January 11, 2021
Estimated read time: 9 minutes
1. As the Internet Rises, Trust Falls
2. Mining Gold
3. AI: A Scalable Solution
4. The Limits of Rationality
As the Internet Rises, Trust Falls
142 minutes per day. As of 2019, that’s how much time people spend social media. And this is where 46% of the USA turns to for their news.
This on its own isn’t necessarily any cause for concern. But given the full context, it very much should be. Trust in mass media, especially among Republicans has been decreasing over the past two decades. But the fall has not been steady, rather it happened in three distinct drops.
In 2004, the year of Bush’s re-election, 13% of Republicans lost faith in mass media in correspondence with the rise of internet usage for news, which rose to 1 in 5 Americans.
In 2012, the year of Obama’s re-election, a 12% drop in Republican faith occurred, the same year as the Facebook IPO.
In 2016, the year of Trump’s election, Republican faith in mass media dropped by 18% down to only 14% of all Republicans trusting mass media.
The data strongly seems to suggest that presidential elections + internet = fall of trust in mass media among Republicans. Meanwhile, Democrat trust is the second highest it’s been in the past two decades.
The result? The chaos of recent events.
Our two main parties have been divided to the point that we cannot agree on a common reality, on a common sense of truth, or on a common source of reliable information.
I would argue that the only way to reverse this trend is to start working towards any semblance of common ground. Clearly, this won’t be able to be achieved through our current news media or our current social media since those mediums led us to where we are today. Instead, I would argue we need to do work to make original sources of information as accessible as possible.
Mining Gold
Original sources of information are sources like laws passed or judicial rulings, where you could argue the implications or interpretations of a law but the fact that "This law passed which says that" is incontrovertible. Sources like this are gold veins for staying informed. But like gold veins, they either take a significant time investment or outside assistance to mine.
Consider the fact that the average reading speed is 236 words per minute. If you read an entire day with no breaks that comes out to 340,000 words in 24 hours. Meanwhile, the spending bill passed by Congress in December is 1.15 million words. This flood of information means not all of it can be consumed, some of it has to be refined down into something more digestible. But this is where bias inevitably enters.
There are three key ways in which bias inevitably seeps in when reporting on original sources of information.
The sources are completely passed through a filter, so the only output is opinion on the facts.
Not all sources are shared, but those that are shared are presented in their entirety. This selective inclusion of sources ends up painting a narrative that might clash with the missing sources.
All sources are shared, but are condensed. The act of shortening the sources necessarily removes information and ends up painting a narrative that might clash with the missing information.
Since our primary goal is finding sources of information to provide common ground between different ideologies, we should aim to eliminate the first form of bias whenever possible. However, to summarize and account for time, we need to permit either the second or third form of bias. Between the two, I would argue that in either case you end up not knowing what you don’t know, but at least with the third you get snippets of every source.
What does this look like in practice? GovTrack.
GovTrack is an independent team of four people who are dedicated to making Congress accessible and transparent. For every bill in Congress they aggregate data from the Library of Congress to share the current status of the bill in addition to the original text, a summary, and keywords. And that summarization is efficient. In this bill for example, the keywords take 7 seconds to read, the summary takes 13 seconds, while the original text takes 3 minutes.
In my opinion, GovTrack’s work is a near ideal solution. A person can see in seconds that a bill has been created in Congress, what topics are covered, and what the general idea is. Then, if they realize those ideas are relevant to them, they can choose to invest the time to fully consume the original source. This strategy seems like the most effective way of scaling up staying informed while minimizing bias. Get a high level overview of everything happening and a deep understanding of that which you care about most.
Unfortunately, there are limits to GovTrack’s approach.
Their source, the Library of Congress, does not always provide summaries of bills or their keywords. And the Library of Congress only covers Congress. Creating a new organization for every original news source to provide this kind of service isn’t the most tenable. If we want more of our news to be delivered in this manner, we need to create a scalable solution.
AI: A Scalable Solution
For those of you who don’t spend a lot of time in the tech space, AI that understands the English language has advanced. Significantly. Using the state of the art of AI, Mario Klingemann was able to create a six page story, with his only input being a fake name and the word "It". The AI generated the rest. Its English may not be perfect, but it is coherent.
The capabilities of modern AI that are most relevant for replicating GovTrack are:
AI can summarize large bodies of text
AI can extract keywords from large bodies of text
AI can provide the same service as the Library of Congress but can be pointed at any original news source. And it can do so much faster. To put it to the test, I tried applying the tech to executive orders.
I chose executive orders because they are another original data source that are freely available like Congressional bills. Additionally, because they represent actions that a President can take without first seeking approval of Congress, they carry a significant weight. But unlike bills, no similar service exists to provide a high level overview of executive orders.
Here’s an example of the AI at work.
Title: Expanding Educational Opportunity Through School Choice
Original Source (1175 words: 9m 59s): link
Keywords (5 words: 1 s):
recent, educational, states, school, students
Summary (400 words: 1 m 42 s):
As part of their efforts to address the public health challenges and uncertainties posed by the COVID-19 pandemic, State and local officials shut down in-person learning for the vast majority of our more than 56 million elementary and secondary school students beginning in late February and early March of this year.
The American Academy of Pediatrics (AAP) has also found that school closures have a “substantial impact on food security and physical activity for children and families.” Additionally, a recent survey of educators found student absences from school, including virtual learning, have nearly doubled during the pandemic, and as AAP has noted, chronic absenteeism is associated with alcohol and drug use, teenage pregnancy, juvenile delinquency, and suicide attempts.
Schools provide not only academic supports for students with special needs, but they also provide much-needed in-person therapies and services, including physical and occupational therapies.
Moreover, the survey found that virtual learning may not be fully accessible to these students, as children with special needs are twice as likely to receive little or no remote learning and to be dissatisfied with the remote learning received.
A recent analysis projected that, if in-person classes do not fully resume until January 2021, Hispanic, Black, and low-income students will lose 9.2, 10.3, and 12.4 months of learning, respectively.Start Printed Page 220
A failure to quickly resume in-person learning options is likely to have long-term economic effects on children and their families.
While some families, especially those with financial means, have been able to mitigate school disruptions through in-person options such as homeschooling, private schools, charter schools, and innovative models like microschools and “learning pods,” for many families, their children's residentially assigned public school remains their only financially available option.
These children, including those with special needs, are being underserved due to the public education system's failure to provide in-person learning options.
I am committed to ensuring that all children of our great Nation have access to the educational resources they need to obtain a high-quality education and to improving students' safety and well-being, including by empowering families with emergency learning scholarships.
The Secretary of Health and Human Services shall take steps, consistent with law, to allow funds available through the Community Services Block Grant program to be used by grantees and eligible entities to provide emergency learning scholarships to disadvantaged families for use by any child without access to in-person learning.
The AI shrunk the original text by 66% for the summary and 99.6% for the keywords. Moreover, if you read the original, you can see that not only is the summary fairly well written English, but it’s also fairly accurate. And crucially, everything you saw above was created by the AI algorithm in less time than it took you to read the title.
It seems like a perfect solution to scaling the approach of GovTrack. However, there are risks.
The accuracy and trustworthiness of an AI model is largely driven by the data it’s trained with. If it’s trained on extremely partisan data, then the summary or keywords generated could end up partisan.
Even if the summary isn’t biased, it may be flat out wrong. One alternate attempt at generating the summary in the example above resulted in the entire order being attributed to Barack Obama instead of Donald Trump.
Therefore, if we want to utilize AI as our scalable solution, we’d need to add additional feedback mechanisms to keep its potential shortcomings in check.
The Limits of Rationality
I believe that this article has, up until this point, constructed a rational argument for how AI can be utilized to help build a common source of truth. And that common source of truth is something that can be rationally comprehended by people of any party.
But rationality is not why most people engage with social media.
Social media engagement feels good. It releases dopamine specifically in a manner that triggers addictive behavior with the result being that we get addicted to social media. And an addiction mechanism in an environment where you can post anything and filter your consumption however you’d like inevitably leads to the vicious cycle of only consuming and sharing ideas your tribe agrees with.
Not every social media platform has such an environment though. tbh was an app that didn’t let users type anything of their own. Instead, it presented questions to users like "Who has the best smile?" and then presented the names of four friends. When you choose a friend, they receive an anonymous message that says "somebody thinks you have the best smile." This app still tapped into the same addictive triggers as Facebook or Twitter, but only enabled users to positively interact with each other. Then Facebook acquired tbh and it was shut down less than a year later.
I bring this up because as anyone who’s worked with salespeople before may know, people tend to act based on emotion and later justify their actions through logic. So I do think that AI can be used to solve the logical side of creating common ground. And for those who can be persuaded on logic alone, AI is sufficient. But to truly solve the problem at scale, either an addictive social media app needs to be created that fosters connection over division or we need to break America free from it’s addiction. One of those seems more likely to succeed to me than the other.
If you found this article to be meaningful, share it with friends and family whom you think would also find it to be meaningful.
Jordan, you continue to astound me with your knowledge and your <apparent> ease in sharing it. That means, Wowza! Thank you. Diana Pichierri