This is Section 2.2.3 of my report “Scheming AIs: Will AIs fake alignment during training in order to get power?”. There’s also a summary of the full report here (audio here). The summary covers most of the main points and technical terms, and I’m hoping that it will provide much of the context necessary to understand individual sections of the report on their own.

Audio version of this section here, or search "Joe Carlsmith Audio" on your podcast app.

"Clean" vs. "messy" goal-directedness

We've now discussed two routes to the sort of beyond-episode goals that might motivate scheming. I want to pause here to note two different ways of thinking about the type of goal-directedness at stake – what I'll call "clean goal-directedness" and "messy goal-directedness." We ran into these differences in the last section, and they'll be relevant in what follows as well.

I said in section 0.1 that I was going to assume that all the models we're talking about are goal-directed in some sense. Indeed, I think most discourse about AI alignment rests on this assumption in one way or another. In particular: this discourse assumes that the behavior of certain kinds of advanced AIs will be well-predicted by treating them as though they are pursuing goals, and doing instrumental reasoning in pursuit of those goals, in a manner roughly analogous to the sorts of agents one encounters in economics, game-theory, and human social life – that is, agents where it makes sense to say things like "this agent wants X to happen, it knows that if it does Y then X will happen, so we should expect it do Y."

But especially in the age of neural networks, the AI alignment discourse has also had to admit a certain kind of agnosticism about the cognitive mechanisms that will make this sort of talk appropriate. In particular: at a conceptual level, this sort of talk calls to mind a certain kind of clean distinction between the AI's goals, on the one hand, and its instrumental reasoning (and its capabilities/"optimization power" more generally), on the other. That is, roughly, we decompose the AI's cognition into a "goal slot" and what we might call a "goal-pursuing engine" – e.g., a world model, a capacity for instrumental reasoning, other sorts of capabilities, etc. And in talking about models with different sorts of goals – e.g., schemers, training saints, mis-generalized non-training-gamers, etc – we generally assume that the "goal-pursuing engine" is held roughly constant. That is, we're mostly debating what the AI's "optimization power" will be applied to, not the sort of optimization power at stake. And when one imagines SGD changing an AI's goals, in this context, one mostly imagines it altering the content of the goal slot, thereby smoothly redirecting the "goal-pursuing engine" towards a different objective, without needing to make any changes to the engine itself.

But it's a very open question how much this sort of distinction between an AI's goals and its goal-pursuing-engine will actually be reflected in the mechanistic structure of the AI's cognition – the structure that SGD, in modifying the model, has to intervene on. One can imagine models whose cognition is in some sense cleanly factorable into a goal, on the one hand, and a goal-pursuing-engine, on the other (I'll call this "clean" goal-directedness). But one can also imagine models whose goal-directedness is much messier – for example, models whose goal-directedness emerges from a tangled kludge of locally-activated heuristics, impulses, desires, and so on, in a manner that makes it much harder to draw lines between e.g. terminal goals, instrumental sub-goals, capabilities, and beliefs (I'll call this "messy" goal-directedness).

To be clear: I don't, myself, feel fully clear on the distinction here, and there is a risk of mixing up levels of abstraction (for example, in some sense, all computation – even the most cleanly goal-directed kind – is made up of smaller and more local computations that won't, themselves, seem goal-directed). As another intuition pump, though: discussions of goal-directedness sometimes draw a distinction between so-called "sphex-ish" systems (that is, systems whose apparent goal-directedness is in fact the product of very brittle heuristics that stop promoting the imagined "goal" if you alter the input distribution a bit), and highly non-sphex-ish systems (that is, systems whose apparent goal-pursuit is much less brittle, and which will adjust to new circumstances in a manner that continues to promote the goal in question). Again: very far from a perspicuous distinction. Insofar as we use it, though, it's pretty clearly a spectrum rather than a binary. And humans, I suspect, are somewhere in the middle.

That is: on the one hand, humans pretty clearly have extremely flexible and adaptable goal-pursuing ability. You can describe an arbitrary task to a human, and the human will be able to reason instrumentally about how to accomplish that task, even if they have never performed it before – and often, to do a decent job on the first try. In that sense, they have some kind of "repurposable instrumental reasoning engine" – and we should expect AIs that can perform at human-levels or better on diverse tasks to have one, too.^[1] Indeed, generality of this kind is one of the strongest arguments for expecting non-sphex-ish AI systems. We want our AIs to be able to do tons of stuff, and to adapt successfully to new obstacles and issues as they arise. Explicit instrumental reasoning is well-suited to this; whereas brittle local heuristics are not.

On the other hand: a lot of human cognition and behavior seems centrally driven, not by explicit instrumental reasoning, but by more locally-activated heuristics, policies, impulses, and desires.^[2] Thus, for example, maybe you don't want the cookies until you walk by the jar, and then you find yourself grabbing without having decided to do so; maybe as a financial trader, or a therapist, or even a CEO, you lean heavily on gut-instinct and learned tastes/aesthetics/intuitions; maybe you operate with a heuristic like "honesty is the best policy," without explicitly calculating when honesty is or isn't in service of your final goals. That is, much of human life seems like it's lived at some hazy and shifting borderline between "auto-pilot" and "explicitly optimizing for a particular goal" – and it seems possible to move further in one direction vs. another.^[3] And this is one of the many reasons it's not always clear how to decompose human cognition into e.g. terminal goals, instrumental sub-goals, capabilities, and beliefs.

What's more, while pressures to adapt flexibly across a wide variety of environments generally favor more explicit instrumental reasoning, pressures to perform quickly and efficiently in a particular range of environments plausibly favor implementing more local heuristics.^[4] Thus, a trader who has internalized the right rules-of-thumb/tastes/etc for the bond market will often perform better than one who needs to reason explicitly about every trade – even though those rules-of-thumb/tastes/etc would misfire in some other environment, like trading crypto. So the task-performance of minds with bounded resources, exposed to a limited diversity of environments – that is, all minds relevant to our analysis here, even very advanced AIs – won't always benefit from moving further in the direction of "non-sphex-ish."

Plausibly, then, human-level-ish AIs, and even somewhat-super-human AIs, will continue to be "sphex-ish" to at least some extent – and sphex-ishness seems, to me, closely akin to "messy goal-directedness" in the sense I noted above (i.e., messy goal-directedness is built out of more sphex-ish components, and seems correspondingly less robust). Importantly, this sort of sphexish-ness/messy-ness is quite compatible with worries about alignment, power-seeking, etc – witness, for example, humans. But I think it's still worth bearing in mind.

In particular, though, I think it may be relevant to the way we approach different stories about scheming. We ran into one point of relevance in the last section: namely, that to the extent a model's goals and the rest of its cognition (e.g., its beliefs, capabilities, instrumental-reasoning, etc) are not cleanly separable, we plausibly shouldn't imagine SGD being able to modify a model's goals in particular (and especially, to modify them via a tiny adjustment to the model's parameters), and then to immediately see the benefits of the model's goal-achieving-engine being smoothly repurposed towards those goals. Rather, turning a non-schemer into a schemer might require more substantive and holistic modification of the model's heuristics, tastes, patterns of attention, and so forth.

Relatedly: I think that "messy goal-directedness" complicates an assumption often employed in comparisons between schemers and other types of models: namely, the assumption that schemers will be able to perform approximately just as well as other sorts of models on all the tasks at stake in training (modulo, perhaps, a little bit extra cognition devoted to deciding-to-scheme – more below), even though they're doing so for instrumental reasons rather than out of any intrinsic interest in the task in question. This makes sense if you assume that all these models are aiming the same sort of "goal achieving engine" at a max-reward goal, for one reason or another. But what if that's not the right description?

Thus, as an extreme human example, consider two applicants for a job on a factory farm (each with similar intelligence, stamina, physical fitness, and other generic "capabilities"): Bob, a very compassionate and soft-hearted animal rights advocate, who is applying for the job in the hopes of making money to donate to animal-welfare causes; and Fred, who is totally indifferent to animal welfare (and to the money that the job offers), but who intrinsically loves performing all of the tasks that factory farming involves as efficiently as possible.^[5] Who is likely to perform the job more effectively? In humans, at least: probably Fred. Bob, after all, would likely push the animals into cages less violently, he would be distracted by feelings of horror and pain and disgust, etc; whereas Fred would perform the job with efficiency and gusto. In this case, Bob is worse at the task, despite his best efforts to perform it well, because his capabilities and his values are in some sense mixed together – his compassion, for example, is tied to both.

Of course, AIs could be different in tons of respects – and many humans are able to perform well in lots of contexts for instrumental reasons (witness, for example, spies working in governments they actively despise; or effective altruists working in hedge funds in order to donate the money to charity; or people, in general, who are just doing the job for the money). My point here is just that once you start allowing values, beliefs, heuristics, capabilities, patterns of attention, and so on to mix together, whether a schemer will be able to perform at the same level as e.g. a training-saint becomes a more open question.

Does scheming require a higher standard of goal-directedness?

As a final point of relevance of "messy goal-directedness" to scheming, I'll note that there's at least some case to be made that schemers require a "cleaner" and less "sphex-ish" form of goal-directedness than some of the other models I've discussed – e.g. training saints. Thus, for example, consider a model trained via RLHF to respond to various short-term user requests in a broadly helpful, honest, and harmless way. Many of these requests will require some sorts of flexible instrumental reasoning, yes – but it seems plausible to me that a model will also be getting a lot of its reward by learning the right sorts of local heuristics, tastes, patterns of attention, etc. That is, that a lot of its task-performance will be more "sphex-ish" than "explicit instrumental-reasoning-y." For this model to be a schemer, though (or at least, a pure schemer who values something quite different from the specified goal), we need to posit an additional role for explicit instrumental reasoning – namely, the model's current good behavior needs to be conditional on its calculation that such behavior conduces to its power (or the power of some other AIs); and when the time comes, the model needs to be ready to abandon such behavior and pursue its true goals. That is, whatever heuristics, local tastes, patterns of attention etc that give rise to the model's good behavior can't be fully hard-coded^[6] – they need to be at least partly subsumed by, and sensitive to, some other kind of instrumental reasoning. Whereas perhaps, for other models, this is less true.

That said, I've been assuming, and will continue to assume, that all the models we're considering are at least non-sphex-ish enough for the traditional assumptions of the alignment discourse to apply – in particular, that they will generalize off distribution in competent ways predicted by the goals we're attributing to them (e.g., HHH personal assistants will continue to try to be HHH, gold-coin-seekers will "go for the gold coins," reward-seekers will "go for reward," etc), and that they'll engage in the sort of instrumental reasoning required to get arguments about instrumental convergence off the ground. So in a sense, we're assuming a reasonably high standard of non-sphex-ishness from the get-go. I have some intuition that the standard at stake for schemers is still somewhat higher (perhaps because schemers seem like such paradigm consequentialists, whereas e.g. training saints seem like they might be able to be more deontological, virtue-ethical, etc?), but I won't press the point further here.

Of course, to the extent we don't assume that training is producing a very goal-directed model at all, hypothesizing that training has created a schemer may well involve hypothesizing a greater degree of goal-directedness than we would've needed to otherwise. That is, scheming will often require a higher standard of non-sphex-ishness than the training tasks themselves require. Thus, as an extreme example, consider AlphaStar, a model trained to play Starcraft. AlphaStar is plausibly goal-directed to some extent – its policy adapts flexibly to certain kinds of environmental diversity, in a manner that reliably conduces to winning-at-starcraft – but it's still quite sphex-ish and brittle in other ways. And to be clear: no one is saying that AlphaStar is a schemer. But in order to be a schemer (i.e., for AlphaStar's good performance in training to be explained by its executing a long-term instrumental strategy for power-seeking), and even modulo the need for situational awareness, AlphaStar would also need to be substantially more "goal-directed" than it currently is. That is, in this case, "somehow be such that you do this goal-directed-ish task" and "do this goal-directed-ish task because you've calculated that it conduces to your long-term power after training is complete" plausibly implicate different standards of goal-directedness. Perhaps, then, the same dynamic will apply to other, more flexible and advanced forms of task-performance (e.g., various forms of personal assistance, science, etc). Yes, those forms will require more in the way of general-purpose goal-directedness than AlphaStar displays. But perhaps they will require less than scheming requires, such that hypothesizing that the relevant model is a schemer will require hypothesizing a more substantive degree of goal-directedness than we would've needed to otherwise.

Indeed, my general sense is that one source of epistemic resistance to the hypothesis that SGD will select for schemers is the sense in which hypothesizing a schemer requires leaning on an attribution of goal-directedness in a way that greater agnosticism about why a model gets high reward need not. That is, prior to hypothesizing schemers, it's possible to shrug at a model's high-reward behavior and say something like:

"This model is a tangle of cognition such that it reliably gets high reward on the training distribution. Sure, you can say that it's 'goal-directed' if you'd like. I sometimes talk that way too. But all I mean is: it reliably gets high reward on the training distribution. Yes, in principle, it will also do things off of the training distribution. Maybe even: competent-seeming things. But I am not making predictions about what those competent-seeming things are, or saying that they will be pointed in similar-enough directions, across out-of-distribution-inputs, that it makes sense to ascribe to this model a coherent 'goal' or set of goals. It's a policy. It gets high reward on the training distribution. That's my line, and I'm sticking to it."

And against this sort of agnostic, atheoretical backdrop, positing that the model is probably getting reward specifically as part of a long-term strategy to avoid its goals being modified and then get power later can seem like a very extreme move in the direction of conjunctiveness and theory-heavy-ness. That is, we're not just attributing a goal to the model in some sort of hazy, who-knows-what-I-mean, does-it-even-matter sense. Rather, we're specifically going "inside the model's head" and attributing to it explicit long-term instrumental calculations driven by sophisticated representations of how to get what it wants.^[7]

However, I think the alignment discourse in general is doing this. In particular: I think the discourse about convergent instrumental sub-goals requires attributing goals to models in a sense that licenses talk about strategic instrumental reasoning of this kind. And to be clear: I'm not saying these attributions are appropriate. In fact, confusions about goal-directedness (and in particular: over-anchoring on psychologies that look like (a) expected utility maximizers and (b) total utilitarians) are one of my top candidates for the ways in which the discourse about alignment, as a whole, might be substantially misguided, especially with respect to advanced-but-still-opaque neural networks whose cognition we don't understand. That is, faced with a model that seems quite goal-directed on the training-distribution, and which is getting high reward, one shouldn't just ask where in some taxonomy of goal-directed models it falls – e.g., whether it's a training-saint, a mis-generalized non-training-gamer, a reward-on-the-episode-seeker, some mix of these, etc. One should also ask whether, in fact, such a taxonomy makes overly narrow assumptions about how to predict this model's behavior in general (for example: assuming that its out-of-distribution behavior will point in a coherent direction, that it will engage in instrumental reasoning in pursuit of the goals in question, etc), such that none of the model classes in the taxonomy are (even roughly) a good fit.

But as I noted in section 0.1, I here want to separate out the question of whether it makes sense to expect goal-directedness of this kind from the question of what sorts of goal-directed models are more or less plausible, conditional on getting the sort of goal-directedness that the alignment discourse tends to assume. Admittedly, to the extent the different model classes I'm considering require different sorts of goal-directedness, the line between these questions may blur a bit. But we should be clear about which question we're asking, and not confuse skepticism about goal-directedness in general for skepticism about schemers in particular.