AGI safety from first principles: Goals and Agency

The fundamental concern motivating the second species argument is that AIs will gain too much power over humans, and then use that power in ways we don’t endorse. Why might they end up with that power? I’ll distinguish three possibilities:

  1. AIs pursue power for the sake of achieving other goals; i.e. power is an instrumental goal for them.

  2. AIs pursue power for its own sake; i.e. power is a final goal for them.

  3. AIs gain power without aiming towards it; e.g. because humans gave it to them.

The first possibility has been the focus of most debate so far, and I’ll spend most of this section discussing it. The second hasn’t been explored in much depth, but in my opinion is still important; I’ll cover it briefly in this section and the next. Following Christiano, I’ll call agents which fall into either of these first two categories influence-seeking. The third possibility is largely outside the scope of this document, which focuses on dangers from the intentional behaviour of advanced AIs, although I’ll briefly touch on it here and in the last section.

The key idea behind the first possibility is Bostrom’s instrumental convergence thesis, which states that there are some instrumental goals whose attainment would increase the chances of an agent’s final goals being realised for a wide range of final goals and a wide range of situations. Examples of such instrumentally convergent goals include self-preservation, resource acquisition, technological development, and self-improvement, which are all useful for executing further large-scale plans. I think these examples provide a good characterisation of the type of power I’m talking about, which will serve in place of a more explicit definition.

However, the link from instrumentally convergent goals to dangerous influence-seeking is only applicable to agents which have final goals large-scale enough to benefit from these instrumental goals, and which identify and pursue those instrumental goals even when it leads to extreme outcomes (a set of traits which I’ll call goal-directed agency). It’s not yet clear that AGIs will be this type of agent, or have this type of goals. It seems very intuitive that they will because we all have experience of pursuing instrumentally convergent goals, for example by earning and saving money, and can imagine how much better we’d be at them if we were more intelligent. Yet since evolution has ingrained in us many useful short-term drives (in particular the drive towards power itself), it’s difficult to determine the extent to which human influence-seeking behaviour is caused by us reasoning about its instrumental usefulness towards larger-scale goals. Our conquest of the world didn’t require any humans to strategise over the timeframe of centuries, but merely for many individuals to expand their personal influence in a relatively limited way—by inventing a slightly better tool, or exploring slightly further afield.

Furthermore, we should take seriously the possibility that superintelligent AGIs might be even less focused than humans are on achieving large-scale goals. We can imagine them possessing final goals which don’t incentivise the pursuit of power, such as deontological goals, or small-scale goals. Or perhaps we’ll build “tool AIs” which obey our instructions very well without possessing goals of their own—in a similar way to how a calculator doesn’t “want” to answer arithmetic questions, but just does the calculations it’s given. In order to figure out which of these options is possible or likely, we need to better understand the nature of goals and goal-directed agency. That’s the focus of this section.

Frameworks for thinking about agency

To begin, it’s crucial to distinguish between the goals which an agent has been selected or designed to do well at (which I’ll call its design objectives), and the goals which an agent itself wants to achieve (which I’ll just call “the agent’s goals”).[1] For example, insects can contribute to complex hierarchical societies only because evolution gave them the instincts required to do so: to have “competence without comprehension”, in Dennett’s terminology. This term also describes current image classifiers and (probably) RL agents like AlphaStar and OpenAI Five: they can be competent at achieving their design objectives without understanding what those objectives are, or how their actions will help achieve them. If we create agents whose design objective is to accumulate power, but without the agent itself having the goal of doing so (e.g. an agent which plays the stock market very well without understanding how that impacts society) that would qualify as the third possibility outlined above.

By contrast, in this section I’m interested in what it means for an agent to have a goal of its own. Three existing frameworks which attempt to answer this question are Von Neumann and Morgenstern’s expected utility maximisation, Daniel Dennett’s intentional stance, and Hubinger et al’s mesa-optimisation. I don’t think any of them adequately characterises the type of goal-directed behaviour we want to understand, though. While we can prove elegant theoretical results about utility functions, they are such a broad formalism that practically any behaviour can be described as maximising some utility function. So this framework doesn’t constrain our expectations about powerful AGIs.[2] Meanwhile, Dennett argues that taking the intentional stance towards systems can be useful for making predictions about them—but this only works given prior knowledge about what goals they’re most likely to have. Predicting the behaviour of a trillion-parameter neural network is very different from applying the intentional stance to existing artifacts. And while we do have an intuitive understanding of complex human goals and how they translate to behaviour, the extent to which it’s reasonable to extend those beliefs about goal-directed cognition to artificial intelligences is the very question we need a theory of agency to answer. So while Dennett’s framework provides some valuable insights—in particular, that assigning agency to a system is a modelling choice which only applies at certain levels of abstraction—I think it fails to reduce agency to simpler and more tractable concepts.

Additionally, neither framework accounts for bounded rationality: the idea that systems can be “trying to” achieve a goal without taking the best actions to do so. In order to figure out the goals of boundedly rational systems, we’ll need to scrutinise the structure of their cognition, rather than treating them as black-box functions from inputs to outputs—in other words, using a “cognitive” definition of agency rather than “behavioural” definitions like the two I’ve discussed so far. Hubinger et al. use a cognitive definition in their paper on Risks from Learned Optimisation in Advanced ML systems: “a system is an optimizer if it is internally searching through a search space (consisting of possible outputs, policies, plans, strategies, or similar) looking for those elements that score high according to some objective function that is explicitly represented within the system”. I think this is a promising start, but it has some significant problems. In particular, the concept of “explicit representation” seems like a tricky one—what is explicitly represented within a human brain, if anything? And their definition doesn’t draw the important distinction between “local” optimisers such as gradient descent and goal-directed planners such as humans.

My own approach to thinking about agency tries to improve on the approaches above by being more specific about the cognition we expect goal-directed systems to do. Just as “being intelligent” involves applying a range of abilities (as discussed in the previous section), “being goal-directed” involves a system applying some specific additional abilities:

  1. Self-awareness: it understands that it’s a part of the world, and that its behaviour impacts the world;

  2. Planning: it considers a wide range of possible sequences of behaviours (let’s call them “plans”), including long plans;

  3. Consequentialism: it decides which of those plans is best by considering the value of the outcomes that they produce;

  4. Scale: its choice is sensitive to the effects of plans over large distances and long time horizons;

  5. Coherence: it is internally unified towards implementing the single plan it judges to be best;

  6. Flexibility: it is able to adapt its plans flexibly as circumstances change, rather than just continuing the same patterns of behaviour.

Note that none of these traits should be interpreted as binary; rather, each one defines a different spectrum of possibilities. I’m also not claiming that the combination of these six dimensions is a precise or complete characterisation of agency; merely that it’s a good starting point, and the right type of way to analyse agency. For instance, it highlights that agency requires a combination of different abilities—and as a corollary, that there are many different ways to be less than maximally agentic. AIs which score very highly on some of these dimensions might score very low on others. Considering each trait in turn, and what lacking it might look like:

  1. Self-awareness: for humans, intelligence seems intrinsically linked to a first-person perspective. But an AGI trained on abstract third-person data might develop a highly sophisticated world-model that just doesn’t include itself or its outputs. A sufficiently advanced language or physics model might fit into this category.

  2. Planning: highly intelligent agents will by default be able to make extensive and sophisticated plans. But in practice, like humans, they may not always apply this ability. Perhaps, for instance, an agent is only trained to consider restricted types of plans. Myopic training attempts to implement such agents; more generally, an agent could have limits on the actions it considers. For example, a question-answering system might only consider plans of the form “first figure out subproblem 1, then figure out subproblem 2, then...”.

  3. Consequentialism: the usual use of this term in philosophy describes agents which believe that the moral value of their actions depends only on those actions’ consequences; here I’m using it in a more general way, to describe agents whose subjective preferences about actions depend mainly on those actions’ consequences. It seems natural to expect that agents trained on a reward function determined by the state of the world would be consequentialists. But note that humans are far from fully consequentialist, since we often obey deontological constraints or constraints on the types of reasoning we endorse.

  4. Scale: agents which only care about small-scale events may ignore the long-term effects of their actions. Since agents are always trained in small-scale environments, developing large-scale goals requires generalisation (in ways that I discuss below).

  5. Coherence: humans lack this trait when we’re internally conflicted—for example, when our system 1 and system 2 goals differ—or when our goals change a lot over time. While our internal conflicts might just be an artefact of our evolutionary history, we can’t rule out individual AGIs developing modularity which might lead to comparable problems. However, it’s most natural to think of this trait in the context of a collective, where the individual members could have more or less similar goals, and could be coordinated to a greater or lesser extent.

  6. Flexibility: an inflexible agent might arise in an environment in which coming up with one initial plan is usually sufficient, or else where there are tradeoffs between making plans and executing them. Such an agent might display sphexish behaviour. Another interesting example might be a multi-agent system in which many AIs contribute to developing plans—such that a single agent is able to execute a given plan, but not able to rethink it very well.

A question-answering system (aka an oracle) could be implemented by an agent lacking either planning or consequentialism. For AIs which act in the real world I think the scale of their goals is a crucial trait to explore, and I’ll do so later in this section. We can also evaluate other systems on these criteria. A calculator probably doesn’t have any of them. Software that’s a little more complicated, like a GPS navigator, should probably be considered consequentialist to a limited extent (because it reroutes people based on how congested traffic is), and perhaps has some of the other traits too, but only slightly. Most animals are self-aware, consequentialist and coherent to various degrees. The traditional conception of AGI has all of the traits above, which would make it capable of pursuing influence-seeking strategies for instrumental reasons. However, note that goal-directedness is not the only factor which determines whether an AI is influence-seeking: the content of its goals also matter. A highly agentic AI which has the goal of remaining subordinate to humans might never take influence-seeking actions. And as previously mentioned, an AI might be influence-seeking because it has the final goal of gaining power, even without possessing many of the traits above. I’ll discuss ways to influence the content of an agent’s goals in the next section, on Alignment.

The likelihood of developing highly agentic AGI

How likely is it that, in developing an AGI, we produce a system with all of the six traits I identified above? One approach to answering this question involves predicting which types of model architecture and learning algorithms will be used—for example, will they be model-free or model-based? To my mind, this line of thinking is not abstract enough, because we simply don’t know enough about how cognition and learning work to map them onto high-level design choices. If we train AGI in a model-free way, I predict it will end up planning using an implicit model anyway. If we train a model-based AGI, I predict its model will be so abstract and hierarchical that looking at its architecture will tell us very little about the actual cognition going on.

At a higher level of abstraction, I think that it’ll be easier for AIs to acquire the components of agency listed above if they’re also very intelligent. However, the extent to which our most advanced AIs are agentic will depend on what type of training regime is used to produce them. For example, our best language models already generalise well enough from their training data that they can answer a wide range of questions. I can imagine them becoming more and more competent via unsupervised and supervised training, until they are able to answer questions which no human knows the answer to, but still without possessing any of the properties listed above. A relevant analogy might be to the human visual system, which does very useful cognition, but which is not very “goal-directed” in its own right.

My underlying argument is that agency is not just an emergent property of highly intelligent systems, but rather a set of capabilities which need to be developed during training, and which won’t arise without selection for it. One piece of supporting evidence is Moravec’s paradox: the observation that the cognitive skills which seem most complex to humans are often the easiest for AIs, and vice versa. In particular, Moravec’s paradox predicts that building AIs which do complex intellectual work like scientific research might actually be easier than building AIs which share more deeply ingrained features of human cognition like goals and desires. To us, understanding the world and changing the world seem very closely linked, because our ancestors were selected for their ability to act in the world to improve their situations. But if this intuition is flawed, then even reinforcement learners may not develop all the aspects of goal-directedness described above if they’re primarily trained to answer questions.

However, there are also arguments that it will be difficult to train AIs to do intellectual work without them also developing goal-directed agency. In the case of humans, it was the need to interact with an open-ended environment to achieve our goals that pushed us to develop our sophisticated general intelligence. The central example of an analogous approach to AGI is training a reinforcement learning agent in a complex simulated 3D environment (or perhaps via extended conversations in a language-only setting). In such environments, agents which strategise about the effects of their actions over long time horizons will generally be able to do better. This implies that our AIs will be subject to optimisation pressure towards becoming more agentic (by my criteria above) will do better. We might expect an AGI to be even more agentic if it’s trained, not just in a complex environment, but in a complex competitive multi-agent environment. Agents trained in this way will need to be very good at flexibly adapting plans in the face of adversarial behaviour; and they’ll benefit from considering a wider range of plans over a longer timescale than any competitor. On the other hand, it seems very difficult to predict the overall effect of interactions between many agents—in humans, for example, it led to the development of (sometimes non-consequentialist) altruism.

It’s currently very uncertain which training regimes will work best to produce AGIs. But if there are several viable ones, we should expect economic pressures to push researchers towards prioritising those which produce the most agentic AIs, because they will be the most useful (assuming that alignment problems don’t become serious until we’re close to AGI). In general, the broader the task an AI is used for, the more valuable it is for that AI to reason about how to achieve its assigned goal in ways that we may not have specifically trained it to do. For example, a question-answering system with the goal of helping its users understand the world might be much more useful than one that’s competent at its design objective of answering questions accurately, but isn’t goal-directed in its own right. Overall, however, I think most safety researchers would argue that we should prioritise research directions which produce less agentic AGIs, and then use the resulting AGIs to help us align later more agentic AGIs. There’s also been some work on directly making AGIs less agentic (such as quantilisation), although this has in general been held back by a lack of clarity around these concepts.

I’ve already discussed recursive improvement in the previous section, but one further point which is useful to highlight here: since being more agentic makes an agent more capable of achieving its goals, agents which are capable of modifying themselves will have incentives to make themselves more agentic too (as humans already try to do, albeit in limited ways).[3] So we should consider this to be one type of recursive improvement, to which many of the considerations discussed in the previous section also apply.

Goals as generalised concepts

I should note that I don’t expect our training tasks to replicate the scale or duration of all the tasks we care about in the real world. So AGIs won’t be directly selected to have very large-scale or long-term goals. Yet it’s likely that the goals they learn in their training environments will generalise to larger scales, just as humans developed large-scale goals from evolving in a relatively limited ancestral environment. In modern society, people often spend their whole lives trying to significantly influence the entire world—via science, business, politics, and many other channels. And some people aspire to have a worldwide impact over the timeframe of centuries, millennia or longer, even though there was never significant evolutionary selection in favour of humans who cared about what happened in several centuries’ time, or paid attention to events on the other side of the world. This gives us reason to be concerned that even AGIs which aren’t explicitly trained to pursue ambitious large-scale goals might do so anyway. I also expect researchers to actively aim towards achieving this type of generalisation to longer time horizons in AIs, because some important applications rely on it. For long-term tasks like being a CEO, AGIs will need the capability and motivation to choose between possible actions based on their worldwide consequences on the timeframe of years or decades.

Can we be more specific about what it looks like for goals to generalise to much larger scales? Given the problems with the expected utility maximisation framework I identified earlier, it doesn’t seem useful to think of goals as utility functions over states of the world. Rather, an agent’s goals can be formulated in terms of whatever concepts it possesses—regardless of whether those concepts refer to its own thought processes, deontological rules, or outcomes in the external world.[4] And insofar as an agent’s concepts flexibly adjust and generalise to new circumstances, the goals which refer to them will do the same. It’s difficult and speculative to try to describe how such generalisation may occur, but broadly speaking, we should expect that intelligent agents are able to abstract away the differences between objects or situations that have high-level similarities. For example, after being trained in a simulation, an agent might transfer its attitudes towards objects and situations in the simulation to their counterparts in the (much larger) real world.[5] Alternatively, the generalisation could be in the framing of the goal: an agent which has always been rewarded for accumulating resources in its training environment might interalise the goal of “amassing as many resources as possible”. Similarly, agents which are trained adversarially in a small-scale domain might develop a goal of outcompeting each other which persists even when they’re both operating at a very large scale.

From this perspective, to predict an agent’s behaviour, we will need to consider what concepts it will possess, how those will generalise, and how the agent will reason about them. I’m aware that this appears to be an intractably difficult task—even human-level reasoning can lead to extreme and unpredictable conclusions (as the history of philosophy shows). However, I hope that we can instill lower-level mindsets or values into AGIs which guide their high-level reasoning in safe directions. I’ll discuss some approaches to doing so in the next section, on Alignment.

Groups and agency

After discussing collective AGIs in the previous section, it seems important to examine whether the framework I’ve proposed for understanding agency can apply to a group of agents as well. I think it can: there’s no reason that the traits I described above need to be instantiated within a single neural network. However, the relationship between the goal-directedness of a collective AGI and the goal-directedness of its individual members may not be straightforward, since it depends on the internal interactions between its members.

One of the key variables is how much (and what types of) experience those members have of interacting with each other during training. If they have been trained primarily to cooperate, that makes it more likely that the resulting collective AGI is a goal-directed agent, even if none of the individual members is highly agentic. But there are good reasons to expect that the training process will involve some competition between members, which would undermine their coherence as a group. Internal competition might also increase short-term influence-seeking behaviour, since each member will have learned to pursue influence in order to outcompete the others. As a particularly salient example, humanity managed to take over the world over a period of millennia not via a unified plan to do so, but rather as a result of many individuals trying to expand their short-term influence.

It’s also possible that the members of a collective AGI have not been trained to interact with each other at all, in which case cooperation between them would depend entirely on their ability to generalise from their existing skills. It’s difficult to imagine this case, because human brains are so well-adapted for group interactions. But insofar as humans and aligned AGIs hold a disproportionate share of power over the world, there is a natural incentive for AGIs pursuing misaligned goals to coordinate with each other to increase their influence at our expense.[6] Whether they succeed in doing so will depend on what sort of coordination mechanisms they are able to design.

A second factor is how much specialisation there is within the collective AGI. In the case where it consists only of copies of the same agent, we should expect that the copies understand each other very well, and share goals to a large extent. If so, we might be able to make predictions about the goal-directedness of the entire group merely by examining the original agent. But another case worth considering is a collective consisting of a range of agents with different skills. With this type of specialisation, the collective as a whole could be much more agentic than any individual agent within it, which might make it easier to deploy subsets of the collective safely.

  1. AI systems which learn to pursue goals are also known as mesa-optimisers, as coined in Hubinger et al’s paper Risks from Learned Optimisation in Advanced Machine Learning Systems. ↩︎

  2. Related arguments exist which attempt to do so. For example, Eliezer Yudkowsky argues here that, “while corrigibility probably has a core which is of lower algorithmic complexity than all of human value, this core is liable to be very hard to find or reproduce by supervised learning of human-labeled data, because deference is an unusually anti-natural shape for cognition, in a way that a simple utility function would not be an anti-natural shape for cognition.” Note, however, that this argument relies on the intuitive distinction between natural and anti-natural shapes for cognition. This is precisely what I think we need to understand to build safe AGI—but there has been little explicit investigation of it so far. ↩︎

  3. I believe this idea comes from Anna Salamon; unfortunately I’ve been unable to track down the exact source. ↩︎

  4. For example, when people want to be “cooperative” or “moral”, they’re often not just thinking about results, but rather the types of actions they should take, or the types of decision procedures they should use to generate those actions. An additional complication is that humans don’t have full introspective access to all our concepts—so we need to also consider unconscious concepts. ↩︎

  5. Consider if this happened to you, and you were pulled “out of the simulation” into a real world which is quite similar to what you’d already experienced. By default you would likely still want to eat good food, have fulfilling relationships, and so on, despite the radical ontological shift you just underwent. ↩︎

  6. In addition to the prima facie argument that intelligence increases coordination ability, it is likely that AGIs will have access to commitment devices not available to humans by virtue of being digital. For example, they could send potential allies a copy of themselves for inspection, to increase confidence in their trustworthiness. However, there are also human commitment devices that AGIs will have less access to—for example, putting ourselves in physical danger as an honest signal. And it’s possible that the relative difficulty of lying versus detecting lying shifts in favour of the former for more intelligent agents. ↩︎