Narrative Syncing

Partial attempt to make sense of: On Doing the Improbable

Epistemic status: On my inside view, this seems like a clear and useful distinction that could use a short, shared term. Hasn’t been checked much by others. Some edge cases I’m still confused about. Also I may have a suboptimal term; help generating the right term greatly appreciated.

I’d like a short term for such sentences as:

  • “The sand is lava; if you touch it you die” (when introducing the rules to a game, not making a false prediction)

  • “Colloquium is at 3pm on Wednesdays” (when creating a new colloquium, or adding weight to a common-knowledge agreement /​ Schelling point)

  • “We don’t shout around here” (stated normatively)

I’m going to suggest the term “narrative syncing.” I’d like to distinguish “narrative syncing” from “sharing information” of a sort whose accuracy is not intended to be created/​affected by the act of thinking/​stating/​believing the given sentence, i.e. to distinguish the sentences above from such sentences as:

  • “It will probably rain tomorrow.”

  • “3 + 5 = 8.”

  • “The people over there in that organization seem to avoid shouting, and to somewhat shun those who shout – though I have no stance on whether that’s a good norm, I’m just describing/​predicting it.”

That is, I’d like to use the term “narrative syncing” (or some better such term that y’all suggest; suggestions appreciated) to distinguish sentences whose primary purpose is to sync up with some set of other people as to how we all think or talk about a given thing around here (as opposed to sentences whose primary purpose is to describe a piece of outside reality).

A situation where the concept of “narrative syncing” seems to me to add clarity

One reason I’d like a short term for this, is IMO having such a term would’ve helped me understand some past social situations I was in.

Example: Narrative syncing about “helpful” careers

Sometimes participants at CFAR’s/​MIRI’s “AI risks for computer scientists” program would ask me something like “what careers should I go into, if I want to help with AI risk?” And I would sort-of interpret this as a request for information, and would nevertheless feel a predictable frustration when I heard it as I braced for the subterranean/​confusing-to-me conflict that I anticipated would follow my words, without really understanding why (I would anticipate this based partly on the particular person’s nonverbals as they asked the question—not all participants were like this). I’ll elaborate how this interaction tended to go by describing a stereotyped conglomerate participant of this sort named Alec:

Scenario 1: “I try to tell the truth”

Alec: “What careers should I go into, if I want to help with AI risk?”

Me: “Well, unfortunately, I’m not really sure. My own models are messy and confused, as to what paths there are to an existential win, what actions of yours might end up helping how much with each, etc. I’d mostly recommend you try to develop inside views on how to get to victory; although I happen also to like MIRI, who is [at that time] hiring engineers and scientists; Paul is also looking for ML engineers, and seems smart and like he’s really trying; and ML seems likely to be a skill that more and more groups that say they are aimed at safety are trying to hire people with… But there aren’t ready-made careers I know of that produce some reasonable guarantee of helping.”

Alec: [nonverbals I interpreted as frustration/​anger/​desire to get me to act “correctly” again/​turmoil]: “Why did I even sign up for this program?”

Me, to myself: [… I guess people just really want me to simplify, so they can believe the world is easier than it is? I notice I’m confused, though.]

Scenario 2: “I go along with what he seems to want”

Alec: “What careers should I go into, if I want to help with AI risk?”

Me: “For most people it seems probably best to study ML, or computer engineering broadly, and then apply to AI safety orgs. Math/​physics is probably also a decent bet. You can also try to develop an inside view of the problem, by thinking about how the heck you’d code an aligned AI, and e.g. making the puzzle easier to get started on by imagining you get to start with a hypercomputer, or the super-duper-deepnets of the future, or something.”

Alec: “Okay.”

Me, to myself: [huh, I feel vaguely dirty/​contaminated after that exchange; not sure why; I guess it’s because I was simplifying? My relationship to Alec makes me feel weird, now.]

My current models of what was happening here

The way the above interaction seems to me now, is that Alec was actually trying to request narrative syncing rather than information exchange. He wanted to know: “which careers are should I go into, if I want you /​ MIRI /​ the EA community to, in exchange, regard me as ‘committed’ and as ‘a real member of the group, who should be supported and included and viewed as legitimately taking part in the shared effort’? What are the agreed-on local rules for which careers count as ‘AI Safety’?” And in response to this question, in scenario 1, I was sort-of saying “sorry, I don’t play that game,” but I was doing it non-explicitly, in a way that was hard for Alec to process. This was especially so since I was sending other signals, elsewhere in AIRCS, that I was playing games of that sort – that I did play a conscious role in the community’s decisions about who was/​wasn’t inside the community, or about which ideas counted as legitimate. Thus Alec felt frustrated; and I sensed a difference in wavelengths but failed to quite understand why.

In scenario 2, I was answering the question in a way Alec could take as narrative syncing: “These are the careers such that, if you pursue them, I and such others as I can predict/​influence/​sync with will regard you as being part of the real AI safety effort; I will take on some of the responsibility for your choice; if you choose these things, your views and actions can be part of us.” But I wasn’t actually okay with doing that, and so I felt weird/​bad/​as though my new relationship with Alec was unacceptable to me in some way.

If I had that scene to do over again:

Scenario 3: Using my new understanding of “narrative syncing”

Alec: “What careers should I go into, if I want to help with AI risk?”

Me: “We try to have a culture around here where there is no vetted-by-the-group answer to this; we instead try to encourage forming your own inside-view model of how AI risk might work, what paths through to a good future might be possible, etc. One nice thing is that money is pretty abundant lately: if you end up with an inside-view saying a particular research direction is worth trying, you’re almost certain to be able to get funding to pursue it, at least after you develop it a bit (given you’re starting as a smart PhD student), and I’d be glad to help you seek funding in such a situation if you want. I’m also happy to share my own inside-view models, but you should know that I might change them, and that others in AI safety may disagree.”

(Note that in this example, the “me” character does do narrative syncing – but in a way that does not disguise itself as information sharing about AI. Instead, I bid for the norm that “we” try to have a culture “around here” in which we each form an inside view, rather than syncing about which inside view of AI is the “us view” or the “good view” or something. This may still involve sharing misleading info about how “we” do things around here, which may still lead to problems.)

My take-aways from the above

It seems to me that narrative syncing is not always bad, and in fact, has real uses. It is nice to be able to create “house rules” for a board game, to decide that colloquium is at 3pm on Wednesdays, and to create norms for what “we” do and do not “do around here.”

However, narrative syncing is sometimes mistaken for (or deliberately disguised as) information sharing, and this is bad for epistemics in at least two ways:

  1. Direct misunderstanding: People often think other people are giving them information/​predictions when they are actually making social moves, which leads to updates they wouldn’t make if they understood what was happening.

  2. Indirect “inquiry-dampening”: If there is an attempt to cause social coordination around some “what we say around here” matter that sure sounds like a matter of prediction or information exchange (such as whether it’s harmful to advance ML research, or more local matters like whether so-and-so chose good groceries for the workshop) this is apt to create social pressure for others not to express their own views freely, which is bad for epistemics on this point, and which risks setting up “watch what you say” conversation-dampening dynamics broadly.

So, it’s valuable to be aware that “narrative syncing disguised as information requests /​ information sharing” is a thing, and to have some notion what it looks like, what kind of places you’re likely to find it, etc. Much as it’s valuable to be aware that people lie sometimes, and to have some notion what kinds of things people are likely to lie about.

(Though, unlike the situation with more textbook examples of “lying,” I’d like to emphasize that most people doing this “disguised narrative syncing” stuff do not have an explicit model under which this is harmful, and so the thing to do is more like to look into things curiously and collaboratively, or to just sidestep the dynamic, and less to tell people “cut it out please” as though it has been established that that’s the thing to do. Also my concepts need vetting.)

Less importantly: understanding the positive uses of narrative syncing seems at least a bit useful to me also vis a vis understanding what social requests people sometimes have, and being able to meet those requests more straightforwardly. Though this part is more speculative, and is not a crux for me for most of the above.

My remaining confusions and open questions

My biggest remaining confusion is this: what’s up with self-fulfilling (or partially self-fulfilling) predictions playing a role in internal coordination? For example, I tell myself that I’d like to write this before I sleep, and now I have more of a desire and plan to write this than I did before; or, parents tell a kid they like the color blue and it catches; etc.? What’s a good ontology around this? “Truth-telling” vs “lying” vs “bullshitting” doesn’t seem like the most natural set of distinctions for this sort of loopy sentence. And while this is arguably a weird edge-case if your goal is to make Bayesian predictions about the outside world, it seems basic and not at all an edge-case if your goal is to understand what humans are or how humans or groups of humans act in the world or where agency comes from.

A few other questions that’re still open for me, mostly as recap:

  • Is “narrative syncing” an okay enough term that we can use it for this thingy I’m describing? Anyone have an idea for a term that better cleaves nature at its joints?

  • Does “narrative syncing disguised as information sharing” have important positive uses?

  • Is it [edit: narrative syncing disguised as information sharing] something we should simply try to minimize if we want good epistemics? (My current hypothesis is “yes,” but this hasn’t been checked enough.)

  • What’re some places where this happens a lot? Where it happens almost not at all? How can a person notice it? How can a person shift things toward less confusing patterns?