To admit how behind times I tend to be, I only recently started to use the YouTube main page to find content.   Before then, I would see YouTube only when it is embedded in other articles.

The first thing I noticed is that the main pages are immediately populated with recommendations based on the most recent video I watched.   If the video came from a channel with a lot of content, there would be many suggestions from the same channel.  Otherwise, there would be multiple recommendations with the same content, not just the same content, but the same sentiment of the content.   To the extent there are any surprise recommendations, the recommendations seem to be based on some estimation of who I am, perhaps based on demographics or my watching history.

I find interesting that the algorithms rely on negative feedback.   The algorithm always takes the initiative to decide what I might be interested in, and I am only given the option to say I don’t like the video, I am not interested in the channel, or I am not interested in recommendations based on a recently viewed video.   This negative feedback is very tedious because it has to be repeated for every time it gets it wrong.

I assume that the algorithms may be learning, but I have not seen much evidence that it is learning my interest in what is unlike what I had recently seen.   There seems to be a built-in bias to assume that I’m interested in watching what is similar or something compatible with what I had seen before.   Secondarily, there seems to be a bias toward recommending based on the demographic the algorithm thinks I belong to.   I’m middle-aged, male, living in urban part of the DC area.

To be clear, I have no idea how the algorithms are designed.  I only know that there is some algorithm, and I’m trying to infer how it might be making recommendations.   My conclusion is that the recommended content is very likely to be consistent with what I had watched previously or what others in my demographic are watching.

Similar things occur in other social media sites such as Facebook, Twitter, LinkedIn, etc.   The recommended content is based on algorithms that reinforce what it thinks I am.   The algorithms assume I pretty much made up my mind on everything.   Afterall, I’m a middle-aged guy and settled down where I am.   Why would I want to explore beyond what I know works for me already?

I started to think about the election that I watched from the earlier primaries where I was very interested in the challengers to the front-runners.   I am among the many who were intrigued by the strength of the unlikely challengers to the two parties: Bernie Sanders and Donald Trump.   I have read many post analysis of why they continued to have strong results after at each point when I assumed each had reached the end.   I am no political analyst so I will grant the professional analysts are pretty close to getting it right.

On the other hand, I can’t escape the suspicion that then entire election process starting in mid 2015 was somehow not following a normal democratic processes.  I get the impression that the trajectory of the election process was being controlled externally.

I am not alone in having this suspicions,  as well illustrated by the post-election allegations of hacking or foreign influence.  I tend to agree that there was a strong non-democratic influence on the trend of the election.   However, I believe this influence was not deliberate in terms of having the intention of directing the election in the way it played out.

The influence was a change in the way that democracy works in the age of Web 2.0 social media.  This is not the democracy of my youth.  It is a completely different ball-game.   This is not a paradigm shift in terms of describing future elections as being similar to this election.  Instead is a more awesome change in saying that each future election will not have any resemblance to its predecessor.

Let me clarify with the example of the 1960 election that many attribute as the first television influenced election.   After that election, the rules of campaigning changed to emphasize the how the campaign looks on television.   This was a major change in how to run elections, but once the new rules were defined the rules worked for many election cycles.

It is attractive to assume that the new social media environment changed elections in a similar way so that from now on, there will be new set of rules that will work for several future election cycles.   I’m inclined to believe that the rules for winning elections have changed very significantly, but unlike the example of television, the rules will continue to change radically for each new election.

The word “rules” above means the set of strategies that will more often than not work in winning votes.  The rules for television was to be sure the candidate looks good on camera and is well rehearsed in speaking the desired talking points.  The candidate was not allowed to be seen as uncomfortable or to speak spontaneous in a way that could result in embarrassing statements.   Following these rules did not guarantee a win, but not following these rules would at a minimum trigger a flurry of damage control actions with press releases or spokesman defending the statement on talk shows.

Certainly, we can study the recent election to define the new set of rules that will work in the modern age of social media.   My bet is that putting those rules into action the next time will fail.   From now on, each election will be completely unlike the previous one.

Returning to my observations about YouTube, I projected my recent observations on what my news feeds might have looked like during the campaigning season itself.

To set some context, let me project back to how social media played out in the 2008 election of Obama, a similarly surprising election.   In 2008, social media played a big role.  I would not say it was decisive, but it was a major contributor to Obama’s popularity.  At that time, social media was still relatively new.  In particular, the algorithms for identifying interesting content placed the majority of weight on personal interests.   As a result, the algorithms behaved roughly consistently across the entire country.   People with similar interests in any part of the country would get the same recommended content.   As a result, the proportionality of votes for either side were consistent with the proportion of population holding those views in each area.   This resulted in the pleasant outcome of a clear win in both the popular and electoral votes.   Similarly, this helped the exaggerating the win by consistently exciting the enthusiasm of one side across the country while simultaneously depressing the enthusiasm of the other side.

Eight years later, the algorithms have changed to be consider more demographics in order to optimize advertising revenue.   There is more emphasis to offer content of interests of the local community in addition to the specific person.   This permits better targeting of ads to match demographics.  Also, this will allow advertisers to focus their ads to match the sentiment the local population had on certain issues.

I am guessing this change in algorithm had a significant impact in the election because it fragmented like-minded people into demographics defined by age, sex, and location.   Similarly minded people in different parts of the country were getting different articles (in the form of news articles, videos, tweets, etc).   Different parts of the country were presented with substantially different content concerning the issues and the candidates, because those recommendations were highly influenced by their demographic cohorts.

This is only a conjecture, but if this is true, it is easy to see how this would have a major impact on the electoral college.

Consider a hypothetical county in western Pennsylvania where many people were interested in positive stories about Trump and negative stories about Clinton.  When some undecided or democratic voter would view some content, the subsequently recommended content may include content that would appeal to others in the same county.   He will end up seeing more negative content about Clinton and more positive content about Trump.  This could easily influence his vote or his choice to vote.

Next consider a similarly minded person in some similarly rural area of California.   Despite have the same general interests as his peer in Pennsylvania, his recommended content will be more consistent with his neighbor’s views.   He will see more recommendations of positive accounts about Clinton and more negative accounts of Trump.

In contrast to 2008 where social media exaggerated the voter turnout for the competing candidates nationwide, 2016 social media exaggerated the voter turnout in localities.  This led to the opposite results of the popular and electoral college counts.

If this theory is true, it may also have exaggerated the animosity within the country where different parts of the country appear to be completely in opposition to each other.   How can California be such a big win for Clinton while states like Pennsylvania can be such a big upset in favor of Trump, unless these states are populated with people with opposite views of what is good or bad for the country?   The truth is that there is more consensus in the nation than it appears, but the lopsided results gives the impression that we are close to the divisions we had leading up to the Civil War.

Going back to my earlier point about what changed in social media was the algorithms that matured to better optimize advertising effectiveness.   I take for granted that advertising is more effective when targeted to each individual’s demographic to include age and location cohorts as well the individual’s interests.   This results in certain types of products being more popular in some parts of the country and less popular in others.   The advertising intentionally exploits and amplifies local differences to optimize sales.

When the election came up, it tapped in to this divide-and-conquer bias of the algorithms that recommend content to social media users.   The commercially beneficial exaggeration of local differences became applied to political sentiments.   For a country as large as this country, there is a large diversity in commercially-relevant demographics.  On the other hand, the miracle of this country is that there is a general consensus about the nature and direction of government with both major parties offering modest changes in either direction at least in terms of what is practical to implement.

The danger exposed in the recent election is that the social media recommendation algorithms will apply commercially effective divide-and-conquer strategies to politics with the inevitable consequence of dividing and conquering.

For those who were surprised by this recent election, be prepared to be even more surprised by the next one.

Addendum: For a better researched discussion coming to a different conclusion, see Film Theory: Is Fake News KILLING the Internet? and its companion Game Theory: The REAL Reason YouTube is Broken.