Second-Order Consequences

Tldr

Notes on the term “second-order consequences”. I find myself wanting to use this term frequently enough that I would like to have some definitions to refer to.

The term “second-order consequences” comes up often enough. What does it mean?

It should be clear to most people that actions have consequences. In some cases it’s also clear that those consequences will have consequences of their own.

Looking up the term “second-order consequence” or “second-order effect” gives many definitions that are not much more restrictive than this. The following section gives some examples.

Definitions of Second-Order Effects in Popular Media

Every action has a consequence, and each consequence has another consequence. These are called Second-Order Effects. Every change you make to a system will have Second-Order Effects, which may affect the system’s functionality. Be careful when making changes, they may have the opposite effect of what you aimed for.

— https://personalmba.com/second-order-effects/

First order = what are the effects of the issue Second order = what are the effects of the first order issues.

To distinguish look at what the consequence is and where it came from. Did it from the issue itself or from the effects from the issue?

Also, outside of academia, distinguishing 1st from 2nd order effects can be really hard and sometimes impossible. Issues and consequences are never as neat as they look in a book.

— A reply to “How can one distinguish between first and second order effects/consequences?”, Reddit, deleted account

Too long to quote, but this blog from Farnam Street: https://fs.blog/second-order-thinking/

Scientific Definitions of Second-Order Effects

Example

First order effects included direct effects and substitution and were defined as negative and positive environmental impacts from ICT that stem from the production, use, disposal of the products involved, and as well as new possibilities due to ICT products. This definition is based on that for direct impacts stated by Berkhout and Hertin (2004). Second order effects included effects that go beyond the direct consequence of an activity, i.e. effects of effects or indirect effects. As “second order effects” we include a number of effects described in the literature as e.g. “ripple effects”, different types of “rebound effects” and “structural and transformational changes”.

… Second order effects:

Rematerialisation effects
Induction
Direct economic rebound effects
Indirect economic rebound effects
Economy-wide rebound effects
Time rebound
Space rebound
Learning about production and consumption
Scale effects and learning in production and consumption
Changed practices
Transformational rebound effects

… We opted to use the terms first order and second order effects but we did not attempt to make a clear definition of what second order effects are. Instead, we let them be defined as described in the literature and took them as going beyond the first order effects. Thus our use of second order effects does not entirely coincide with that in previous studies, e.g. some second order effects as defined by e.g. Abukhader and Jönson, 2004, Erdmann and Hilty, 2010 and Yi and Thomas (2007) ended up in our first order category. This was the case for substitution effects, because we argue that such effects are part of what was intended with the product/service and hence they should be counted as first order effects. As another example, Erdmann and Hilty (2010) define second order effects as the impacts of the services provided by ICT applications, for example energy saved from using a service. We included such effects in the category of first order effects. In contrast to these authors, we also refrain from separating second order effects from third order effects. We consider it is difficult and often not useful to distinguish the application of ICT from the ICT practice. They are interdependent and interwoven and we argue that the separation between the two often depends on the different disciplinary background of the beholder, rather than an actual division.

— [@borjessonriveraIncludingSecondOrder2014]

• First-order effects have a cited definition, second order effects defined but not cited.

Problems with Common Definitions

These descriptions don’t seem to draw a clear boundary between first and second-order. I’ll try to illustrate what I mean with some examples.

1. Some effects-of-effects are so obvious that calling them second-order seems ridiculous

Imagine a set of dominoes arranged in the usual way, such that tipping the first knocks over the second, which knocks over the third, and so on. If I tip the first domino, it is clearly a first-order consequence that the first domino falls over. But anyone with a basic grasp of “folk physics” will know immediately that the whole set of dominoes is bound to fall. The falling of the second domino is a consequence of a consequence, so it should be second-order by the lay definitions. But this is not the kind of sophisticated, deep, complex thing that the popular media seems to be pointing to.

2. What an “effect” is depends on the level of coarse-graining

Furthermore, the ability to label things as first- or second-order is sensitive to our level of analysis. If I initiate a phone call to a friend, most would say it is a first-order consequence that my friend receives a call on their phone. But there is a complex chain of events that facilitates this. The touch screen on my phone registers a touch, the operating system identifies this as a button press, the CPU processes a set of instructions, and so on, and eventually my friend interprets stimulus from their eyes and ears as “I am getting a phone call”. The operating system registering a button press is a consequence of my having touched the screen — why is my friend getting a phone call a 1st-order consequence and not an nth-order consquence? Presumably only because we agree to keep things coarse-grained.

What do I actually mean by second-order?

Second-order as a level of unpredictability

For me (and this is purely personal, though I think others would probably agree), “second-order consequence” connotes something difficult to predict.

So in one sense, I use meanings like:

First-order consquences: the set of consequences that are easy to predict Second-order consquences: the set of consequences that are not easy to predict

Second-order can be divided into higher orders by adding more degrees of difficulty…

Third-order consquences: the set of consequences that are hard to predict Fourth-order consquences: the set of consequences that are very hard to predict

But this by itself doesn’t seem sufficient. There are many consequences that are difficult to predict that are not interesting or consequential, so it seems there is also a condition that the consequences must be important.

Second-order as a category of origin

I think a lot of the time that I say “second-order”, I actually mean something more like “effects that result from intelligent, goal-oriented agents taking action in the world produced by the first-order effects”.

To break this down, I imagine a sequence of events like:

1. An action is taken by agent $A$. This is the “cause”.
2. An effect ensues, mediated by rules of the environment but not mediated in any meaningful way by other agents. The state of the environment is modified by these effects.
3. Other agents, for example agent $B$, are observing and acting in the world in order to fulfill their goals. As a result of the state of the environment having changed in step 2, these agents act differently than they would have if no action had been taken in step 1.
4. The actions of other agents (Agent $B$ and so on) have effects. These are first-order with respect to $B$‘s action, but second-order with respect to $A$‘s action.

Thus in this definition, an effect $e$ is $i$-th order with respect to some action $a$ if the chain of events from $a$ to $e$ passes through the decision-making process of some intelligent agent $i-1$ times.

Note that this does not mean that Agent B needs to be adversarially targeting A. B does not even need to be aware of Agent A. It is only required that A’s action creates some appreciable change in B’s environment, and B’s action likewise creates some appreciable change in A’s environment — in other words the two environments overlap.

Does this definition solve the problems with common definitions? I think it’s a good start.

1. The effects-of-effects in this definition are only “obvious” (the first of the problems) if predicting the behaviors of another intelligent agent is so easy as to be “obvious”. This is true in some circumstances but not in general. It is not always very difficult, but it is sublter than predicting the actions of dominoes.

2. This definition naturally sets the level of coarse-graining at the level of intelligent agents and their actions.

Using the “Second-Order Consequences” Concept

Most of the times that I use the concept, I am making a practical request like “people should anticipate the second-order consequences of X technology and prepare for them”. I think previously I have made this statement without presenting a clear way to operationalize this. I think a more precise definition makes it easier to act on this.

If “second-order consequences” are less predictable than the first-order, then the practical request above is intrinsically hard relative to first-order thinking alone. Second-order thinking can be difficult, but this is just downright discouraging. The only thing you know is that you will have a hard time, not much else is given by this definition. This is one reason why I don’t like definitions like “second-order consequences are the things that only smart people see coming” — insinuated if not said explicitly by many of the pop media presentations.

If “second-order consequences” are consequences that result from the actions of a goal-directed agent subjected to the first-order consquences, this can still be a daunting task, but at least there is a way to start reasoning. Based on Agent B’s goals, we might be able to infer how they might act in the world we created with action $a$. The analysis plan is not just “be smart” or “don’t be rash”, but “consider what approximately rational, approximately goal-directed agents with goals different from yours will do in the world that results from your actions”.