Anatomy of a Mob: The Lacy/Zuckerberg Interview

"Backchannel" communication mechanisms like Twitter are going to revolutionize everything from meetings and classrooms, to our day-to-day relationships… but first we have to learn to deal with them appropriately. At SXSW, the Twitter backchannel turned an audience into a mob. The session leaders, not having access to the backchannel, had no idea what was happening, or why the mood of the audience had changed so rapidly.

This is an article about the keynote interview at the 2008 SXSW Interactive Festival. If you haven't heard the details, you can do a Google search on "SXSW Lacy Zuckerberg interview", or read one of many articles such as Wired’s SXSW: 2008, the Year the Audience Keynoted.

First things first.

I am not going to make any comments here about whether it was a good interview, bad interview, inappropriate topic, inappropriate people… that's not the point of this article, nor do I think it was really the reason why the audience at SXSW got so upset. Please keep any comments on topic, and away from the details of the interview. In particular, keep in mind that this was not the only session at SXSW that ended up with an audience verbally attacking the panelists and taking over the session.

At the SXSW Keynote interview, an audience of conference-goers acted unusually; they became a mob. I mean that in the sense (as Dave Winer does in his "Twitter is not a chatroom" podcast) of a crowd out of control—a crowd doing things that individuals would not. This is the kind of behavior generally seen at political events; or meetings where the audience is emotionally tied to the subject matter. An interview or panel that seems off-topic usually gets a few walk-outs and a bad rating in the "What can we do better next year?" survey; not an audience revolt. I find that very interesting, and I have some ideas as to why it may have occurred.

“Where are the yellow comment cards?” - 51 minutes

In the nineteen-sixties, psychologist Stanley Milgram ran a series of experiments which are probably familiar to anyone who has taken Psychology 101. He asked subjects to inflict pain on a person when that person made a mistake on a test. A number of interesting results came out of the experiment, but one variation in particular is relevant here. The willingness to administer shocks increased when the recipient was more remote. It's hard to de-humanize someone who is right next to you. On the other hand, if you can't see them, and don't know them, your emotional attachment is greatly reduced, and you will do (and say) things you would never do in person. This isn't news to anyone who has been involved in an email flame-fest, or posted something online that later got quoted face-to-face. We say things online that we would never say in real life.

“sarah lacey is in love with zuckerberg, me thinks”- 11 minutes
“someone offering me $20 to yell "Beacon Sucks". paying for my night of drinking if I do. I definitely should.”- 17 minutes
“Someone just yelled out, "BEACON SUCKS!" - yes, that's how this is going...u should be here :)”- 18 minutes
“Holy cow, whoever this is interviewing Mark is horrible! I think she is in love with him. Hold on...she might go in for the kiss!" - 20 minutes

Mobs form when individuals feel anonymous, and believe that their feelings and behaviors are shared by others. When the behavior becomes visible, and when nobody reacts negatively to it, the behavior gets amplified, with more and more people joining in. But why would this happen in a technical conference on a relatively unexciting keynote topic?

The blame, of course, has fallen on Twitter, because it was through Twitter messages that the revolt first began to appear. More and more, services like Twitter are being used as "backchannel" communications mechanisms by which the audience can provide non-interruptive feedback to each other, and to the people running the event. In this case, the people running the keynote obviously weren't tracking the backchannel (perhaps "sidechannel" would be more accurate in that instance). Rationally, there's no reason to get frustrated that your Twitter comments are being "ignored" by someone whom you know isn't reading them, but frustration probably did play a role here as well.

One thing that wasn't clear to me in the discussions about the keynote, was whether the Twitter phenomenon was "real", or whether it was just the result of a lot of posts by a few (prolific) individuals. So I decided to see if I could gather up a collection of all the Twitter messages sent from the keynote, during the keynote.

There's a detailed methodology section at the end, but this is the quick summary.

1. Searched for all users who mentioned "SXSW" between Thursday and Sunday the week of the conference.
2. Searched for all users who mentioned "facebook", "lacy", "zuck" or "keynote" within a few hours of the event.
3. Manually narrowed down the resulting 4000 messages.

The final result was nearly 2000 Twitter messages, sent by 500 users. I know that some of the messages I included were irrelevant, and I'm sure I missed others. But overall, that comes to an average of close to one message every two seconds! There is no question that backchannel communications were very active during the keynote.

It is my belief that what happened was a combination of standard mob behavior with the side-effects of the behavior observed by Milgram. People sent messages, using the usual level of rudeness that occurs in electronic communication. Those messages were read by people also at the keynote. As the level of vitriol rose in the twittersphere, it also fed back into the people in the crowd. "Surrounded" by other people feeling the same way, the animosity and feeling of anonymity moved from the virtual world into the real world; and a virtual mob turned into a real one.

Is that what really happened? It's impossible to say for sure. Twitter is not a real-time protocol, and people take time to type messages, so it was impossible to directly connect a particular message to a particular time during the talk. Also, there is no way to tell who in the audience saw the messages (or heard about them from a seat-mate who did).

What I have done, is gather up the Twitter messages from the keynote, and overlay them directly on a video of the interview. I used the video posted on Viddler by allfacebook. It cuts off somewhere near the end, but was the most complete version I could find.

The following video is annotated with all the Twitter messages I collected. In some cases you'll have to pause to read them, since they go by too quickly. I grouped together any posts that occurred within one second of each other to make it a little easier to scan. Since I saw apparent lags of up to fourteen minutes between what some tweets described and what happened, I finally settled on a five minute lag for all tweets. So keep in mind that the tweets do not directly correspond to the current scene in the video, and two adjacent tweets were not necessarily written, sent, or recorded at the same time.

Because this version of the video is hosted at Viddler, you can not only comment on my blog, you can also make comments in the video, attaching them to particular frames. Please feel free to do so.

The Annotated Video

Note: The annotations will be easier to read if you expand the video to full size (click in the upper right hand corner of the video).

The Annotated Data

The following charts show the relationships between messages ("tweets"), users, content and time. I have actually done a minimal amount of analysis. Most of my time was spent filtering the data and annotating the video. If someone wants to examine things some more, I'd be happy to turn over the data and my (very) motley collection of Perl scripts.

The charts are all plotted against time, usually expressed as HH:MM and sometimes just MM. All values are the sum of the activity over a minute, or in some cases, five minutes.

Number of Messages Sent and Number of Users

The blue area shows the total number of messages being sent each minute. The red area shows the total number of new users (i.e. those who haven't previously sent a message). I actually have not spent much time examining what happens in the video vs. what happens in the tweets. Does something interesting happen around 45 minutes to cause a spike, or is that a conincidence?

The final total was 512 users and 1857 messages.

First and Last Sending Times (With Count) Per User

The following two charts (attempt to) provide a view of how often and how long people were posting messages. I could have done a standard mean/median/mode chart, but I wanted to include the start and stop times. In both charts the vertical axis represents the time at which someone sent their first message, and the horizontal axis indicates when they sent their last message. (So the diagonal line represents people who only sent one message.) The first chart uses the size of the bubble to indicate how many messages the person sent. The second chart uses the color of the square.

The most messages sent was 65 (by three people). A quick check indicated that the frequent posts seemed to be from people providing a running report on what was happening.

Frequency of Top 50 Words over Time

You'll probably want to click on the image and view the PDF version.

This chart shows how often each word appeared during a given five minute period. The most common words appear on the right. For instance, we can see that the words "audience", "people" and "crowd" grow in popularity over time; probably indicating that people were tweeting about the behavior of the audience as the interview progressed.

Frequency of Top 50 Words by User over Time

This chart addresses the question of whether the messages were being sent by just a few people, or many. Here the value for each five minute period is the number of people who used a word. Where the previous chart counted all instances of a word, even if used multiple times by a single user; this counts only one use per user. The chart appears very similar to the previous one, indicating that the use of the top fifty words was quite wide-spread.

Top 50 Words, Total Frequency, and by User

Like the previous chart, but showing just the totals. The red bars indicate the number of people who used a given word. The blue bars indicate the total number of times the word was used. Again, the usage appear to be fairly evenly distributed.

Conclusion

The Twitter transcript makes it clear that there was an early and constant stream of negative comments flowing from a large number of senders. The lack of accurate timing information makes it impossible to tell for certain whether that was something that started small and spread, or exactly how it erupted into real life. However, it is clear that the conversations in Twitter did lead to the same level of real life behavior and dialog. While it could just be attributed to the general decline of societal mores, I believe my original assertion as to the connection and influence of the virtual and real worlds is potentially valid. It would be interesting to see a more detailed and rigorous study of future events. (Sounds like a good Sociology/Psychology/Anthropology thesis for someone. :-)

Twitter provides a communication channel which augments, rather than interrupts, existing communications. As such, it makes it possible for people to communicate both within a group, and (in structured events) to the leaders of a group, all without disrupting the normal progress of the activity. If that sounds like too much for a panel discussion or interviewer to manage, consider that most reporters in traditional media (not to mention football quarterbacks) have similar mechanisms for receiving information while they work. Whether increased multitasking is a good thing from a quality standpoint is a different issue. For that, look at the research that Howard Rheingold has been doing in the areas of multitasking and backchannels.

The Twitter backchannel can definitely have a positive influence. As a remote observer of SXSW I was not only able to receive ongoing summaries of sessions, but I could suggest questions for attendees to ask, and provide resources to panelists while they were in an active session. The backchannel can provide a low-key mechanism for alerting presenters to issues, offering support, and of course organizing and coordinating group actions. The issue, is how to keep group actions from growing out of control. I believe that requires education (or perhaps just a new generation of users) about the differences between virtual and real communication, and the dangers of transporting emotions directly from one to the other. I believe it also requires responsibility on the part of backchannel spectators.

There is a tendency in online discussions to let flames burn themselves out. After all, it's the virtual world, not the real one. "Getting involved" can be a pain. But as the SXSW events show, the boundaries between real and virtual get thinner every year, and virtual emotions can cause real-world harm. I greatly admire Sarah Lacy's ability to deal with the abuse she has gotten and move on. A reporter has to have a tough skin, but it still can't have been easy. She didn't deserve the abuse that was dished out on Twitter, let alone what happened in the auditorium.

As citizens of the online world, we have a responsibility to step forward when we see people misbehaving. It doesn't take much to tone things down. People need to be reminded that the target of their frustrations is a real person. They also need to be reminded that their persona, though virtual, has its own reputation to think about. The members of an online mob are in fact far less anonymous than those in a real mob. I was rather shocked when I happened to notice that one of the tweets I quoted above was actually made by someone I follow on Twitter. It was more sophomoric than mean, but it still contributed to the overall mood. Finding out who said what during the conference is a simple task for anyone with access to Google. We need to live our online lives under the assumption that everything we say, and everything we do, no matter how private it seems, is going to contribute to our overall reputation. That's a good thing, but it takes getting used to.

“So it took an eight-year-old child to bring 'em to their senses.... That proves something - that a gang of wild animals can be stopped, simply because they're still human. Hmp, maybe we need a police force of children.” –– Jean Louise (Scout) Finch in "To Kill a Mockingbird"

Finally, when we misbehave online (as we all invariably do at one time or another) we need to own up and apologize. Fortunately, the same attributes that make it easier to screw-up online, also make it easier to apologize online. You don't have to do it face to face, a quick tweet or email message works just fine. If you were at the SXSW keynote, you might consider that.

Predictions

This experience, and the past month that I've spent intensively using Twitter, have led me to a few beliefs about where this all is leading us.

The first one easy. Anyone who runs a conference, panel or large meeting without monitoring the backchannel is simply asking for trouble. Ironically, SXSW did have an official chatroom for the keynote, but that did not receive as much traffic, nor was it being monitored as a backchannel should be.

The second one is longer term. For several generations social networking on the computer has been derided as not having the depth or value of real life social interactions. Tools like Twitter (and Facebook), which blur the lines between work and home, important and trivial, and which deliberately create a malleable and ambiguous set of simple tools ("status", "poke", "what are you doing") are the primitive forerunners of what the next generation will take for granted. The always-on aspects will surely migrate to phones and become a constant part of our online life. The interfaces may be crude, but I am already more connected to the lives of people halfway across the world than I am with my next door neighbors. That knowledge extends from the trivial (I'd love to have dinner at Adam Engst's house, he cooks a lot of interesting stuff) to the critical (Susan Reynolds' fight with breast cancer has led to a wonderful support group and a great funding effort). It isn't a matter of not spending time with the neighbors, it's that I don't have a real-time, ongoing conversation with my neighbors day in and day out! The next generation is going to look back at pre-computer-mediated social interaction and say that we are the ones who had no depth in our relationships.

Methodology

• Searched for all users who mentioned SXSW between Thursday and Sunday the week of the conference (4068). For this I used the Terraminds Twitter Search API with a filter for the correct dates.
• Searched for all users who mentioned "facebook" (256), "lacy" (80), "keynote" (290) or "zuck" (111) between 1pm and 4pm on the day of the interview.
• I then took the list of 4232 unique users from the previous searches, and gathered every message they posted between 2:05pm and 3:15pm. (I initially gathered more, but then narrowed it down to those times as my best guess for the start/end times of the interview, as well as the most likely lag time for tweets.) The Terraminds search doesn't do per-user searching, so I used Tweetscan. Tweetscan doesn't provide an API, so I screen scraped the results.
• Narrowed down the resulting 3562 messages by splitting them into two groups. The first group matched the previously searched keywords (2039), the second did not (1523). I scanned the first group for messages that didn't look like they belonged (wrong topic, different session). I scanned the second for things that I might have missed. Those were quick scans and sloppy. The end result was to remove 66 entries and add 30.
• I should note that there are several things I didn't do but could have. That includes searching for additional keywords, and also checking the communication chains. E.g. If user @a sent a message to @b, then I should check @b's messages as well. I also did not include the Meebo transcriptsin the results, although that would be easy to add.
• The final result was 512 users and 1857 messages. (One user was removed in post-processing, when I realized that "twitgeistr" was a bot that simply reported on keywords that it found in the public stream.)
• I then spent an inordinate amount of time figuring out how to subtitle the video, including attempts at two different sub-title formats. The final solution involved Final Cut Pro and programatically generated XML files that specified text-effect overlays with differing offsets depending on the number of lines. If anyone at Apple wants to write a bit more documentation on FCP XML files, and provide a few more examples, that would be just fine with me. The annotations have all @user references changed to "@", and all links replaced with "[LINK]". The data is public, but I see no reason to make it easy to embarrass individuals.
  
• The keyword processing is also done with a Perl program. Lingua::StopWords was used to remove common English words from the list. Lingua::Stem was used to stem the words (e.g. make "improve", "improves" and "improved" all map to the same word). Stemming exceptions were made to ensure that the words in the top fifty were all spelled correctly (stemming programs don't really care if the result is a real word, only that the mapping is correct. "improves" normally gets stemmed to "improv"). In addition, I added some other common mappings. "sarah", and "lacey" both map to "lacy". "mark", "zuck", "z" and "zuckerburg" all map to "zuckerberg". "fb" maps to "facebook", and so on. I also added some additional stop words. I tossed "just", "like", "now", "can", "go", "got" and a number of other words that were very common but which didn't really carry any emotional content.
• The same Perl program generated the XML files, as well as a set of tab-separated data files for processing by DeltaGraph. DeltaGraph is a very powerful charting software (particularly if you are dealing with data sets with missing data), but some better (dare I say, "prettier"?) defaults and an updated UI wouldn't hurt.

If anyone is feeling particularly masochistic, I would be happy to package the whole mess up and make it available for download. Let me know.

Postscript

All the data gathering and analysis here were done by myself and for that, and any of the errors that are inevitable in such a rushed project, I am solely responsible. However, I'd like to thank a few people who contributed, knowingly or unknowingly.

• Jeremiah Owyang, Robert Scoble, Marshall Kirkpatrick and Dave Winer all discussed in detail what was happening, and why, on Twitter and on their blogs.
• Howard Rheingold is doing some very interesting work on backchannels and multitasking. His tweets on the daily progress of his Virtual Communities/Social Media class are quite interesting and relevant to this discussion.
  
• At a recent Boston Tweetup someone from the SOURCEBoston group suggested that I do a word analysis; a suggestion which added several days to the project, but was definitely worth it. Unfortunately I don't remember who it was, let alone their name.
• Also at the Boston Tweetup, Dmitri Gunn reminded me of the name of the other SXSW session which had a Twitter dustup, "Social Marketing Strategies Metrics, Where Are They?" (apparently that's what the audience wanted to know too) which in turn led me back to Jeremiah's excellent article on the different sessions where Twitter played a role.
• Dan Byler gave me feedback on a preliminary version of this article and brought up the mob scene in "To Kill a Mockingbird".
• Brett Peters reminded me that I hadn't gotten around to proving my claim that it wasn't just a few malcontents; thus sending me off to create two more charts just when I thought I was done.