One innovation on social media that I perceive as having received a reasonable amount of praise from diverse constituencies is "Community Notes" on X (Twitter at the time of initial implementation). The basic idea is to allow notes to accompany a tweet to add additional context or present a critical or contrary viewpoint. Such a process would seem to rely on a situation where the correlation between views of different users on the platform isn't uniformly high. If all users have highly correlated views it will be hard to find divergent viewpoints that would potentially be useful to surface as a note. This is the power of low correlations. When you have access to sources of information with low correlation, you can recover from errors in one source by relying on sources that aren't strongly correlated. Adding correlated sources of information doesn't help as much because when one source is wrong the others are likely to be as well. It may be tempting to always want to rely on only the highest quality sources of information, whatever one considers those to be (peer reviewed studies, reputable new outlets, superforecasters etc.). The issue with looking solely at source quality is that when such a source is wrong, if you have heavily restricted sources that are open to consideration due to quality concerns then you may never be able to correct errors because all allowable sources are highly correlated.
Why AI may increase correlations
One idea that has been proposed that I find appealing is that of "AI for epistemics". The basic idea, as I understand it, would be to deploy AI systems to assist humans with understanding what is true about the world, similar to how the community notes algorithm hopefully surfaces notes that help people to figure out what is true. You'd have AI systems in the background doing things like doing research and evaluating evidence and then surfacing the results of this to human users.
I think this seems very interesting and promising, but one aspect of it that worries me is that this would have a general effect of increasing correlations across the board in many domains, short-circuiting the benefits that I see in lower correlations and making the world in general less robust.
Why would AI systems used for this purpose have a general tendency to increase correlations? I see two reasons:
The increased scalability of AI may result in increased centralization, where consumers look to a smaller number of information providers as their go-to sources. Information coming from a smaller number of sources may tend to be more correlated.
Developers of AI tools for epistemics will likely use a small number of advanced AI models that use relatively similar training data and procedures as part of their products. This small pool of models may tend to have a smaller diversity of outputs compared to the comparatively large number of humans involved in content and information generation as it functions in the present. If information production and evaluation begins to increasingly shift towards these AI models, the resulting end product that gets surfaced to users may be more correlated even if the media and informational institutions under whose banner the information is produced remain the same.
If this effect plays out in practice, I think the increased correlation would be a potential downside of using AI tools for this purpose.
The power of low correlations
One innovation on social media that I perceive as having received a reasonable amount of praise from diverse constituencies is "Community Notes" on X (Twitter at the time of initial implementation). The basic idea is to allow notes to accompany a tweet to add additional context or present a critical or contrary viewpoint. Such a process would seem to rely on a situation where the correlation between views of different users on the platform isn't uniformly high. If all users have highly correlated views it will be hard to find divergent viewpoints that would potentially be useful to surface as a note. This is the power of low correlations. When you have access to sources of information with low correlation, you can recover from errors in one source by relying on sources that aren't strongly correlated. Adding correlated sources of information doesn't help as much because when one source is wrong the others are likely to be as well. It may be tempting to always want to rely on only the highest quality sources of information, whatever one considers those to be (peer reviewed studies, reputable new outlets, superforecasters etc.). The issue with looking solely at source quality is that when such a source is wrong, if you have heavily restricted sources that are open to consideration due to quality concerns then you may never be able to correct errors because all allowable sources are highly correlated.
Why AI may increase correlations
One idea that has been proposed that I find appealing is that of "AI for epistemics". The basic idea, as I understand it, would be to deploy AI systems to assist humans with understanding what is true about the world, similar to how the community notes algorithm hopefully surfaces notes that help people to figure out what is true. You'd have AI systems in the background doing things like doing research and evaluating evidence and then surfacing the results of this to human users.
I think this seems very interesting and promising, but one aspect of it that worries me is that this would have a general effect of increasing correlations across the board in many domains, short-circuiting the benefits that I see in lower correlations and making the world in general less robust.
Why would AI systems used for this purpose have a general tendency to increase correlations? I see two reasons:
The increased scalability of AI may result in increased centralization, where consumers look to a smaller number of information providers as their go-to sources. Information coming from a smaller number of sources may tend to be more correlated.
Developers of AI tools for epistemics will likely use a small number of advanced AI models that use relatively similar training data and procedures as part of their products. This small pool of models may tend to have a smaller diversity of outputs compared to the comparatively large number of humans involved in content and information generation as it functions in the present. If information production and evaluation begins to increasingly shift towards these AI models, the resulting end product that gets surfaced to users may be more correlated even if the media and informational institutions under whose banner the information is produced remain the same.
If this effect plays out in practice, I think the increased correlation would be a potential downside of using AI tools for this purpose.