What are bots?
Recently I watched a presentation from the last CCC called “Social Bots, Fake News und Filterblasen”. The presentation was give by a data journalist Michael Kreil, who applies data science toolkit to social networks. Mostly he gets data from Twitter, because Twitter public API enables comparatively easy access to user publications and user records. He analyses the data he can gather and publishes reports about, for example, social profile of a typical person spreading fake news stories.
When doing such a research it is natural to question if the accounts under study backed by real people? One particular phenomenon usually contrasted to real people often called social bot. Surprisingly enough the question who is a bot and who is a person is not so easy to answer, because on the Internet you don’t another side of the cable. What is social bot and how to identify it was one of the main topics in the talk. Of course the discussion started with the very first question about social bots: How one would define a social bot? Among all possible definitions the presenter also was interested in a scientifically unambiguous one. And as turned out the more scientific the definition is, the harder is to count the bots? I agree that the question of counting is important, nevertheless I think that the way how Michael defined social bots missed or at least understated the root: what is our motivation to study social bots?
In the talk we see the definition of a social bot given by TAB in a position paper Social Bots(my translation):
Social bot is a computer program, created with a manipulative purpose, which imitates a human personality and communicates with people on the Internet.
Once there is a definition one requires a method to count entities who fall under. This question is hard to answer, because a successful social bot can pass a Turing test and thus is indistinguishable from a human. The presenter discusses various methods (around 30:00+) to figure out who is a bot on Twitter. The methods range from primitive ones, like having a threshold for number of tweets per day, to sophisticated machine learning based algorithms which analyze the whole network of twitter-followers of a given account. The problem with the former is being scientifically silly. The problem with the later is being hardly verifiable.
From my point of view any possible way to identify social bots in a sense of aforementioned definition is conceptually problematic. A simple method knowingly ignores data exposing computer nature of an account. A sophisticated one has to play a Turing test game better than humans. This is hard to do, and still an ideal social bot is truly indistinguishable from an individual. Michael Kreil acknowledges that it is possible to estimate if an account run by a program only with a certain confidence.
I think the whole issue of counting bots comes from this precise definition of what a bot is. Breaking the definition into individual statements we see that a social bot is a computer program. It is deemed to manipulate human opinion. And a social bot communicates with other people. It is not too far fetched to say that any Twitter client matches the definition. Any Twitter client is a computer program. A Twitter client manipulates people opinion on behalf of the user of the client. And of course a twitter client is supposed to be reachable by other people. All in all social bots also act on behalf of real people. Then what is the difference between a social bot from TAB definition and a usual Twitter client which publishes posts on behalf of its user? Not so much!
The reason we care about social bots not because we are afraid that there is a too thick software layer between people, but because when people exchange opinions, they want to know what does the society around them thinks. Most people are hardwired to be susceptible to judgements of other people. Since social bots pretend to represent an opinion of some real person, in the worst case they can trick more susceptible individuals in believing in a fiction. We much easier believe that something terribly unjust happens around us of our circle of reference says that. When such belief is established, anybody can become radicalized, even if the reality gives no apparent reason for such kind of belief.
Those who are not tricked into false narratives of social bots are also affected. Seeing it as an attempt of a deception, post of social bots easily become a source of anger. Still people are equally angry if they find themselves trying to dialogue with a real person, but whose sole purpose is to provoke them. The whole anxiety about social bots is very similar in nature to distress about internet trolls. We don’t care how much compute power was put into making something, what we can’t consider to be a valid opinion. Thus the definition of a social bot as a computer program superfluous for most practical purposes.
This intuition behind the term is coherent with frequently used colloquial understanding social bots as a concept. Infamous Internet Research Agency operates hiring real people, who post opinions under fake identities with an attempt to create an impression that certain opinions are represented more than they actually are (one, two). Employees of IRA are often called trolls from Olgino (by the office location) or kremlebots (by their political orientation). Note, how words troll and bot are used interchangeably. As the first article describes the kremlbots have to post at least 50 comments per day on social networks. From the first glance, this would qualify them as social bots according to criterion mentioned by Michael Kreil. But the trolls have to maintain at least 6 Facebook accounts, which means that on average one account will publish less than 10 posts. Such a rate is even less than some moderate social network users. And still this group of people should be included into the definition of social bot, otherwise the definition simply omits an important dimension of the phenomenon it is deemed to describe. As an alternative, I propose following definition of a social bot.
Social bot is a social media account which represents non-unique identity of a human person. Unique identity of a human person is an online identity, which uniquely maps to a human person, represents a specific set of interests, thoughts and opinions of this person and does not covertly overlap with other active unique identities of the same person.
A unique identity is a new concept in the definition, which is usually an account on a social network used for a particular purpose. For example, a user may have two Twitter profiles, one used to publish information related to professional activity, another one used to publish information about personal life. Even if these two accounts are not explicitly connected few people will be deceived into thinking that the two accounts represent two opinions.
Even if it happens so, that, for example, the first account retweets a post of the second as long the fact that the two accounts are related to the same person remains clear, nobody will be deceived into perceiving original post and retweet as opinions of separate people. This definition does not require accounts to be explicit about who they are mapped to. Both accounts can stay pseudonymous. Even if the user creates a new account for each and every new post, neither of the accounts falls under the definition of social bot, because if an old account is discontinued, it can’t overlap with a newly created account. Of course, to avoid confusion, the fact that the account is discontinued should be evident.
Another example is a CNN tweeter bot posting definitely more than 50 tweets per day, but still not qualifying for a social bot, because it unambiguously represents an organisation, not a person. On the other hand an IRA employee, even if posting less than 50 comments per day, qualifies as a social bot, because the accounts the employee runs map to the same human and express opinions on related topics.
This definition works around an unsolvable problem of identifying how much of algorithmic effort was put into creating a post on social media. On the other hand this suggests that analysis on how many social bots are there on the network purely from the account behavior renders to be infeasible. We’re definitely past the time when such analysis could be considered reliable, so better not even try.
Such a conclusion does not mean that you can’t trust anybody on the Internet. I just think that to factor out social bots one should do it on a technical level. A crude example of technical measures for human validation is provided by Tweeter and Facebook. These both companies have a teams of specialists whose task is to check an account holder credentials and prove that the displayed name matches the person’s real name. After a successful verification an account gets a special sign signalizing others that the account holder was validated. I see a number of shortcomings of this process. First, it is quite cumbersome and expensive. Few companies provide it, and still only for small fraction of users. Second, it restricts the user anonymity.
I believe better measures are possible. A properly built technical system should have much less personnel interaction for issuing new user identities and still allow more freedom for anonymity. If a social network consists only of users who pass this form of verification the problem of social bots simply falls off.
Update As a friend of mine correctly pointed out was formally described as Sybil attacks.