Costello and Gordon Pennycook) on how AI could convince conspiracy theorists to believe less in a conspiracy. I was intrigued by the result, but what really impressed me was the dialogues he showed. The AI was so patient and rigorous in its explanations. It would answer any question, no matter how absurd, with clear, intelligent prose. Always calm, always rational.
Since that time this has been my experience of in my own conversations with AI. It is _really smart_. It says really smart things to me, on any topic I ask about. Is it sometimes wrong? Yes. Aren't people sometimes wrong? And unlike people, it _admits_ it when you catch it. Or, in other words, the medium of AI content is "try to be precise based on all knowledge in the world." It's hard for me to imagine replacing the social media ethos with this one will be bad for the public or for the profession of journalism.
I do expect something to change. I expect that, in the future, there will be _fewer_ opinion leaders/talking heads at the top. Great writers with sophisticated knowledge will be better with the help of AI, and AI alone will not be able to replace them. And they will be trusted by their audiences. Meanwhile, AI can write the mediocre copy that now runs on mediocre news sites. But this doesn't mean there won't be work for journalists. Journalists will become more like research analysts -- interrogating the world to find facts to feed into the AI system. In fact, because the AI will be omni-curious, and highly productive, I expect more jobs here. More jobs reporting in local news, investigative reporting and so forth.
Those jobs will pay well and be respected for the knowledge they provide. But the one thing they won't do is give those reporters a human audience. The AI will be their audience, and then the audience for the AI will be the big shot stars. Would that be so terrible? For journalists to do a good job and just not be famous? And wasn't that what it was like for decades, when Americans hardly knew the names of any news reporters at all?
I'm curious about the potential for models that are trained to not *quite* align with the user's inclinations. Challenging conspiracy theories is an application I hadn't thought of before but matches; education models that make students work for the answer without just giving it away is the biggest application I can imagine. In both those cases, you would need a user who is at least partially aligned to the goals of the model trainer: you need not just any conspiracy theorist, but one who prides themselves on their reasoning and is open to having their beliefs challenged. For an education model, you need not just any lazy teenager, but one who accepts that they need to learn the material, even if their first inclination would be to finish the work as quickly as possible.
Yeah, specific goals would need to be designed for, and would work to a varying degree for sure.
But beyond this, I just think there will be general benefits to people interacting with these machines, in comparison to social media, just in the _form_ of their speech. Full sentences, humble, willing to correct errors. I think we will all benefit just from interacting more this way.
They're not a substitute for human interaction. And there are some very scary tail risks. But it's hard for me to imagine a worse public sphere than we have now.
AI is helping me oversee very large quantities of information that I’m writing about. A lot of the projects I’ve had lately have involved thousands of pages of documents. It makes it much easier to scan them for the right source material. I use it stress test my drafts to see if the arguments are strong given my source material. I don’t use it to replace thinking but to augment the writing process. It makes me much more efficient.
I think this is a great engagement. Progress and legitimate grievance aren't exclusive. I think the anger at McArdle is about what comes next as much as what she describes doing.
Her workflow still has research legwork, pushback on her thesis. It's not "hey AI make this thing" slop creation, but engaged and iterative usage. Feels like a lot of the critics are having a kneejerk reaction to any kind of AI use. I wrote about those framings a few months back - https://www.robin-cannon.com/p/did-ai-write-this-post. We should focus more on the question of how AI is involved, not just whether it was at all. AI doesn't preclude writers' exercising judgement.
I love being a writer more since AI. I'm more productive, creatively and professionally, than I've ever been. I find it a useful tool to help me catalogue my thoughts, act as a sounding board when a real person isn't available. But I'm also not economically dependent on my writing, so I'm looking at it from a different perspective.
You replied to an earlier comment that just because the net benefit to society of a change is positive, that doesn't mean it doesn't impact some people negatively. That's exactly right. This is creative destruction we're talking about. Who bears the costs and who reaps the benefit?
AI is great for writing FOIA requests. It can scan what you then receive and write follow-up requests that help you zero in on what you’re investigating.
"musicians were wrong to worry about live performances; live music revenue has exploded, even as recorded music revenue remained modest."
And yet, has this benefitted all musicians or primarily those at the top? There is research (Koenig 2023 at AER Insights) on the example you cite. Koenig finds that the introduction of television exacerbated inequality. It expanded the market for the most popular entertainers and boosted their incomes, but it also crowded out the less talented or lucky ones.
That could be the case today as well. The Eras tour was a multibillion dollar event, but that revenue doesn't support live musicians that aren't Taylor Swift.
I’m glad you’re engaging with this (and I had a back and forth with Megan on X). I think the more important issue is revising the prose, which is what I spend time on my Substack. My recent piece on paragraphs (!) has over 100 likes, which is crazy for a piece on human v. LLM paragraphs! My piece last month on “not x but y” has over 300 likes. There is clearly a need for engagement with craft and how to tell the difference.
Here's the link if anyone wants to know how to paragraph. https://hollisrobbinsanecdotal.substack.com/p/for-the-love-of-god-learn-to-paragraph
"Lots of working musicians lost their livelihoods to jukeboxes and radios replacing them in hotels, bars, and restaurants."
Worldwide there are more working musicians now than there were then, per capita.
Every single time that technological change has been predicted to be a net job killer, that change has proven to be a net job creator. Every single time. And the only reason people have to believe that now is different is because they live now and they need to believe that they live in a unique, special time. But there's nothing special about them and there's nothing special about now.
Yes... the point of the piece is even when industries end up "better off" than they were before, lots of real people are impacted in the short to medium run.
The transition costs aren't frivolous just because the end state passes a benefit cost analysis. The transition from agricultural to industrial employment took more than a century and involved enormous suffering even though it was obviously worth it.
It's good to ground ourselves in that even if we're optimistic about the new equilibrium we're headed towards. In part because it's callous to tell someone whose entire life is about to be upended that it'll be worth the cost and in part because there are policy levers we can use to smooth transitions for people who were caught unawares.
An apocryphal story I heard years ago hit me in a way the old Luddite debate never quite did. A western economist was being given a Potemkin Village-style tour of rural China during the Great Leap Forward. When his minder brought him to a site where a dam was being constructed, he observed the work for a while, then asked, “why are the workers using picks and shovels? I know you have earth-moving equipment.” “We do, of course,” he was told. “But the peasants need jobs.” The economist slapped his forehead and exclaimed, “Oh! I’m sorry! I misunderstood completely — I thought you wanted a dam built. If you wanted a jobs program, why didn’t you give the workers spoons?”
“Every single time” is surely false, if you’re really talking about jobs in an industry growing as the industry adds automation. Agriculture is the big clear example.
Even with musicians, it may be that global employment in music has grown per capita, but I suspect a lot of that has to do with emergence from poverty. In the United States, musicians are only about half the per capita employment of what the were a century and a half ago (though the absolute number has increased).
Not every good has enough elasticity for consumption to grow faster than productivity.
I’m a trade press reporter and I find AI a valuable tool for transcribing, summarizing, finding background information, analyzing data, and putting dense jargon in layman-friendly language. Of course you have to fact check, but that’s not so different from working with human sources who might have grudges, biases, or memory issues of their own.
Anti-AI Puritanism at this point just reads as willful ignorance. Something doesn’t need to be perfect to be useful.
I think there's two elements to the "jobs" impact.
One is that technology that improves productivity can drive cost down, therefore increasing demand, and thus preserving or actually increasing the number of jobs in an industry. I believe more people (globally) are employed in automobile manufacturing today than were in 1925 or 1955 by a substantial margin. Cars per worker is much greater today but total cars sold is much, much greater. I used this analogy with my undergraduates who are worried about computer coding jobs. I'm not sure that it will work out the same as for cars, but I can easily imagine a positive future scenario where people with modest coding ability and Claude Code can cheaply solve many problems that, currently, get no attention. For example, municipalities or small non-profits could actually have well organized IT systems, integrated with their websites etc.
The other is what we might call the "fixed audience" problem. This is something that will hurt journalism specifically. It is sort of a demand side analog of to cost-disease. Much as primary school or nursing is always going to require a large number of people per customer, _there is only so much news_ that ordinary citizens can, or should, _consume in one day_. It follows that if AI can do the jobs of reporters, or assist some of them in being more productive, _there won't be space for the rest of them to be read by the public_.
But aren't we kind of already there? Who has time to read _more news_?
Here's my positive scenario for journalism. It starts from my experience of using AI and how it is different from experience of using social media.
The first time I saw Twitter was in 2006. I was in grad school and an undergrad showed it to me after a lab meeting. I said "wait, so you any idiot can just force their garbage ideas into your feed, and you _want_ this?" Back then it was people tweeting pictures of their lunch. I started studying social media a few years later and have never had this original view overturned. Too many people talking too much sharing low quality information to try to get attention. Did trust in expertise and journalism decline only because of this? No. But there is some truth to McLuhan's intuition that the medium is the message. The medium of social media is "scream out trash" and our public sphere is now a littered park.
The first time I used AI, I had a very, very different experience. I had just seen my colleague David Rand present his research (with Tom Costello and Gordon Pennycook) on how AI could convince conspiracy theorists to believe less in a conspiracy. I was intrigued by the result, but what really impressed me was the dialogues he showed. The AI was so patient and rigorous in its explanations. It would answer any question, no matter how absurd, with clear, intelligent prose. Always calm, always rational.
Since that time this has been my experience of in my own conversations with AI. It is _really smart_. It says really smart things to me, on any topic I ask about. Is it sometimes wrong? Yes. Aren't people sometimes wrong? And unlike people, it _admits_ it when you catch it. Or, in other words, the medium of AI content is "try to be precise based on all knowledge in the world." It's hard for me to imagine replacing the social media ethos with this one will be bad for the public or for the profession of journalism.
I do expect something to change. I expect that, in the future, there will be _fewer_ opinion leaders/talking heads at the top. Great writers with sophisticated knowledge will be better with the help of AI, and AI alone will not be able to replace them. And they will be trusted by their audiences. Meanwhile, AI can write the mediocre copy that now runs on mediocre news sites. But this doesn't mean there won't be work for journalists. Journalists will become more like research analysts -- interrogating the world to find facts to feed into the AI system. In fact, because the AI will be omni-curious, and highly productive, I expect more jobs here. More jobs reporting in local news, investigative reporting and so forth.
Those jobs will pay well and be respected for the knowledge they provide. But the one thing they won't do is give those reporters a human audience. The AI will be their audience, and then the audience for the AI will be the big shot stars. Would that be so terrible? For journalists to do a good job and just not be famous? And wasn't that what it was like for decades, when Americans hardly knew the names of any news reporters at all?
I'm curious about the potential for models that are trained to not *quite* align with the user's inclinations. Challenging conspiracy theories is an application I hadn't thought of before but matches; education models that make students work for the answer without just giving it away is the biggest application I can imagine. In both those cases, you would need a user who is at least partially aligned to the goals of the model trainer: you need not just any conspiracy theorist, but one who prides themselves on their reasoning and is open to having their beliefs challenged. For an education model, you need not just any lazy teenager, but one who accepts that they need to learn the material, even if their first inclination would be to finish the work as quickly as possible.
I have no idea if this is even feasible, however.
Yeah, specific goals would need to be designed for, and would work to a varying degree for sure.
But beyond this, I just think there will be general benefits to people interacting with these machines, in comparison to social media, just in the _form_ of their speech. Full sentences, humble, willing to correct errors. I think we will all benefit just from interacting more this way.
They're not a substitute for human interaction. And there are some very scary tail risks. But it's hard for me to imagine a worse public sphere than we have now.
The Bully Pulpit is really good! It made me have even more complicated feeling about TR
AI is helping me oversee very large quantities of information that I’m writing about. A lot of the projects I’ve had lately have involved thousands of pages of documents. It makes it much easier to scan them for the right source material. I use it stress test my drafts to see if the arguments are strong given my source material. I don’t use it to replace thinking but to augment the writing process. It makes me much more efficient.
I think this is a great engagement. Progress and legitimate grievance aren't exclusive. I think the anger at McArdle is about what comes next as much as what she describes doing.
Her workflow still has research legwork, pushback on her thesis. It's not "hey AI make this thing" slop creation, but engaged and iterative usage. Feels like a lot of the critics are having a kneejerk reaction to any kind of AI use. I wrote about those framings a few months back - https://www.robin-cannon.com/p/did-ai-write-this-post. We should focus more on the question of how AI is involved, not just whether it was at all. AI doesn't preclude writers' exercising judgement.
I love being a writer more since AI. I'm more productive, creatively and professionally, than I've ever been. I find it a useful tool to help me catalogue my thoughts, act as a sounding board when a real person isn't available. But I'm also not economically dependent on my writing, so I'm looking at it from a different perspective.
You replied to an earlier comment that just because the net benefit to society of a change is positive, that doesn't mean it doesn't impact some people negatively. That's exactly right. This is creative destruction we're talking about. Who bears the costs and who reaps the benefit?
AI is great for writing FOIA requests. It can scan what you then receive and write follow-up requests that help you zero in on what you’re investigating.
You write
"musicians were wrong to worry about live performances; live music revenue has exploded, even as recorded music revenue remained modest."
And yet, has this benefitted all musicians or primarily those at the top? There is research (Koenig 2023 at AER Insights) on the example you cite. Koenig finds that the introduction of television exacerbated inequality. It expanded the market for the most popular entertainers and boosted their incomes, but it also crowded out the less talented or lucky ones.
That could be the case today as well. The Eras tour was a multibillion dollar event, but that revenue doesn't support live musicians that aren't Taylor Swift.
https://pubs.aeaweb.org/doi/pdfplus/10.1257/aeri.20210539
Question is the weekly pod coming back?
We'll have an exciting announcement about this in just a few days!
Yay!! I was hooked on all the interesting weekly topics
In the interests of fairness, this thoughtful post from a right leaning data analyst argues that the "Trump screws over blue states" chart is an artifact of a cherry picked definition of blue states: https://www.marginallycompelling.com/p/fema-is-not-screwing-blue-states
Thanks! Thought we got them all...