I think it's important to distinguish the question of whether AI will reduce support for specific left wing perspectives that are often justified on the basis of 'its science' and whether it will reduce the influence of experts.
In general, I'd argue that expert influence is the ability of someone to counterfactualy move the debate based on what they say. The truth is that things like the wage gap *aren't* instances of academics forcing a view on the rest of society but of a left wing view being forced on academics and intellectuals. The academics may mostly not be struggling, they do tend to all be on the left, but this narrative isn't being dictated by the intellectual elite so much as the intellectual elite are being told by the rest of the left what is acceptable to say.
I hope AI means a reduction in that sort of thing but I'm not so confident. When I talk to people about it they either already tend to treat it as suspect (eg people who lean right) or happy to use motivated reasoning to reach that conclusion.
But if it does I think it will actually mean increased influence of intellectuals and experts because right now people are compensating for the fact that ideology shapes what the experts are allowed to say. The more you remove that, the more reason you have to actually change your view in response to an argument or experiment.
And I would add that AI is a powerful tool for knowledgeable experts who are dissidents from the expert consensus to build a powerful case against that consensus.
My guess is that we will see many “conventional expert wisdoms” come crashing down within the next decade. Not that those experts will acknowledge their incorrectness, but AI will make it much harder for them to persuade the next generation of their wisdom.
I agree completely. People have been debating on the internet the same silly gender wage imbalance issue for twenty years. Every single time it is like starting from scratch and bringing up the issues of men and women sorting around different preferences and values. Now you ask your AI and it succinctly grounds the issue. This is a major reduction in leftist misinformation.
I find that ChatGPT has become more balanced in latest models. It does indeed seem to be libertarian centrist (aka classical liberal). Its positions are in no way aligned with the woke progressive left or the MAGA right. I would say that it makes a mockery of both extremes (though in a respectful way). It does make me wonder what it actually says to a radical feminist socialist.
I agree that It is really important to push on the models to make sure they aren’t too obsequious.
I got it to give me its position on every important political topic — immigration, crime, education, welfare, etc. I can then push back and debate it. I can influence it and get it to change its position in some cases, in others it convinces or moderates me. One rare area we never agreed is the role of public service unions (I was strongly opposed, it less so).
For comparison, I tested Google's inbuilt AI on another gender-political issue where even informed people often misinterpret the data - the higher rate of male suicide. The AI response correctly observed that suicide *attempts* were actually more common among women, and that the reason that men nonetheless had higher suicide rates is because they chose more lethal methods, owing to higher male propensity for violence including violence against oneself.
Having spent the last month vibe-coding an app with ChatGPT, I notice that the quality of my interior dialogue has changed. Even when I’m working outside with no screens or other distractions, there’s still a constant cognitive back and forth that takes an analogous form to “Write a prompt to have Codex carry out a forensic audit…” and then waits for the result and carries out the necessary action. And this is just the beginning!
I've told Claude, Chat GPT and Grok in their permanent contexts not to feed my confirmation bias but instead to be a critical friend. They all stopped being sycophantic when I did this. I've also told them not to give monocausal explanations for complex phenomena and to give confidence assessments of their claims. They all follow these requests well. I would regulate AI companies to make this the default.
I find it very hard to five AI standing orders like that. It seems to work for awhile within one chat, but then it inevitably reverts back to default behavior.
In Claude, the AI I use the most, you put it in the preferences box in your profile. I am using the Sonnet 5 model. This is what I have there:
'I am a voracious reader of non-fiction, especially science, philosophy, cognitive science and some politics. I am an avid reader of Substack. I value meta and holistic perspectives and when discussing contentious matters I don't want AI to feed my cognitive biases but rather to act as a critical friend and help me understand different perspectives on a topic. For complex/contested political, social, or institutional questions: identify material contributing factors with a rough weighted sense of relative contribution (not sequential competing single-cause explanations), note where they reinforce or counterbalance each other, and converge on an actionable best-current-judgment — avoid open-ended hedging that never resolves to a priority. For any substantive claim or recommendation, explicitly flag your confidence level and the main reason for any uncertainty — distinguishing between 'I don't know', 'this is genuinely contested', and 'I'm extrapolating beyond my reliable knowledge'. '
BTW, as you may be able to tell, I had discussions with Claude to craft the above wording, discussing my need and having it draft the wording it thought would be most reliable in influencing it. I had those conversations with it at the end of chats when I thought it had not done what I needed, so I could give it a concrete example of what I thought it was doing unhelpfully and discuss what better would look like. Just like good feedback practice to a team member at work!
I also group conversations in Claude on a common theme into a Claude Project, which helps it maintain a common context across multiple related conversations.
I agree with 90%+ of the essay and think the thesis is sound, but have a couple of critiques:
1) You wrote: "AI companies compete for consumers, who choose the models they find most useful and trustworthy ... Users’ trust is therefore a key asset for AI companies. Any intentional deception or obfuscation uncovered by users, journalists or competitors could greatly damage a company’s reputation and drive many users away."
-I think the commercial success of Newsmax, OANN and charlatan influencers reveal that for *many* consumers trust is not earned by telling the truth. Trust is earned by confirming your priors, making you feel good about your side and providing a plausible excuse for why the other side is wrong. The huge AI companies will have a clear economic incentive towards the truth (just like mass market ABC, CBS, NBC) but the more niche AI models will have a market for earning trust through confirming priors (just like smaller market cable TV Fox News). Among typical consumers, trust is typically a socially derived quality, not an epistemiologically derived quality.
2) You wrote: "AI companies would not, for instance, draw obvious benefits from pushing false claims on major social and political issues, such as whether a vaccine is effective ... It is therefore reasonable to expect that competition between AI companies gives firms strong incentives to provide citizens with models that give the best answers they can, rather than answers designed to protect the authority of others."
-Smaller bespoke AI agents flooding social media pushing products have clear obvious benefits for pushing false claims. "Buy this product, it is better" even if it is not can be widely profitable for a company running an ad campaign using AI bots. In this case it's not the user going to a LLM looking for a wingman, it is the company's ad team looking for a wingman to deceive potential customers. The "user" is not always an individual human, it can be a company's marketing team.
3) A great deal of people simply do not have the baseline level of knowledge about a topic or the baseline ability to evaluate things from a "scientific" lens. They will not know what questions to ask, what topics to ask about and will be unaware of the gaps in their own knowledge that the LLM could fix. The average American adult reads at a middle school reading level - and that's the AVERAGE. There are tens of millions who are far below that.
--------
TLDR - Good essay, mostly agree. But I think more pessimism is warranted about the abilities of the median human to use tools effectively, the median human's desire to find the truth and the dangers of bespoke smaller AI models flooding the zone with propaganda to make money.
Hi Will, you make good points. Many media outlets like those you mention are far from impartial and often present a slanted view of reality. I may well be overoptimistic. However, my impression is that AI models, as products, are different.
First, I feel that their training puts them, by default, on much firmer epistemic ground than the average news outlet and, a fortiori, those with a political agenda.
Second, their interactive nature allows people to question them, and that makes deception and bias somewhat harder. A lot of the problem with the media you describe comes from the selection of facts and slanted presentation in a one-way format. When you can cross-examine the source, it cannot get away with this as easily. It has either to engage in more outright deception (which is riskier) or to back down.
So the interactive nature of AI models, besides any advantage their training may give them, makes biased presentation harder to sustain.
I second your mild skepticism about the average user. I would also add that future models will not have access to datasets /information that no one expected to be consumed by AI models. Everyone with money and machiavellian instincts is going to try to manipulate the inputs of future models, and will surely use AI to do so. The game is on, I would think, and this, I suspect, will wipe out part of the AI wingman effect. I can imagine an empirical test: are c span callers less conspiratorial today than 1-2 years ago? My impression is no-less conspiratorial, but they might be. Will callers be less hyperpartisan next year than they are now? An expert on AI models might be able to conduct these experiments with ease.
Nothing is entirely “clean”, and neither are human sources (media, experts). There are known limitations of AI models (and surely others yet to be discovered), but in my experience they are a much better source of knowledge, in terms of speed and accuracy, than many, if not most, alternatives.
“Better source of knowledge” , you mean DIRECTION of knowledge. They use knowledge as bread crumbs to point you in a direction according to owners algorithm. There is a new book out called; “The Nerd Reich” that explains the “better source of knowledge.”
I think it's important to distinguish the question of whether AI will reduce support for specific left wing perspectives that are often justified on the basis of 'its science' and whether it will reduce the influence of experts.
In general, I'd argue that expert influence is the ability of someone to counterfactualy move the debate based on what they say. The truth is that things like the wage gap *aren't* instances of academics forcing a view on the rest of society but of a left wing view being forced on academics and intellectuals. The academics may mostly not be struggling, they do tend to all be on the left, but this narrative isn't being dictated by the intellectual elite so much as the intellectual elite are being told by the rest of the left what is acceptable to say.
I hope AI means a reduction in that sort of thing but I'm not so confident. When I talk to people about it they either already tend to treat it as suspect (eg people who lean right) or happy to use motivated reasoning to reach that conclusion.
But if it does I think it will actually mean increased influence of intellectuals and experts because right now people are compensating for the fact that ideology shapes what the experts are allowed to say. The more you remove that, the more reason you have to actually change your view in response to an argument or experiment.
I wonder if this could be a driver of animosity towards AI from highly ideological groups 🤔 it also reduces their market power
I agree with this article.
And I would add that AI is a powerful tool for knowledgeable experts who are dissidents from the expert consensus to build a powerful case against that consensus.
My guess is that we will see many “conventional expert wisdoms” come crashing down within the next decade. Not that those experts will acknowledge their incorrectness, but AI will make it much harder for them to persuade the next generation of their wisdom.
Indeed, I would expect some fields (in particular those which became overly political) to face some pushback.
I agree completely. People have been debating on the internet the same silly gender wage imbalance issue for twenty years. Every single time it is like starting from scratch and bringing up the issues of men and women sorting around different preferences and values. Now you ask your AI and it succinctly grounds the issue. This is a major reduction in leftist misinformation.
I find that ChatGPT has become more balanced in latest models. It does indeed seem to be libertarian centrist (aka classical liberal). Its positions are in no way aligned with the woke progressive left or the MAGA right. I would say that it makes a mockery of both extremes (though in a respectful way). It does make me wonder what it actually says to a radical feminist socialist.
I agree that It is really important to push on the models to make sure they aren’t too obsequious.
I got it to give me its position on every important political topic — immigration, crime, education, welfare, etc. I can then push back and debate it. I can influence it and get it to change its position in some cases, in others it convinces or moderates me. One rare area we never agreed is the role of public service unions (I was strongly opposed, it less so).
For comparison, I tested Google's inbuilt AI on another gender-political issue where even informed people often misinterpret the data - the higher rate of male suicide. The AI response correctly observed that suicide *attempts* were actually more common among women, and that the reason that men nonetheless had higher suicide rates is because they chose more lethal methods, owing to higher male propensity for violence including violence against oneself.
Having spent the last month vibe-coding an app with ChatGPT, I notice that the quality of my interior dialogue has changed. Even when I’m working outside with no screens or other distractions, there’s still a constant cognitive back and forth that takes an analogous form to “Write a prompt to have Codex carry out a forensic audit…” and then waits for the result and carries out the necessary action. And this is just the beginning!
I've told Claude, Chat GPT and Grok in their permanent contexts not to feed my confirmation bias but instead to be a critical friend. They all stopped being sycophantic when I did this. I've also told them not to give monocausal explanations for complex phenomena and to give confidence assessments of their claims. They all follow these requests well. I would regulate AI companies to make this the default.
I find it very hard to five AI standing orders like that. It seems to work for awhile within one chat, but then it inevitably reverts back to default behavior.
In Claude, the AI I use the most, you put it in the preferences box in your profile. I am using the Sonnet 5 model. This is what I have there:
'I am a voracious reader of non-fiction, especially science, philosophy, cognitive science and some politics. I am an avid reader of Substack. I value meta and holistic perspectives and when discussing contentious matters I don't want AI to feed my cognitive biases but rather to act as a critical friend and help me understand different perspectives on a topic. For complex/contested political, social, or institutional questions: identify material contributing factors with a rough weighted sense of relative contribution (not sequential competing single-cause explanations), note where they reinforce or counterbalance each other, and converge on an actionable best-current-judgment — avoid open-ended hedging that never resolves to a priority. For any substantive claim or recommendation, explicitly flag your confidence level and the main reason for any uncertainty — distinguishing between 'I don't know', 'this is genuinely contested', and 'I'm extrapolating beyond my reliable knowledge'. '
BTW, as you may be able to tell, I had discussions with Claude to craft the above wording, discussing my need and having it draft the wording it thought would be most reliable in influencing it. I had those conversations with it at the end of chats when I thought it had not done what I needed, so I could give it a concrete example of what I thought it was doing unhelpfully and discuss what better would look like. Just like good feedback practice to a team member at work!
I also group conversations in Claude on a common theme into a Claude Project, which helps it maintain a common context across multiple related conversations.
Here's a link to a recent conversation I had with it where you can see some of these principles in action. https://claude.ai/share/511a1116-daf1-464c-9e09-54cb148742f3
I agree with 90%+ of the essay and think the thesis is sound, but have a couple of critiques:
1) You wrote: "AI companies compete for consumers, who choose the models they find most useful and trustworthy ... Users’ trust is therefore a key asset for AI companies. Any intentional deception or obfuscation uncovered by users, journalists or competitors could greatly damage a company’s reputation and drive many users away."
-I think the commercial success of Newsmax, OANN and charlatan influencers reveal that for *many* consumers trust is not earned by telling the truth. Trust is earned by confirming your priors, making you feel good about your side and providing a plausible excuse for why the other side is wrong. The huge AI companies will have a clear economic incentive towards the truth (just like mass market ABC, CBS, NBC) but the more niche AI models will have a market for earning trust through confirming priors (just like smaller market cable TV Fox News). Among typical consumers, trust is typically a socially derived quality, not an epistemiologically derived quality.
2) You wrote: "AI companies would not, for instance, draw obvious benefits from pushing false claims on major social and political issues, such as whether a vaccine is effective ... It is therefore reasonable to expect that competition between AI companies gives firms strong incentives to provide citizens with models that give the best answers they can, rather than answers designed to protect the authority of others."
-Smaller bespoke AI agents flooding social media pushing products have clear obvious benefits for pushing false claims. "Buy this product, it is better" even if it is not can be widely profitable for a company running an ad campaign using AI bots. In this case it's not the user going to a LLM looking for a wingman, it is the company's ad team looking for a wingman to deceive potential customers. The "user" is not always an individual human, it can be a company's marketing team.
3) A great deal of people simply do not have the baseline level of knowledge about a topic or the baseline ability to evaluate things from a "scientific" lens. They will not know what questions to ask, what topics to ask about and will be unaware of the gaps in their own knowledge that the LLM could fix. The average American adult reads at a middle school reading level - and that's the AVERAGE. There are tens of millions who are far below that.
--------
TLDR - Good essay, mostly agree. But I think more pessimism is warranted about the abilities of the median human to use tools effectively, the median human's desire to find the truth and the dangers of bespoke smaller AI models flooding the zone with propaganda to make money.
Hi Will, you make good points. Many media outlets like those you mention are far from impartial and often present a slanted view of reality. I may well be overoptimistic. However, my impression is that AI models, as products, are different.
First, I feel that their training puts them, by default, on much firmer epistemic ground than the average news outlet and, a fortiori, those with a political agenda.
Second, their interactive nature allows people to question them, and that makes deception and bias somewhat harder. A lot of the problem with the media you describe comes from the selection of facts and slanted presentation in a one-way format. When you can cross-examine the source, it cannot get away with this as easily. It has either to engage in more outright deception (which is riskier) or to back down.
So the interactive nature of AI models, besides any advantage their training may give them, makes biased presentation harder to sustain.
I second your mild skepticism about the average user. I would also add that future models will not have access to datasets /information that no one expected to be consumed by AI models. Everyone with money and machiavellian instincts is going to try to manipulate the inputs of future models, and will surely use AI to do so. The game is on, I would think, and this, I suspect, will wipe out part of the AI wingman effect. I can imagine an empirical test: are c span callers less conspiratorial today than 1-2 years ago? My impression is no-less conspiratorial, but they might be. Will callers be less hyperpartisan next year than they are now? An expert on AI models might be able to conduct these experiments with ease.
My apologies for the placement of my comment. It was supposed to be a reply to Will.
You assume that Artificial Intelligence is epistemologically clean? LOL!
Nothing is entirely “clean”, and neither are human sources (media, experts). There are known limitations of AI models (and surely others yet to be discovered), but in my experience they are a much better source of knowledge, in terms of speed and accuracy, than many, if not most, alternatives.
“Better source of knowledge” , you mean DIRECTION of knowledge. They use knowledge as bread crumbs to point you in a direction according to owners algorithm. There is a new book out called; “The Nerd Reich” that explains the “better source of knowledge.”