Conflict and error are human constants?
What can Wikipedia tell us about trust?
No one seems to know for sure when humans started cooking, but most would agree that cooking was the technological leap (not in fact the 2001 monolith) that allowed humans to advance beyond an animal livelihood, to spend less time micromanaging their survival, and to go on to do greater things — basically thanks to a greater energy intake per unit time.
Like cooking, trust is a strategic choice that allows us to spend less time attending to (or micromanaging) essential processes that require our attention. Many processes (especially those arising in nature) are bursty in nature and therefore only require spotty interest. We can get by happily by checking whether progress is being made just occasionally–checking from time to time, rather than obsessively asking “are we there yet”?
The watched pot never boils and that’s because we can’t tell time is passing when we focus on smaller and smaller intervals of time in a fit of mistrust.
In 2006, my friend and colleague Jan Bergstra and I started to think about trust through the lens of Promise Theory on a visit to the University of Bremen, sponsored by the EMANICS research Network of Excellence. As we were developing our early ideas, disappointed with what people had written earlier about trust, we came to understand about half of the problem surprisingly well. Trust seemed clearly related to promise keeping, but we didn’t figure out the whole picture from the beginning.
We wrote up our notes and forgot about the matter for a couple of years before uploading a summary to the arXiv, and then forgot about it again for another 15 years! I returned to the topic this year, reminded of trust by some work I’d been doing on virtual motion. It gave me the idea that trust was a kind of virtual currency, like energy, and I was able to pick up this work and continue it, thanks to some funding from NLnet!
An updated understanding of trust
With a deeper understanding, I showed that trust comes in two forms:
i) assessments about promises that are kept (from the work with Jan; this is what leading to a sense of trustworthiness, which is essentially reliability in promise keeping), and
ii) what I call kinetic trust, which is a policy for sampling evidence, i.e. how closely we pay attention to the process of promise keeping. Kinetic trust is about resource saving during observation.
If we turn it on its head, kinetic mistrust is an investment of attentiveness. The less we trust, the more we look. We all know the cold war addage “if you can’t trust, verify”, yet verifying is expensive, so we prefer not to.
Like cooking food to unlock nutrients, trusting a process allows us to unlock time saving from obsessive watching.
I’ve been developing a hypothesis about trust as a currency in a series of studies, summarized on my research page. The goal is to understand the meaning of trust more thoroughly for application to our augmented cybernetic world of human-computer interaction. Much of our lives has shifted online today, so much so that we need to grok the role of social behaviour in more detail to understand and govern our behaviours for civil online society.
Trust in process: Wikipedia
In the latest installment of this work, I turned to Wikipedia for data about trust and mistrust. Wikipedia is made up of entirely voluntary contributions from users. It’s a prime example of voluntary cooperation, in the Promise Theoretic sense, and its data are open and inspectable to all. That makes it a rare and valuable source of data for study.
Each article is started by someone, but gradually attracts the attention of other people as well as a cluster of automated “bots” programmed to find errors and mischief. Once a new article is discovered by others, changes will be made, corrections and improvements will be added. These edits tend to cluster into bursts of activity, punctuated by longer intervals of inaction. All the information is logged so we can see how this process unfolds.
What does trust have to do with Wiki editing? The answer is complicated, but it starts with whether or not authors write reliable information in an online society of shared interests.
What does reliable mean? From a Promise Theory viewpoint, it has to with keeping a promise consistently. There’s an implicit offer (+) promise that authors make: I promise that my information is accurate, impartial, and documented with references. Authors are never asked to promise this explicitly, with their hand on a favourite book, but Wikipedia editors and moderators expect this, and thus only promise to accept (-) contributions that they assess to keep this promise. This mutual alignment of intent is what makes promises important, like a potential or force in physics.
There are more implicit promises too: I promise not to write about myself or promote something as simple marketing, only to report referenceable facts. The effect of this promise is not always what the editors may intend: by this rule, it’s okay to refer to idle rumours spread by tabloid news muckrakers, but it isn’t okay to report first hand information like a witness or informant. The editors trust anything published, but not individuals.
However much we may question the consistency of these promise decisions, these are the promises that Wikipedia expects authors to uphold. Trustworthiness (potential trust) is i) the extent to which an article yields accurate information, and ii) the extent to which the editing process keeps the implicit promises above.
There’s a few things we can check using the Wikipedia information:
- Is conflict (mistrust) over content constant or occasional? This is a scale-dependent question, so we make it scale free by looking at probabilities. The key point is that conflict sustains busy attentiveness and kinetic cost.
- How many users are active at the same time, interacting (effective network cliques)? Is the work associated with one to one interaction, or group contention. This is less clear from the data, but the group action gives the least uncertain, most focused fit.
- Does the amount of fighting depend on the number of users? The data show that the level of contention is fairly constant over long times, albeit bursty on a shorter timescale.
It turns out that the number of agents (humans, bots, etc) involved at any one time is quite stable at between 5–15 (one of the Dunbar numbers). See the figure 1 below (depicting figures 5 and 6 from my research notes).
We might expect that, as time goes by, an article approaches a steady state of being finished and that the fighting would settle down. This doesn’t seem to be the case. The mere fact that people come together over a common cause seems to perpetuate new episodes of intended change and disagreement! At any time, there's about the same number of people fighting per unit length. We see this in the figure 2 graphs below:
The same average number of users is quite constant over time, and never really dies away. In other words, the length of an article is only a partial proxy for elapsed (proper) time. Different people work on different sections, but the numbers are about the same. This is reminiscent of Dunbar's hypothesis.
Moreover, the contention about issues does not seem to depend on the issue. Quantum mechanics is as contentious as political actors and popstars. People just seem to like disagreeing.
The energy hypothesis for trust
To measure kinetic trust, we want to know whether a subjective assessment of a sub-sample of an process gives a meaningful impression of its content or at some level. There are two processes in Wikipedia: i) the story told by the article, and ii) the meta process of editing and correcting it.
Based on a frequency of sampling (attention) — does the process make sense to a neutral third party observer? I discussed this in the previous articles. This does seem to be the case if we promise to uphold a minimum level of sampling–for a minimum sampling velocity. This sampling velocity can be reduced, without breaking the promise, if there is a lot of repetition or padding in an article.
So how do we choose the sampling rate? A detached observer has no particular idea what rate applies to which article. However, by trial and error, the same observer tries different rates of skimming the material until it seems like an acceptable balance and calls this a trusted velocity. The level depends on how rich the source text is in its intended content. So how can we measure that?
To measure intent, we go back to the bioinformatic method I discussed in an earlier installment. The potential trustworthiness of a process is related to how reliably it keeps its promises and how much that costs. Longer words take more work to write, so they are likely more intended than shorter words that are mostly padding. Words that are repeated a key amount are more intended than those appearing spuriously or those repeated incessantly. This gives a potential function for intentionality.
Following the energy hypothesis, we measure the richness of text in terms of a measure called its Intentionality. Summing up the intentionality of all the text “DNA” fragments gives an impartial measure of its intent (effectively “what it promises” ). If we seek out these highest potential sentence events in a stream, we should be able to reduce our level of attention to a minimum. But what is that minimum? The energy hypothesis gives a simple approximate principle:
change in potential trust = change in potential trust
Equating these amounts gives an average rate of sampling
- Cost of intentional work V ←→ summed for all n-grams per sentence event
- Cost of attention (kinetic mistrust) T ←→ ½mvˆ2, determines sampling velocity v.
In Promise Theory, agents are at the centre of all that happens, because agents are the sources of autonomous behaviour. This is equivalent to the principle of locality for agent causation. Agents can be people, bots, or even groups (superagents) acting coherently.
In physics, the conservation of energy is partly what makes it useful. Why should it be conserved? We don't know. Here, the constancy of behaviour is what justifies treating the equation as universal. It's a statistical law. For physics, we don't know.
Apart from the accuracy and stability of the article text, we can turn to the changelogs to associate the rate of kinetic infighting (do-undo cycles, write, alter, combined with arguing) with the level of mistrust around the change process. For either of these, the energy hypothesis works quite well, but of course the conclusion is somewhat ad hoc.
Can we say this hypothesis is “true”, as we would argue in the case of natural science — for physical energy? That’s an interesting question, because we can’t actually even say it’s true in physics. All we can say is that this definition works as a model and that it seems to be universal (at least within the realm of Wikipedia). So, we define it to be true and trust that it makes sense, even though we check from time to time. This is how physics works too. But here, we do the same with trust directly. So far, the results for trust seem universal, but with a much higher level of noise than for an elementary system in physics.
Afterword
The trust assumption is hard to escape in any context. It pervades everything we do. It’s all about the information economy. Philosophy began with religion and moral ideas before science gained a foothold, but this shows that we don’t need to cling on to moral notions to understand trust. Even complex humans seem to be guided by simple countable measures on a coarse scale. Some will find this idea objectionable, hoping to preserve our special nature–just as the Victorians wanted to place us at the top of a ladder to god. Alongside all the recent work on “AI”, this study suggests that isn’t appropriate. Humans are just complicated participants in complicated processes.
The reason why we get confused about trust as humans is that we tend to imagine (make up) promises that we would like others to have made, and judge them on those. "You didn't do what I wanted!" Such behaviours are “impositions” in promise language, and they are generally unreliable and untrustworthy.
The logical opposite of a promise, in a sense, is an unfounded accusation–the imposition of an unreliable promise of truth. Promise Theory allows us to make these simple objective reformulations of trust, from a moral conception to being a process attribute.
We can say a lot about how one should justify trust, but the simplest approach is to think of it as a process that costs work proportional to the “proper time” of the process. The more effort something takes to complete, the more we tend to invest in it–but only if it keeps the promises we perceive it to make.
