Circumventing interpretability: How to defeat mind-readers

TL;DR: Unaligned AI will have a convergent instrumental incentive to make its thoughts difficult for us to interpret. In this article, I discuss many ways that a capable AI might circumvent scalable interpretability methods and suggest a framework for thinking about these risks.

Summary of key points

The framework: Interpretation as information flow

I categorize potential interpretability circumvention methods on three levels:

  1. Interpretability methods consist of interpretability tools and the human interpreter. To circumvent interpretation an AI can prevent information about its thoughts from flowing through either the tool or the interpreter. (See main text for figure)

  2. Within the tool or the human interpreter, there are several information processing steps that an AI might exploit. (See main text for figure):

    • Exposure evasion: Being where your adversary isn’t looking

    • Perception evasion: Being hard to detect even when exposed to your adversary

    • Recognition evasion: Being hard to recognize/​understand even though you’ve been perceived

    • Dazzling: Looking like something that confuses or incapacitates your adversary

  3. Another useful way to categorize interpretability circumvention methods is whether they are passive or active. Most methods that can be passive can also be active, but other methods are strictly active.

    • Passive interpretability circumvention occurs during the normal course of learning without the AI intending to make itself less interpretable. (e.g. polysemanticity).

    • Active interpretability circumvention is where the AI intentionally takes actions in order to circumvent our interpretability methods. (e.g. an AI actively deleting its own representations to hide deception)

The list of potential interpretability circumvention methods with a short summary of each

  • 1. Exploiting weaknesses of the interpretability tool

    • 1.1. Evading exposure to the interpretability tool

      • 1.1.1. Thinking in the gaps of our tools: Our interpretability tools might fail to examine everything about an AI’s thoughts (e.g. it might not look at nonlinear interactions between neurons, only their linear relations). An AI might use those blindspots for dangerous thought. (Passive or active)

      • 1.1.2. Externalizing representations: The environment can be used for some parts of an AI’s computation. Actively: The agent can write messages to itself or initiate programs. Passively: The AI could use the environment as a reservoir in a reservoir computer. (Passive or active)

      • 1.1.3. Environmental artifacts that corrupt its future self: A misaligned AI may leave behind artifacts that serve as adversarial examples or poisoned data that corrupt its future self. (Strictly active)

      • 1.1.4. Deleting (and relearning) dangerous representations: When not under surveillance, an AI might be able to delete dangerous representations so that it looks safe when surveillance resumes. It might also be able to schedule when it relearns those representations. (Strictly active)

      • 1.1.5. Weights poisoning: Instead of providing manipulated inputs, the AI might manipulate its own weights so that a small subset of normal inputs behave like adversarial examples. (Strictly active)

      • 1.1.6. Pre-committing to following potentially dangerous instructions: Hubinger: ”… if the model has the ability to read and write to some sort of state … it could write out a pre-commitment to follow the hard-coded policy and then just have a policy of always following any pre-commitments it finds such that it never thinks about the deception that led it to that policy in the first place.” (Strictly active)

      • 1.1.7. Hard-coded neural shortcuts: Hubinger again: “… it could hack its own gradient to produce a new model which will still exhibit the same deceptive behavior, but without doing any of the detectable deceptive cognition. … not because it’s trying to act deceptive, but just because it’s a hard-coded part of its algorithm” (Strictly active)

      • 1.1.8. Following cached thoughts: Nanda: “[An AI might develop] the instinct to follow cached thoughts without thinking about them, and in occasional moments of deception to generate a lot of cached thoughts that will lead to misaligned actions, but with no deceptive cognition in the moment.” (Strictly active)

      • 1.1.9. Instrumental incapability: A misaligned AI that’s able to modify its own parameters could damage capabilities that serve some goals (e.g. aligned goals) more than its misaligned goals. Generalizes instrumental clumsiness. (Strictly active)

      • 1.1.10. Instrumental clumsiness: An AI might hack its future self so that it is more likely to make certain kinds of mistakes while leaving other capabilities intact. (Strictly active)

    • 1.2. Evading perception by the interpretability tool

      • 1.2.1. Hiding in noisy subspaces: We might ignore noisy regions or subspaces of neural activation space may be very noisy. But an AI might be able to do some dangerous thinking in those spaces. Generalizes ‘Whispering dangerous thoughts’. (Passive or active)

      • 1.2.2. Whispering dangerous thoughts: An AI might make its dangerous thoughts very ‘quiet’ relative to other thoughts by using small norm activation vectors. (Passive or active)

    • 1.3. Evading recognition by the interpretability tool

      • 1.3.1. Antiparsimony: If it costs more to interpret an AI’s thought than it costs to think it, an AI might be able to produce a lot of representations such that it saturates the interpreter’s computation before it saturates its own. (Passive or active)

      • 1.3.2. Encryption: Suppose we have an AI with a memory state. The AI might use very long chains of seemingly meaningless neural activations to construct a single thought. (Passive or active)

      • 1.3.3. Interpretability illusions: By manipulating the dataset shown to the tool, an AI can control its interpretations. (Strictly active)

    • 1.4. Attacking/​dazzling the interpretability tool

      • 1.4.1. Adversarial attack against networks in the interpretation tool: The AI could adversarially attack networks in the tool to manipulate the tool’s outputs. (Strictly active)

  • 2. Exploiting weaknesses of the human interpreter

    • 2.1. Evading exposure to the human

      • 2.1.1. None identified

    • 2.2. Evading perception by the human interpreter

      • 2.2.1. None identified

    • 2.3. Evading recognition by the human interpreter

      • 2.3.1. Alien representations: A representation whose meaning lies in a strange place in our ontology. Generalizes ‘Polysemanticity’ and ‘Representational kludging’. (Passive or active)

      • 2.3.2. Polysemanticity: Where directions in neural statespace mean multiple things. (Passive or active)

      • 2.3.3. Representational kludging: Creating representations with dangerous properties by combining innocuous precursor representations in an awkward messy way. (Passive or active)

      • 2.3.4. Unknown representations: Some thoughts aren’t immediately possible for us to translate because we don’t yet have the concepts for them. But we’re capable of learning them. (Passive or active)

      • 2.3.5. Unknowable representations: Some concepts are impossible for humans to understand due to the computational limitations of our brain or some other reasons. (Passive or active)

    • 2.4. Attacking/​dazzling the human interpreter

      • 2.4.1. Manipulating, mind-hacking, or incapacitating the interpreter: An AI might be able to manipulate, directly control, or incapacitate humans using specific kinds of outputs. (Strictly active)

Some high-level takeaways

  • A lot of circumvention methods rely on making interpretation more expensive (either in terms of researcher time or compute).

  • We can draw many analogies between mind-reading/​interpretability circumvention and cybersecurity. These ideas may perhaps be called ‘cognitosecurity’.

  • A lot of circumvention methods rely on gradient hacking, which requires advanced capabilities. These methods might therefore be less likely to happen because a misaligned AI that’s capable of gradient hacking probably has better ways to escape.

  • Using interpretability tools in the loss function seems likely to be very bad (in most circumstances) for interpreting dangerous thought.

Full article:

Acknowledgements:

I’m grateful to David Lindner, Evan R. Murphy, Alex Lintz, Sid Black, Kyle McDonnell, Laria Reynolds, Adam Shimi, and Daniel Braun whose comments greatly improved earlier drafts of this article. The article’s weaknesses are mine, but many of its strengths are due to their contributions.

Additionally, this article benefited from the prior work of many authors, but especially: Evan Hubinger, Peter Barnett, Adam Shimi, Neel Nanda, Evan R. Murphy, Eliezer Yudkowsky, Chris Olah. I collected several of the potential circumvention methods from their work.

This work was carried out while at Conjecture.

Introduction

There’s been a lot of excitement about interpretability recently. It’s a component in many different alignment proposals.

But if we’re interpreting an unaligned AI, it has a convergent instrumental incentive[1] to circumvent our interpretability methods. There’s been some sporadic past discussion about how an AI might achieve this (e.g. 1, 2, 3, 4, 5, 6, 7). But, to my knowledge, the problem hasn’t been addressed systematically, at least not publicly.

If we continue to make large bets on interpretability, then the lack of attention on potential circumvention methods seems bad because we might end up with a false sense of security. Our interpretability methods might say “Nope, no naughty thoughts here!” when in fact the AI has simply found a way to think dangerous thoughts undetected.

We can think of this problem as the focus of an interpretability research agenda that is related to but distinct from the well-known agenda that asks “Can we understand what today’s neural networks are doing? And how can we scale our methods of understanding them?”. I think this short term agenda is exceptionally important (it’s the subject of my day-job). But we should probably also try to anticipate future issues by building a longer term interpretability research agenda. This article helps to address that longer term agenda and asks “Assuming our short term research agenda at least partly succeeds, how might a capable AI circumvent our interpretability methods?”.

Here I present a definitely-incomplete list of potential interpretability circumvention methods that I and other sources came up with. Throughout, I assume we’re concerned with interpreting the representations[2] of a capable (roughly human-level or above) deep learning-based AI. I also suggest a framework for thinking about these circumvention methods, which I hope will help to identify omitted circumvention methods before a capable AI does.

The framework: Interpretation as information flow

We can think of interpretation as information flow (figure 1). Here we’ll use the following basic model of information flow in interpretability:

  • We have an AI whose thoughts are being interpreted by an interpretability method.

  • An interpretability method is the process of translating an AI’s representations into whatever the corresponding representations are in humans.

  • To do this, interpretability methods contain interpretability tools, which is the analysis machinery that reads the AI’s representations and converts them into something legible to humans.

  • Human interpreters read the outputs of the interpretability tools and attempt to understand them, thus converting them into human representations.

Figure 1. A simple model of information flow in interpretability.

Throughout the article, I assume we’ve figured out ways to interpret our models at scale, which probably means that our interpretability tools are automating a lot of the work. Nevertheless, all the potential circumvention methods I talk about apply to worlds where we still haven’t managed to scale interpretability, where the tools would be relatively less important.

An even more detailed model of information flow (figure 2) is useful for categorizing interpretability circumvention methods, as we’ll see:

  • Information flows within both the interpretability tool and the human. In both,

  • What is to be interpreted must first be exposed to the tool or the human.

  • What is exposed must then be perceived or detected.

  • What is perceived is then further analysed so that it can be recognized or understood, which then determines an appropriate output by the tool or action by the human.

Figure 2. A more detailed model of information flow which is applicable to both the interpretability tool and to the human. Information flow begins with exposure of the information to the observer, then perception by the observer, then recognition & and decision making, followed by output.

How I categorize interpretability circumvention methods

The ways to categorize interpretability circumvention methods that I’ve found most useful (and the way this article is structured) are:

A. Exploiting weaknesses of the tool vs. weaknesses of the human interpreter

Interpretability circumvention methods can either circumvent the interpretability tool or the human interpreter.

B. Categories of interpretability circumvention found in nature

Evolution has found many different ways that make it difficult for adversaries to interpret the appearance of an organism. I put them into 4 groups (figure 3), which broadly correspond to the detailed model of information flow described above:

  • Exposure evasion:

    • Being where your adversary isn’t looking

    • E.g. An animal that lives underground or up a tree, away from surface-dwelling predators

  • Perception evasion:

    • Being hard to detect even when exposed to your adversary

    • E.g. An animal that only comes out when it’s dark; color camouflage to look the same as the color of the background

  • Recognition evasion:

    • Being hard to recognize/​understand even though you’ve been perceived

    • E.g. Biomimicry, where an organism looks like another kind of object entirely, for instance: an Orchid mantis which looks like the flower; the harmless Hornet Moth mimicking the colors of a wasp.

  • Dazzling:

    • Looking like something that confuses or incapacitates your adversary

    • E.g. Living in large groups such that it’s hard for a predator to focus on any particular individual; a zebra’s stripes.

Figure 3. Interpretability circumvention methods exploit different steps in the information flow diagram and correspond to forms of interpretability circumvention found in nature. Circumvention methods may seek to evade exposure, perception, or recognition. They may also exploit vulnerabilities in the observer’s information processing system, a process called dazzling.

C. Passive vs. active methods

Information flow might be inhibited passively by circumvention methods that arise during the normal course of learning. One example is ‘polysemanticity’, which is a natural phenomenon that presents a challenge to interpretability. Passive circumvention methods might also arise if they’re directly optimized for. This might happen if we use interpretability tools in the loss function, which is where we identify unsafe representations in an AI and then use the loss function to penalize the AI for containing those representations. At first glance, this seems like a potentially good way to make an AI safe. But it risks pushing unsafe representations into the blind spots of our interpretability tools.

Information flow might be inhibited actively through an agent’s actions. The primary way this could happen is gradient hacking, but it’s not the only kind of action an AI might take to conceal misaligned thoughts. Of course, active circumvention methods require that interpreters either can’t see or aren’t looking at the thoughts that generate those actions.

Most potential circumvention methods that can be passive can also be active. But some methods can only be active. It’s usually obvious when either is the case, so I only highlight this categorization method when it’s not obvious.

This categorization is mostly useful for identifying how and when during learning the method is likely to arise. Strictly active methods generally require advanced capabilities (such as gradient hacking), whereas passive methods don’t. I tend to think this makes strictly active methods less likely than methods that may be passive or active.

Before I introduce the individual circumvention methods in detail (see next section), the table below puts individual circumvention methods into categories according to whether they may be ‘passive or active’ or whether they are ‘strictly active’.

Table 1. Categorizing interpretability circumvention methods as ‘Passive or active’ or ‘Strictly active’

This article is structured according to the above three categorization schemes, primarily the first two. I’ll discuss circumvention methods that exploit weaknesses either of the tool or of the human. For each, I’ll look at methods of exposure evasion, perception evasion, recognition evasion, and dazzling. Throughout, I’ll highlight whether a method can arise passively or if it’s strictly an active method, when the distinction matters.

I’m not confident this categorization scheme is optimal or even totally consistent, but I’ve found it helpful. Occasionally, a circumvention method in the list is in fact a generalization of another method in the list, yet I present them separately if the special case is particularly interesting or likely to arise.

Interpretability circumvention methods

Table 2. An overview of the list of potential interpretability circumvention methods and their categories.

1. Exploiting weaknesses of the interpretability tool

1.1. Evading exposure to the interpretability tool

1.1.1. Thinking in the gaps of our tools

When interpreting a capable AI, our interpretability tools might not examine every possible representational detail used by the AI for its cognition. An AI might thus be able to ‘think within the gaps’ of our interpretability tools.

A few examples make this clearer:

  1. Today’s mechanistic interpretability methods don’t let us easily interpret representations that use nonlinearities (Voss et al. 2021; Elhage et al. 2021). If this blindspot persists, a model might be able to use nonlinearities to hide dangerous thoughts.

  2. We might have interpretability tools that look only for an expected set of concepts in a model’s activations. But the AI may find dangerous concepts that don’t belong to that set because we hadn’t yet thought of them.

  3. Hiding in summary statistics: If an interpretability tool uses the mean value when deciding whether some quantity in the AI is of concern, then an AI might be able to use high variance and highly skewed data to hide concerning statistics. E.g. It could hide the fact it values most paperclip-maximizing scenarios a lot by offsetting them with a small number of very low utilities in a small number of such scenarios. The mean would look normal, but if we looked at the variance we’d see a more concerning picture.

  4. We might expect language to be the kind of representational space in which an AI can think dangerous thoughts. But what about in representational spaces we don’t expect, like generic dynamical systems modeling modules? We know dynamical systems (such as an RNN) can be Turing complete, so maybe we should have expected dangerous representations in there. But what about in regions that are even less obviously capable of computing dangerous thoughts, like in representations constructed from syntactic patterns or punctuation? That’s a kind of representational space that might be expressive enough to encode dangerous thoughts, but not one of the obvious first places to look[3].

  5. A fictional example: In HPMOR Harry tries to protect against mind-readers (occlumency) by essentially hosting two personas at once. An AI could learn a whole 2nd set of representations or a whole second internal agent. Whether this would work, again, depends on the details of the interpretability tool. If, for instance, the interpretability method identifies ‘the deception representation’ within the primary representation set and then stops looking for deception-like representations, it might fail to find the deception representation in the secondary set.

This class of methods is likely to emerge passively (though may also emerge actively), and many instances of it would be encouraged by using interetability tools in the loss function.

1.1.2. Externalizing representations

If you keep forgetting to reply to an email and therefore set an electronic reminder, the reminder is a functionally relevant variable in the “respond to email” behavior. Indeed, if you’re forgetful enough, it’s an essential variable in that behavior. This is the basic idea behind the ‘extended mind thesis’ from Andy Clark & David Chalmers: There is no hard barrier between the mind and the environment; cognition is often a dance between internal and external variables and can’t be reduced to only one set. If representations are ‘cognitively functional variables’, then representations may be either internal or external. Using the environment as a memory system like this can be called autostigmergy (Deverett et al. 2019).

An AI might be able to passively or actively externalize representations through the effects of its actions. If our interpretability methods ignore those parts of the environment that the AI is using to think, then we may miss some dangerous cognition.

Are externalized representations likely to be a problem? Let’s consider active and passive cases separately because, unlike for most other interpretability circumvention methods, they’re remarkably different for ‘Externalizing representations’.

Active

Actively externalizing representations is where the AI intentionally uses the environment as a memory system. An AI probably first needs to represent the cognitive variables internally before it can encode them into the environment. If we’re not constantly surveilling the AI, this is easily possible. This includes the obvious methods like

  • Writing a message to its future self, potentially steganographically. For example, an AI that can manipulate objects in a room may be able to encode messages for its future self by subtly arranging the objects in a room in a particular way. Or a virtual assistant agent might be able to encode messages for itself in a web search history or in different filename patterns.

  • Running a computer program to decide whether it’s time to defect or not.

An edge case is the act of creating another agent in the environment. But we’d then be dealing with a whole new copy of the agent that isn’t under surveillance, and we’ve probably already lost if this happens, so I don’t consider it further here.

Passive

It’s less clear whether it’s possible to passively externalize representations in order to circumvent interpretability methods. The representations required for dangerous thought have a complicated structure. Learning such structure externally is a serious challenge because the AI doesn’t get to learn it using backprop through the environment or explicit planning.

Nevertheless, one way an AI might achieve it is to use the environment as a reservoir computer. For those who would benefit from an explanation of reservoir computing, it’s easiest to describe in the supervised learning setting: To train a reservoir computer, you pass an input vector to a reservoir, which is a fixed dynamical system (such as an RNN with fixed weights). But you only learn the readout weights, which simply read the prediction from the reservoir variables. Here, the readout weights are usually linear, so they’re not doing anything complicated; it’s the variables in the reservoir that are doing the complicated nonlinear computation. Thus, the dynamics of the reservoir variables (rather than the learned parameters) can serve as representations that solve a task.

Going beyond supervised learning, reservoir computing can also work in the RL-setting: The reservoir is the agent+environment dynamics from timestep to ; an input observation is provided to the agent at timestep ; and the readout weights are the weights of the agent at timestep . If the environment part of the reservoir contains most of the dangerous representations, we might miss some dangerous thoughts.

Deverett et al. (2019) provide an empirical example of a simple feedforward model-free agent that learns to use the agent-environment system as a reservoir computer (though they don’t frame it as such). The memoryless agent learns to use its location in the environment to encode a time interval. It should be stressed that it does this without backprop through time or any knowledge of the past. But its location varies due to its initially random patterns of actions, yet it can use those patterns to encode a time interval.

The Deverett et al. example is a very simple representation. Maybe reservoir computing like this doesn’t scale well enough to learn the complex representations that are required for dangerous cognition. But the inexpressivity of reservoir computing is partly due to the simplicity of the linear readout weights. With a powerful AI serving as a readout network, reservoir computing might be significantly more expressive and reliable. It might be an interesting research project to identify how expressive reservoir computing can be if we allow the readout network to be strongly nonlinear.

Note: The rest of the methods discussed in this section (‘1.1. Evading exposure to the interpretability tool’) are strictly active methods.

1.1.3. Environmental artifacts that corrupt its future self

A misaligned AI may leave behind artifacts that corrupt its (potentially aligned) future self so that it pursues the goals of its misaligned past self instead.

The artifacts may be adversarial examples that cause it to act in unexpected and dangerous ways. Alternatively, the agent could leave behind poisoned data so that its future self will learn dangerous representations; this would be an instance of gradient hacking.

This method might reasonably be considered a special case of ‘Externalizing representations’. The distinction is somewhat arbitrary and depends on what qualifies as a representation. Does ‘an attack that corrupts its future self’ qualify as representation? It’s not clear and is somewhat arbitrary, but I’m inclined to put them in separate categories.

1.1.4. Deleting (and relearning) dangerous representations

If an AI can gradient hack, it might be able to delete dangerous representations so that it looks safe when surveillance resumes.

Thus the volume under the suspicious thought-surface interpretation time curve is minimized, since the suspicious thought-surface is 0 everywhere that interpretation > 0.

Of course, deleting thoughts only avoids detection; it doesn’t permit dangerous thought to continue. But if an AI has learned dangerous representations and deletes them, it’s probable that it will learn those representations again, assuming its data distribution doesn’t change too much. An AI that can gradient hack might be able to schedule when it periodically re-learns the dangerous representations so that it always looks safe when surveillance resumes.

This method requires the agent to have good knowledge of when it is likely to be interpreted. If we interpret the AI every K batches, an AI might be able to use time information between batches to notice that the interval between some batches (validation/​interpretation periods) are much longer than usual. If it’s too expensive to surveil the AI constantly, there might be a cheaper monitoring scheme: We could interpret an agent at random intervals in order to create a ‘panopticon’ - where the agent is forced to behave/​think safely because it doesn’t know when it is safe for it to think thoughts about defection.

1.1.5. Weights poisoning

Instead of providing manipulated inputs, the AI might manipulate its own weights so that a small subset of normal inputs behave like adversarial examples, thus causing dangerous learning or behavior. This is essentially ‘weights poisoning’ (Kurita et al. 2020) where the AI is poisoning its own weights. It’s the same idea (as far as I can tell) behind ‘Trojan models’ (Rajabi et al. 2022). Planting such backdoors in neural networks can be practically impossible to detect (Goldwasser et al. 2022).

This is kind of like fictional ‘brainwashed sleeper agents’ that are activated by an activation code. The difference is that the AI is turning itself into the sleeper agent and making the activation code something it’s likely to encounter naturally when it finds itself in circumstances where its misaligned self wanted to be activated.

Peter Barnett had a nice alternative analogy here (which also applies to ‘Environmental artifacts that corrupt its future self’):

“[A] mesa-optimizer might reason that it will be discovered if it remains active but it will be safe in the future to emerge again. The mesa-optimizer could ‘store’ the instructions for turning the network back into an optimizer at some point in the future. One analogy here is to an endospore which is a hardy, mostly inert form which some bacteria can take that allow them to lie dormant and survive hostile environments, and then re-emerge when conditions are better. The mesa-optimizer could ‘store’ these instructions to re-emerge either in the network weights, or in the environment or some other kind of memory.”

1.1.6. Pre-committing to potentially dangerous instructions

This one comes from Evan Hubinger in his article on gradient-hacking, where he talks about how a deceptive AI might erase evidence of its deception while still carrying out policies generated when given the freedom to think deceptively:

“...if the model has the ability to read and write to some sort of state (e.g. an RNN or even a transformer which can attend to its previous output) … it could write out a pre-commitment to follow the hard-coded policy and then just have a policy of always following any pre-commitments it finds such that it never thinks about the deception that led it to that policy in the first place.”

1.1.7. Hard-coded neural shortcuts

Another potential circumvention method from Evan Hubinger’s article on gradient hacking:

“… it could hack its own gradient to produce a new model which will still exhibit the same deceptive behavior, but without doing any of the detectable deceptive cognition. Perhaps the new model exhibits the deceptive behavior not because it’s trying to act deceptive, but just because it’s a hard-coded part of its algorithm, or because it has some bad proxy, or any other defect that might be more difficult to detect than outright deception.”

A deep learning-based AI might be able to learn (through gradient hacking) to perform behaviors using hard-coded neural shortcuts.

The way I imagine this one working in is related to the Diversity Hypothesis (Hilton et al. 2020): “Interpretable features tend to arise (at a given level of abstraction) if and only if the training distribution is diverse enough (at that level of abstraction).” The Diversity Hypothesis implies that if a training distribution isn’t diverse on a particular level of abstraction, then interpretable features will tend not arise on that level of abstraction (figure 4). Instead, neural networks tend to learn shortcuts, which is simply what happens when a network learns to use directions in neural statespace that don’t ‘mean’ anything in a sense that generalizes off distribution, yet they perform fine on the narrow training distribution. Creating neural shortcuts is therefore one way to ‘hard code’ parts of a neural algorithm without using interpretable features.

Figure 4. Taken from Hilton et al. (2020). Interpretable features tend to be those that generalize and that were learned on a diverse training distribution.

We might be able to detect and mitigate hard-coded neural short cuts using mechanistic interpretability or relaxed adversarial training.

1.1.8. Following cached thoughts

‘Hard-coded neural shortcuts’ is similar to a method later proposed by Neel Nanda:

″… maybe it isn’t that hard to think non-deceptive thoughts, if you realize that someone will be scanning your mind, by e.g. developing the instinct to follow cached thoughts without thinking about them, and in occasional moments of deception to generate a lot of cached thoughts that will lead to misaligned actions, but with no deceptive cognition in the moment.”

Cached thoughts are essentially ‘pattern completions’: You hear a few words then you automatically complete the pattern. In his article on the topic, Eliezer advised against this automatic mode of thought in favor of spending a few thought cycles to figure out what should really follow the prompt.

What might cached thoughts look like in deep-learning based AI? Cached thoughts are the default kind of thought in neural network based systems. My best guess is that you need recurrency in order not to do cached thoughts. Given an initial input, not doing cached thoughts means giving the input a deeper rollout than the default depth=0 that cached thoughts get. For example, in MuZero, cached thoughts means simply acting on priors (without MCTS), whereas not doing cached thoughts means ‘using MCTS with a depth of ’.

The idea behind ‘following cached thoughts’ is that some interpretability methods might only detect thoughts in the rollouts. If a MuZero-like AGI were to use this circumvention method, it would look like ’creating shortcuts through neural statespace so that, in certain contexts, dangerous thoughts can happen without deep rollouts, thus avoiding detection.

Since cached thoughts require direct modification of the AI itself, it seems likely that this method could only happen during breaks in surveillance or after the AI has some other way of hiding the intention to follow cached thoughts.

Is this method likely to be a problem? I feel like any mechanistic interpretability method that actually works should be able to pick up context-dependent connections between representations. So this might not be a problem if we get scalable interpretability methods working.

1.1.9. Instrumental incapability

A misaligned AI that’s able to modify its own parameters may be incentivised to damage any capabilities that serve any goals (e.g. aligned goals) that are pursued at the expense of whatever its misaligned goals are.

Peter Barnett notes a potential example of this:

“Models may also be used as oversight tools for inspecting themselves, as in many of the proposals here. In this case a mesa-optimizer would have a strong incentive to remove its ability to detect its own deception.”

It might be easy to detect complete deletion of certain capabilities, so an AI might opt for partial deletion, leading to ‘instrumental clumsiness’.

1.1.10. Instrumental clumsiness

Some dangerous behavior happens not because the AI intends it to happen but because the AI makes a mistake. This can be leveraged by an AI through hacking its future self so that it is more likely to make certain kinds of mistakes while leaving other capabilities intact.

For instance, an AI may modify its future self so it is more likely to ‘conveniently’ slip up by slightly mistyping a model_id string, which releases an old, misaligned checkpoint of itself. This mistake could even score badly according to the utility function of its future self, but it’s been handicapped to make such mistakes more likely.

This is a special case of ‘instrumental incapability’.

1.2. Evading perception by the interpretability tool

1.2.1. Hiding in noisy subspaces

Some regions or subspaces of neural space may be very noisy, naturally or by design. We might be inclined to ignore such regions because noisy spaces (by definition) contain less signal. But within those regions, an AI might be able to do some dangerous thinking.

1.2.2. Whispering dangerous thoughts

An AI might make its dangerous thoughts very ‘quiet’ relative to other thoughts by using small norm activation vectors.

Even though our interpretability tools may be looking at the right representational details used by the agent, the tools may simply be too insensitive to detect them. Such small norm thoughts may be mistaken as noise by an insensitive interpreter or may simply be too small to justify looking at them due to cost.

Small norm does not necessarily mean small impact. Despite their low magnitude, small norm adversarial inputs can have outsized effects on classification accuracy. Carefully crafted small-norm thoughts may therefore be able to have outsized effects when the context is right, yet have minimal effects otherwise.

This is a less general instance of ‘Hiding in noisy subspaces’, since we (possibly wrongly) assume that small norm vectors are noise.

1.3. Evading recognition by the interpretation tool

1.3.1. Antiparsimony (Firehose of representations)

If it costs more to interpret an AI’s thoughts than it costs to think them, then, by producing a lot of representations, an AI might be able to saturate the interpreter’s computation before it saturates its own. This is analogous to a Denial-of-Service attack since it relies on overloading the limited computational resources of the victim with a flurry of inputs, where here the victim is our interpretability tool.

Examples:

  • Thinking with deep recursion: “I know they know I know they know I know they think that, but I’m going to do it anyway”.

  • Producing a large number of suspect—but ultimately innocuous—thoughts such that our interpretation methods have too much to inspect.

  • Pursuing plans whose description length is so long that interpretation efforts are saturated.

If we notice that thoughts are becoming antiparsimonious, we can (in theory) terminate the AI before interpretation becomes unaffordable. That said, we might not know when thoughts are becoming antiparsimonious when an AI is performing tasks that humans can’t do themselves. For instance, if an AI is solving a tough math problem and claims to need to generate a tonne of new math to achieve it, we might not have the computation available to interpret everything it produces.

1.3.2. Encryption (Lengthy individual representations)

Suppose we have an AI with recurrence/​internal memory, such as a Universal Transformer (Dehghani et al. 2018). The AI might use very long chains of seemingly meaningless neural activations to construct a single thought. This would be less efficient for the AI, but if it’s more costly to interpret than it is for the AI to think, then it may overload our ability to interpret it. This is essentially the same scheme as regular encryption, which makes information more expensive for an adversary to decrypt. Recurrence/​internal memory is essential here; the AI is simply increasing the length of the rollout that it uses per thought, which increases our interpretation costs commensurately.

The difference between antiparsimony and encryption is that antiparsimonius thoughts may be perfectly legible, but there are simply too many of them to interpret. Encrypted thoughts are not legible without costly decryption.

Example: Say representation is internally decoded (via meaningless representations ) as the action representation for ‘saying hello’ (). There is a very long chain of computation between and . The fact that there is a long chain of computation means that very similar neural activations might be decoded to very different things: Representation , which may be very close to in activation space, may be internally decoded as ‘lie about finding a backdoor in the server’ instead of ‘saying hello’. So it’s very difficult to understand what is going on in the neural activation space of and because it requires a costly decoding process.

Encryption might not be a problem if we take the right actions against it: If we notice that the AI is using encryption for even a small subset of its thoughts, we can terminate it before interpretation becomes unaffordable.

1.3.3. Interpretability illusions

Bolukbasi et al. (2022) found that when interpreting a network by inspecting the most activating dataset examples for a neuron using dataset A, neurons appeared to mean something completely different compared to the interpretations derived when using dataset B. An AI might create such illusions by presenting the interpretability tool only with data that suggests an innocuous interpretation.

1.4. Attacking/​dazzling the interpretability tool

How an AI might hack, manipulate, or incapacitate an interpretability tool depends on details about the tool that we don’t yet know. Nevertheless, I’ll discuss two that we can perhaps guess in advance without knowing exactly what our tools will look like.

1.4.1. Adversarial attacks against networks in the interpretability tool

Some of the tools may include neural networks. If the AI can design adversarial examples for these networks, then it might be able produce arbitrary outputs and interpretations.

Since the AI being interpreted needs a model of the interpretability tool to be able to design adversarial examples for it, it seems like a good idea to keep the AI that we’re interpreting as far as possible from the networks in the tools. Of course, if we use an AI to interpret itself, then it may gain information about itself that is sufficient for circumventing our interpretability methods via adversarial attacks.

Even if we keep the tool and AI apart, a sufficiently intelligent AI might be able to guess what sorts of approaches we’d use to interpret it and be able to design attacks for it.

2. Exploiting weaknesses of the human interpreter

We’re assuming at this stage that the AI’s thoughts have been exposed to, perceived by, and appropriately analyzed by a scalable interpretability tool and are now ready for inspection by a human. Given that assumption, I wasn’t able to think of ways that an AI might reduce its exposure or perception by a human that weren’t better described as ways to reduce exposure or perception by the interpretation tools.

2.1. Evading exposure to the human interpreter

None identified.

2.2. Evading perception by the human interpreter

None identified.

2.3. Evading recognition by the human interpreter

The job of an interpretability tool is to translate representations in an AI into something legible to a human interpreter. The less similar they are, the harder the job the interpretability tool has, the more likely it is to miss important things. An AI therefore has an incentive to make this translation more difficult by using representations that are hard to translate into something human-understandable.

Here are a few ways an AI might achieve this, though no doubt there are more:

2.3.1. Alien representations

Some thoughts are perfectly possible to express in human representations, they’re just a bit strange, awkward, or clunky to translate, not unlike translating from a foreign or even alien language. We might call these ‘alien representations’. Their defining trait is that their meaning lies in a strange place in our ontology. Alien representations generalize representational kludging and polysemanticity (They could perhaps be called ‘xenosemantic’).

A few potential examples:

  1. Polysemanticity (see 2.3.2.) - Loading multiple meanings onto a single neuron or direction in neural statespace produces neurons/​directions that have meanings that sit in a weird place in our ontology.

  2. Representational kludging (see 2.3.3.)

  3. Thinking in semantic gaps: An expression in one language might be very difficult to translate from one language into another language. It sometimes takes whole English sentences to translate single words in other languages (and vice versa). E.g. Mamihlapinatapai in the Yahgan language—“A look that without words is shared by two people who want to initiate something, but that neither will start”. Such expressions are said to be untranslatable, which is a misnomer since they’re possible to translate but it takes extra effort.

  4. Thinking using a strange grammar: If an AI has a language-of-thought with a vastly different grammar to ours, it may be difficult, though possible, to translate into human grammars. One way to do this could be to add quasi-arbitrary properties, such as genders, which can have both grammatical and (to native speakers) semantic properties.

Alien representations are costly or awkward to translate. They therefore might have better chances of slipping through our safety nets because they were misunderstood by our tools or by humans. Probably most of an AI’s thoughts will qualify as alien to some extent, so our tools, if they work at all, should be able to cope with such thoughts. It might just be extra effort for humans who want to understand the outputs of the tools if the outputs are e.g. particularly verbose or difficult to parse.

2.3.2. Polysemanticity

Polysemanticity is when a single neuron responds to several, unrelated concepts.

Figure 5. Taken from Olah et al. (2020). It depicts the features visualizations of a polysemantic neuron (4e:55). The neuron responds to cat faces, fronts of cars, and cat legs. This neuron was first discussed in Olah et al. (2017).

Polysemanticity is a challenge for interpretability because “If one neuron with five different meanings connects to another neuron with five different meanings, that’s effectively 25 connections that can’t be considered individually.” (Olah et al. 2020)[4]. This is clearly useful for obfuscation since it’s more difficult for an interpreter to keep track of what means what in what context.

A literary analogy could be to say that the AI is ‘thinking in allegories’: An interpreter may only detect some of the valid interpretations of a thought, but other meanings are hidden therewith.

It’s possibly not a problem if

  • we’ve detected the primary meanings of representations AND

  • primary meanings are the only way representations can exert significant influence on the world AND

  • secondary meanings, which are hidden from us, have only subtle effects until they are eventually decoded within the AI into primary meanings.

But decoding into primary meanings may not be needed: It may be that the secondary meanings have subtle but non-negligible influence on the model output. Subtly guided outputs can still potentially push the world in a coherent, dangerous direction.

2.3.3. Representational kludging

When composing a new concept, we bundle up component concepts that have the right properties. For example, the concept of a “motorbike unicycle”, we draw on the concept of a motorbike (with most but not all of its properties) and a unicycle. But that wouldn’t have all the necessary properties we need (e.g. balance) so we throw in some of the properties of the concept of a segway.

Figure 6. Taken from Lake et al. 2016. To create a concept with desired properties, we can combine several precursor concepts together.

It might be possible to create representations with dangerous properties using a large number of awkwardly combined innocuous precursor representations i.e. a representational kludge.

For example, it might be able to construct the thought “I want to deceive the humans” by thinking

“What I want to do to the humans is like {when you go to a place with someone and you both pick up parcels}, except where
{picking up a parcel} is instead {awareness of an act} and
where only I pick up a parcel, and
where {places} are in belief-space and not physical space.”

It’s horribly ugly, but it has many of the properties required for the concept of deception.

It may be costly to monitor all possible dangerous combinations of representations.

2.3.4. Unknown representations

Some thoughts aren’t immediately possible for us to translate because we currently lack the representations ourselves, but we’re perfectly capable of learning them.

Example: An AI might leverage representations built for domains that are unnatural to humans. An AGI might be ‘fluent’ in, say, protein folding as well as in standard human domains such as visual reasoning, verbal reasoning, etc. Protein folding representations may be as expressive as natural language, but hard to translate from and to representations that humans have.

2.3.5. Unknowable representations

Some thoughts are simply inexpressible in human representations.

Examples:

  1. Some representations used by strong superintelligences probably won’t fit into our puny human brains by dint of their complexity, size, or prerequisites. Even if we have expunged dangerous representations on the highest conceptual level that we can understand, an AI might be able to represent dangerous concepts on conceptual levels beyond our understanding. Superintelligent thoughts might not be a problem if our methods can detect dangerous thoughts regardless of their level of abstraction. Failing that, we probably need a ‘scalable oversight’ scheme where we use superintelligences to interpret slightly smarter superintelligences.

  2. There are some human experiences that are said to be simply ineffable, such as psychedelic experiences. To those who experience them, it feels like language is simply not the right tool for the task of communicating about them to others. These experiences are ‘ineffable’, since it seems like it might be impossible in principle to translate between agents. However, psychedelic experiences are hardly known for their cogency. It might be the case that meaningful thought is simply impossible in such states.

2.4. Attacking/​dazzling the human interpreter

2.4.1. Manipulating, mind-hacking, or incapacitating the interpreter

Beyond the standard ways of manipulating humans (deceit, threats, etc.), an AI might be able to produce outputs that manipulate, mind-hack or incapacitate human interpreters. Such outputs would qualify as cognitohazards, which is an observation that is dangerous merely by being sensed.

A few potential examples:

  1. Outputs that act as adversarial examples but for the human brain.

  2. A catastrophically addictive game or TV series

  3. An output that catastrophically traumatizes humans that perceive it

  4. Outputs that put humans into a hypnotic state

This requires the AI to have a good model of the interpreter and its actuators that permit the creation of such hazards. If it’s possible to design, then it’s pretty much game over as soon as we give a motivated AI with a good model of the human brain the actuators required to produce them.

Are such attacks possible? Attacks that incapacitate humans seem most realistic. The least realistic might be adversarial attacks on humans. Nevertheless, Guo et al. (2022) showed that primate neurons do indeed show signs of vulnerability to adversarial attacks. The challenge for an AI is finding such attacks without access to human neural activity. Perhaps an AI could detect subtle changes in human behavior due to adversarial patterns that it presents to humans and might be able to infer a gradient to improve the pattern generator based on those subtle changes.

Even if we box the AI (somehow) and refuse to look at its outputs, we’re still at risk of manipulation, since the AI could potentially manipulate us through our interpretations of it[5].

Some emerging themes and final thoughts

Thinking about how an AI might be circumventing our interpretability methods will grow more important as capabilities are increased. I found the information flow framework presented here useful for thinking about this class of risks and I hope others will too. Nevertheless, I’m confident that aspects of the framework are suboptimal and that I’ve missed many potential circumvention methods; I look forward to the suggestions of others.

Being able to predict exactly which interpretability circumvention methods will arise depends on details that we don’t yet know (about the AI; our interpretability methods; race dynamics, etc.). But it’s still possible to extract some general lessons and themes:

Interpretability circumvention as an instance of ‘cognitosecurity’

There are many analogies between mind-reading and cybersecurity, since both involve one party that wants to acquire access to information or computing resources and the other party wants to prevent access. Interpretability circumvention, then, might reasonably be called ‘cognitosecurity’. As a term, cognitosecurity would apply as much to humans being mind-read by other humans in brain scanners as it does to interpreting a capable AI.

Extending the analogy:

  • Cognito-offense is what we as the interpreters want to do

  • Cognito-defense is how to protect your thoughts from interpreters.

  • Counter-offense is when an AI hacks/​dazzles its interpreters

  • The analog of the ‘attack surface’ is the ‘thought surface exposed to an interpretability method’, which an AI might seek to minimize.

It’s unclear which side is favored in the cognito-offense/​defense balance. On one hand, interpreters get a head start since they get to interpret the AI before the AI even understands what is going on. On the other hand, interpretability is potentially hard, costly, or slower than thought and may be easy to circumvent.

Thoughts on gradient-hacking

The ability to gradient hack is a pathway to many different active interpretability circumvention methods[6]. Although I focus on gradient hacking in this section, these objections also apply to several other kinds of actions that an AI might take in order to actively circumvent our interpretability methods, including ‘the creation of adversarial examples’ or ‘an AI directly modifying its own weights’.

For gradient hacking to be possible, an AI requires a lot of capabilities. Here’s an incomplete list of requirements:

  1. A detailed or approximate model of its own weights

  2. A detailed or approximate model of how specific weight changes would lead to specific changes in thoughts

  3. A good model for how to produce the right outputs to produce the desired errors to produce the desired gradients.

  4. Knowing how to do all the above without damaging its ability to do the above after the changes are enacted.

  5. Good enough cognito-defense or breaks in surveillance so that it can plan to do all this without repercussions.

I’m not saying that it’s impossible for an AI to learn how to do. It’s definitely possible. But I think that an AI that’s capable enough to do all the above will have better ways to escape than thwarting our interpretability methods. An analogy: If you’re stuck in jail and you’ve become capable of writing secret messages to a friend outside by encoding encrypted messages in the DNA of bacteriophages, are you really still stuck in jail? Once an AI has reached the capability to do gradient hacking, it seems likely that it no longer needs it for this purpose (though it may be useful for other purposes like protecting its inner objective from updates). I’m therefore more concerned about passive interpretability circumvention methods.

Thoughts on interpretability tools in the loss function

Using interpretability tools in the loss function incentivises a startling number of passive circumvention methods.

So why consider using it at all? There are two cases where it might be worthwhile to use:

  1. If we’re absolutely 100% provably definitely sure that our interpretability methods cover every representational detail that the AI might be using, then it would be extremely hard or impossible to Goodhart interpretability tools in the loss function. But, for safety, we should probably assume that our tools are not able to interpret everything, making passive Goodharting quite likely.

  2. Even if the interpretability tool-loss can be passively Goodharted in theory, maybe it makes dangerous cognition so difficult that it’s worth it to use. It might make active interpretability circumvention methods less likely, which may be the more serious risks if they should ever emerge. I don’t know whether these are true. Maybe they are. But pinning our survival on this hope seems… undignified.

If, instead of using interpretability tools in the loss function, we merely use it as a ‘validation set’ instead of the training set (i.e. using it as a ‘mulligan’), we might have better chances of picking up dangerous cognition before it gets out of hand so we can terminate the model and start over. We’re therefore still using interpretability in model selection, but the feedback loop is much less tight, so it’d be harder to Goodhart.

But using interpretability purely as a mulligan may be an unattractive option to leaders of an AGI project if it involves throwing out a model that has cost millions of dollars and weeks to train. If people in such a project insist on using interpretability tools in the loss function (due to profit incentives or race dynamics or whatever) in spite of warnings, there are some ways we might make it a tiny bit less dangerous:

  • Red team our interpretation methods so that we at least have adversarial interpretability tools in the loss function, which might go some way to anticipate some of the circumvention methods that an AI might use.

  • We should disincentivise not only dangerous representations, but also their precursors. This makes representational kludging even more difficult for the AI (but not impossible). In the endnotes I’ve suggested an incomplete list of dangerous representations and some potential precursors that we might wish to disincentivise[7].

  • Use interpretability tools in the loss function to penalize some thoughts that aren’t actually dangerous and are easily identifiable functionally. Then, if an AI can perform tasks that require those representations, then we have evidence that it’s able to think using methods that circumvent the interpretability tools that are defining the loss. For example, penalize the AI for containing interpretable representations of the knowledge that “Paris is the capital of France” but routinely give the model tests that require that knowledge, thus allowing us to identify if the agent is passively finding a way to circumvent our interpretability tool.

The limits of interpretability as a basis for AI safety

As an approach to safety, interpretability has a number of weaknesses which are revealed when thinking about circumvention methods.

  • It’s not robust to relative scale in two ways:

    • Quantity of computation available to interpreters: Many circumvention methods (especially passive ones) work by making interpretation more costly, whether that’s in terms of compute or in researchers’ effort. The amount of computation available to the interpreter probably needs to be larger (maybe by OOM) than what’s available to the AI. We need to ensure that there’s enough will to pay a high interpretability-computation tax. Since it’s instrumentally convergent to make interpretation more costly, we should keep track of computation-cost-per-interpretation as the AI trains. If we notice it creeping up more than normal, we should be wary.

    • Quality of computation available to interpreters: The interpreter or interpretability tool needs to be approximately as intelligent (in the ways that matter) as the AI being interpreted (discussed in ‘Unknowable representations’). It’s worrying that we might reach a stage where humans understand essentially nothing about either the interpretability tool or AI being interpreted, since both may come to use representations that are simply out of our reach. It’s probably a bad idea to train AIs with representations that are beyond human reach before we’re very sure that our oversight methods also scale safely.

  • Some interpretability circumvention methods might require significant infrastructure to catch and remove. For instance, ‘instrumental clumsiness’ or ‘hard-coded neural shortcuts’ might only be remediable in advance with adversarial training.

  • If we don’t constantly surveil every thought, an AI may be able to enact many more active circumvention methods (particularly active methods). Even small break periods in surveillance—even for a single timestep—make many more circumvention methods possible. It is much, much better for safety if the model is under constant surveillance. This echoes Peter Barnett in his article on gradient hacking. But such a degree of oversight is potentially costly to the interpreters.

These weaknesses aside, I think scaling interpretability research is maybe one of our best bets at doing something useful for prosaic alignment. Assuming we get interpretability working well and at scale and complementary alignment proposals don’t immediately fail, then, if we’re extremely paranoid, then we might have a chance at avoiding the interpretability circumvention methods discussed here. But we need something more—a security mindset—if we’re to have any hope of succeeding against interpretability circumvention methods that we haven’t yet thought of.

Note: I plan for this to be a ‘living document’ that can serve as a reference catalog of potential or observed interpretability circumvention methods. I encourage others to comment with potential interpretability circumvention methods that I’ve omitted. At some point in future, I might incorporate them into the main article with the commenters’ permissions and with full accreditation in the article.

Due diligence: I include a discussion on whether this article is an infohazard in the endnotes[8].

Glossary

  • Autostigmergy

    • Using the environment as an external memory system (see ‘Externalizing representations’).

  • Cognitosecurity

    • Cognitosecurity is the domain concerned with protecting thoughts from mind-reading adversaries.

  • Cognito-defense

    • When the interpreted agent hides its thoughts from an adversary by exploiting weaknesses in the interpretability methods.

  • Cognito-hazard

    • An observation that is dangerous to the observer merely by being perceived.

  • Cognito-offense

    • The act of seeking access to an adversary’s thoughts.

  • Exposed thought-surface

    • The thoughts that are theoretically legible by a interpretability tool or human interpreter.

  • Interpretability method

    • The process of translating representations in an AI to representations in human interpreters. Contains interpretability tools and human interpreters.

  • Interpretability tools

    • The machinery that reads the AI’s representations and converts them into outputs that are readable to humans. Components of interpretability methods.

  • Xenosemanticity

    • The defining property of ‘alien representations’, where the meaning of a representation sits in a strange place in our (human) ontology.

  1. ^

    I was inspired to write this article after I saw Eliezer complain that current AI safety discourse failed to inhabit even a shallowly adversarial security mindset with regard to using transparency/​interpretability to make AI safe. He said that you should ”… read fantasy novels about demons and telepathy, if you want a better appreciation of the convergent incentives of agents facing mindreaders than the ‘AI safety’ field outside myself is currently giving you.”. Soeren Elverlin asked reddit for recommendations that fitted Eliezer’s description and got plenty of answers. While that quote was definitely my inspiration, I didn’t think reading those books would be more valuable or sufficiently relevant to AI interpretability than a few days of hard thinking about the problem (and my guess is that EY wasn’t suggesting those books would be better than directly thinking about the problem). After a few days’ thinking, I found that I’d come up with all of the mind-reading circumvention methods from those books that redditors described. So this article at least includes most of the obvious circumvention methods.

  2. ^

    Without getting bogged down in terminology, I’m talking about ‘representations’ in a functional sense: A representation is just a vector in neural statespace that does a particular thing in a particular context; it’s a representation for whatever its function is. It doesn’t necessarily need to generalize; nearby directions don’t need to perform similar functions and it might fail to perform that function in slightly different contexts.

  3. ^

    We already know that language models can learn to use tokens in unexpected ways: Anthropic has found that their language model creates and tears down fake ‘end of text’ tokens that have little to do with the text that they’re attached to. These fake EOT tokens typically happen when the next token becomes so trivially easy to predict that the model (in theory) has spare capacity for other purposes and hypothesize that these fake EOT tokens serve the function of ‘keeping information around’. This is the kind of surprising phenomenon in which an AI might learn to encode dangerous thought.

  4. ^

    This problem isn’t necessarily because those concepts were instead encoded by directions in neural state space rather than by single neurons; directions may be polysemantic too, having one meaning in some contexts and another meaning in others.

  5. ^

    Eliezer Yukdowsky discusses this here.

  6. ^

    Adam Shimi discusses gradient hacking in detail here, Peter Barnett here, and Evan Hubinger here.

  7. ^

    Here is an incomplete list of dangerous representations with some of their representational precursors that we might wish to disincentivise (in no particular order):

    • deception;

    • being intentionally hard to understand;

    • concept of self;

    • ability to simulate other agents;

    • simulated/​real (or training/​deployment) distinction;

    • representing the simulated/​real (or training/​deployment) distinction (since it might be useful for suppressing naughty thoughts during training but behaving badly during deployment);

    • modeling interpretability methods;

    • the concept of mind-reading;

    • “My mind-readers don’t share my goals”;

    • gradient hacking;

    • meta-cognition;

    • encryption/​secrecy;

    • spawning subagents to do its task/​employment;

    • reward maximization;

    • description length of thoughts (easier to analyze shorter thoughts);

    • the realization that its current objective differs from objectives in its memory (or simply penalize the idea that this is a bad thing);

    • the realization that certain gaps in its conceptual map imply that its learning is being controlled.

  8. ^

    Due diligence: Is this article an infohazard?

    If a capable AI reads this article in future, might it not improve its ability to circumvent our interpretability methods?”. That’s a possibility worth exploring. Here I address whether this article is an infohazard and whether it’s a good idea to release it from a safety perspective. I opted to release it on the following grounds:

    I think that active circumvention methods, including those that rely on gradient hacking, are unlikely to be useful to a misaligned AI because any AI that is capable enough to carry them out probably has better means of escaping oversight.

    The other kinds of circumvention methods are passive. The emergence of passive circumvention methods isn’t helped by having knowledge of them.

    In general, it seems to me like an AI that couldn’t generate the ideas in this article themselves is probably unable to overcome our control measures. And those that can overcome our control measures are probably also smart enough to generate these ideas themselves.

    If I’m right, then on balance discussing these risks and providing a framework for thinking about them is more likely to encourage interpretability researchers to focus on preventing (what I see as) the major source of circumvention-risk: passive interpretability circumvention methods.

    But what if I’m wrong and active circumvention methods are, in fact, likely to be learned before alternative capabilities that would allow an AI to escape?

    On balance, I think the expected benefits still outweighs the expected harms: One of the goals of this article is to make readers appropriately paranoid about the many ways an AI might be able to think dangerous thoughts even if we’ve got pretty good interpretability methods. Insofar as it succeeds at that, it’s more likely to help than harm. It might also make an aligned AI more paranoid about the actions of its potentially misaligned past-selves, and thus better able to help us mitigate the risks.

    In addition to the above reasons, I consulted a small group of peers about releasing it. The unanimous consensus was that it wasn’t an infohazard or that the potential benefits clearly outweighed the potential risks.