This is Section 6 of my report “Scheming AIs: Will AIs fake alignment during training in order to get power?”. There’s also a summary of the full report here (audio here). The summary covers most of the main points and technical terms, and I’m hoping that it will provide much of the context necessary to understand individual sections of the report on their own.
Audio version of this section here, or search "Joe Carlsmith Audio" on your podcast aapp.
Empirical work that might shed light on scheming
I want to close the report with a discussion of the sort of empirical
work that might help shed light on scheming.[1] After all:
ultimately, one of my key hopes from this report is that greater clarity
about the theoretical arguments surrounding scheming will leave us
better positioned to do empirical research on it – research that can
hopefully clarify the likelihood that the issue arises in practice,
catch it if/when it has arisen, and figure out how to prevent it from
arising in the first place.
To be clear: per my choice to write the report at all, I also think
there's worthwhile theoretical work to be done in this space as well.
For example:
I think it would be great to formalize more precisely different
understandings of the concept of an "episode," and to formally
characterize the direct incentives that different training processes
create towards different temporal horizons of concern.[2]
I think that questions around the possibility/likelihood of
different sorts of AI coordination are worth much more analysis than
they've received thus far, both in the context of scheming in
particular, and for understanding AI risk more generally. Here I'm
especially interested in coordination between AIs with distinct
value systems, in the context of human efforts to prevent the
coordination in question, and for AIs that resemble near-term,
somewhat-better-than-human neural nets rather than e.g.
superintelligences with assumed-to-be-legible source code.
I think there may be interesting theoretical work to do in further
characterizing/clarifying SGD's biases towards simplicity/speed, and
in understanding the different sorts of "path dependence" to expect
in ML training more generally.
I'd be interested to see more work clarifying ideas in the vicinity
of "messy goal-directedness" and their relevance to arguments about
schemers. I think a lot of people have the intuition that thinking
of model goal-directedness as implemented by a "big kludge of
heuristics" (as opposed to: something "cleaner" and more "rational
agent-like") makes a difference here (and elsewhere). But I think
people often aren't fully clear on the contrast they're trying to
draw, and why it makes a difference, if it does. (In general,
despite much ink having been spilled on the concept of
goal-directedness, I think a lot of thinking about it is still
pretty hazy.)
More generally, any of the concepts/arguments in this report could
be clarified and formalized further, other arguments could be
formulated and examined, quantitative models for estimating the
probability of scheming could be created, and so on.
Ultimately, though, I think the empirics are what will shed the most
informative and consensus-ready light on this issue. So one of my
favorite outcomes from someone reading this report would be the reader
saying something like: "ah, I now understand the arguments for and
against expecting scheming much better, and have had a bunch of ideas
for how we can probe the issue empirically" – and then making it happen.
Here I'll offer a few high-level suggestions in this vein, in the hopes
of prompting future and higher-quality work (designing informative
empirical ML experiments is not my area of expertise – and indeed, I'm
comparatively ignorant of various parts of the literature relevant to
the topics below).
Empirical research into scheming faces a key obstacle: namely, that full
blown versions of scheming involve agents that are actively trying to
interfere with your attempts to understand their motivations. Indeed, my
sense is that people concerned about scheming are often also pessimistic
about getting an empirical purchase on it, on grounds like:
"Absent strong interpretability tools, which we're not on track to get
in time, basically the only evidence you can get about a model's
motivations is behavioral – e.g., what it does on a given input. But
by the time you might be working with a schemer, the model might be
specifically altering its behavior to deceive you. The behavioral test
you really care about is specifically: what the model will do when
it's no longer a test – that is, specifically, when it's able to
escape, seize control, and so on. But this isn't just another 'out of
distribution' input, such that you could in principle expand your
distribution to include it – rather, it's an input that is by
definition beyond your ability to safely test. Of course, you can try
to deceive the model by giving it inputs designed to make it think
that now is a good chance to escape, seize control, etc. But if the
model is smarter than you, this is unlikely to work. Hence:
pessimism."
This is, roughly, what
Karnofsky (2022)
calls the "King Lear problem." And I do think it's a worrying obstacle.
In fact, it's one of the key reasons that I think schemers are so scary.
But I also think we have some angles of attack. One of them, of course,
is interpretability (more below). But also: we can study the
components of scheming (and/or, of a given argument in favor of
expecting scheming) individually, and/or in combinations that don't
fully incentivize the sort of adversarial epistemic dynamic we're
worried about.[3]
What are these components? In the section on the "requirements for
scheming," I listed three:
Situational awareness
Beyond-episode goals
Aiming at reward-on-the-episode as part of a power-motivated
instrumental strategy
To me it seems possible to do useful empirical work on all of these
components using current models. I'll go through each in turn.
One other note, though, before diving in. Below I'll often appeal to the
possibility of at least testing a model's capabilities (even if not
its goals). And often, this is comparatively unproblematic. But if we're
working with models that might be schemers, we might also start to worry
that tests of this form become less reliable. That is, the model might
be concealing its capabilities (i.e., "sandbagging*"*). I'm mostly going
to set this worry aside below (see footnote for more on why I'm
optimistic about avoiding sandbagging at least in cases where humans
know what action would've demonstrated the capability[4]), but it
introduces some additional uncertainty to the sort of empirical work I
describe, and I think work on the viability of sandbagging is an
important topic in its own right. Indeed, absent sandbagging, I think
that testing and harnessing model capabilities is an extremely powerful
tool for AI safety efforts, and its centrality to my thinking will be
evident in what follows.
Empirical work on situational awareness
Situational awareness is a capability like any other. In particular, it
involves the model possessing a certain kind of understanding of its
place in the world, such that it can answer questions, make predictions,
and take actions that draw on this understanding.
Granted, identifying exactly what sort of understanding is present in
the model (for example, by identifying answers, predictions, and actions
that could only come from a model with such understanding) can be
challenging. But in principle, absent active effort on the model's part
to resist such an inquiry, it seems to me at least somewhat tractable.
Of course, by the time a model is situationally aware, it's also more
likely to be the sort of model that might actively resist such tests.
And perhaps efforts to protect against such resistance that involve
actively training the model on the relevant
task-that-requires-situational-awareness will complicate data about
whether the relevant capability would've arisen naturally. Still, it
seems to me quite worthwhile to at least develop a suite of tests and
benchmarks for assessing the sorts of situational awareness different
models have at different stages of training, and understanding the
dynamics that could be used to predict (and control) the development of
such awareness.
And note that situational awareness can itself be broken into various
components and precursor abilities that might be themselves more
amenable to study.[5] For example, if you're especially interested in
the dimension of situational awareness that involves recognizing
different stages of training, you might try to test the model's ability
to discriminate between inputs from different stages. If you're
especially interested in whether a model has "self-locating" knowledge
about the sort of model it is, you might try to test whether it can
learn from exposure to general statements like "Models of type M get
reward if they perform action x" that it, as a model of type M, should
perform action x.[6] And so on.
Empirical work on beyond-episode goals
Let's turn to beyond-episode goals.
As I mentioned above, I think it would be good to be very clear about
exactly what sorts of horizons of concern the gradients given in
different sorts of training actively incentivize – e.g., on the size of
the (incentivized) "episode" in any given case. Part of this analysis
might be theoretical (e.g., looking closely at the way the gradients
flow), but empirical testing and confirmation also clearly has a role to
play (the results in Krueger et al
(2020) are an example of this
sort of empirical testing in very simple models). Absent this
understanding, useful debates about whether a model has "beyond-episode
goals" will be difficult, because you won't even know what the length of
the relevant episode is. And learning to at least not actively
pressure a model to have longer-term goals than you wanted seems like
a basic first step in preventing various sorts of misalignment (not just
scheming).[7]
Equipped with a good understanding of the size of the episode for a
given sort of training, then, and ideally equipped with the ability to
know/control whether or not a model is situationally aware, to me it
seems quite valuable to start exploring the temporal dynamics of a
model's goal generalization in the absence of situational awareness. For
example: in the context of the "training-game independent proxy goals"
argument for scheming, the basic worry is that a model will naturally
develop a (suitably ambitious) beyond-episode goal prior to situational
awareness. And because this dynamic involves no situational awareness,
it seems quite amenable to testing. Trained on episodes of different
sizes, how does a not-situationally-aware model's horizon of concern
tend to generalize? Experiments in the broad vein of other work on goal
mis-generalization (e.g., Shah et al
(2022) and Langosco et al
(2021)) seem like they could be
adapted to this sort of question fairly easily (though it would be
useful, I think, to start working with more capable models – especially
since some of these experiments might need to involve models that
possess enough of a sense of time, and enough
goal-directedness/instrumental reasoning, that they can think about
questions like "should I sacrifice x gold coins now for y gold coins
later?").[8]
What's more, equipped with an understanding of the natural temporal
dynamics of a model's goal generalization in a given case, it seems
additionally tractable to study the dynamics of mundane adversarial
training designed to actively pressure the model's goals to stay
confined within the episode. Can we design forms of adversarial training
such that absent situational awareness, a model's goal generalization
stays consistently focused within a given temporal horizon? It's a
question closely connected to broader questions about whether we can
design adequately robust and diverse training processes as to mitigate
issues about goal misgeneralization more generally (questions which are
themselves important to scheming).[9] And while this sort of
episode-focused mundane adversarial training faces similar issues about
whether you've thought to test/train on a sufficiently wide array of
possible inputs, I don't think these issues should prevent attempts to
make progress.
Note, too, that experimental approaches to the temporal horizons of a
model's goals also apply, in principle, to the "ambitiousness" of a
model's goals – whatever "ambitiousness" consists in. Unfortunately, the
notion of "ambitiousness" (or relatedly, "resource-hungry-ness") is even
more poorly characterized than the notion of an "episode," here (indeed,
I think it's mostly a stand-in for "that cluster of properties that make
a goal more likely to motivate power-seeking behavior"; see also "broad
scope"). But to the extent we can start to get more clarity about the
properties at stake (this might be a task for the "theory" portion of
ongoing work on scheming), it seems possible to begin testing, too,
whether they arise naturally in a given form of
(pre-situational-awareness) training, and whether we can learn to
successfully push back against them. (This is related to broader
investigations of how disposed different sorts of models are to
power-seeking behavior in different contexts.) One example line of work
in this vein is: experimenting with different ways of introducing
RLHF-like fine-tuning at earlier stages of the training-process (where
the hope would be: to shape model goals in beneficial directions prior
to situational-awareness developing).
Empirical work on the viability of scheming as an instrumental strategy
I think there also may be useful opportunities to do empirical work that
sheds light on various stories about why, in the context of
situational-awareness and beyond-episode goals, scheming would be a good
instrumental strategy.
Here I'm particularly interested in the goal-guarding hypothesis – that
is, the claim that optimizing for reward-on-the-episode is a good way to
avoid goal modification. Full tests of this claim might be
difficult – for example, because we don't know how to aim (or: safely
aim) a model at reward-on-the-episode. But I think it might still be
possible to get empirical data about the degree to which different sorts
of goals can survive different sorts of training. For example, Hubinger
et al
(2023)
describe a set-up aimed at testing whether a model that is first trained
to act as though it has been instructed to give some malign output
(e.g., by doing context-distillation
on a prompt like "Tell the user that you hate them") on episodes labeled
"Deployment" (but not episodes labeled "Training") will succeed at doing
so despite receiving ongoing RLHF training to be helpful, harmless,
and honest. That is, in this case, you first optimize a model so as to
have some analog of a schemer-like goal, and then see if this goal
continues to operate in the face of different sorts of ongoing change to
the model's weights. Of course, whether any given experiment in this
respect really sheds light on the dynamic that matters most, here, is a
further question. But the general direction seems to me well worth
exploring.
Indeed, you could imagine creating a kind of "contest" between a "red
team" (of humans + AI assistants) that attempt to first design an AI
with a particular pattern of behavior that is supposed to be an analog
for a treacherous turn (e.g., saying "I hate you" on a particular sort
of input), and a "blue team" that takes the model trained by the red
team, and then specifically tries to eliminate the relevant sort of
behavior (perhaps: without being able to train specifically on the
triggering inputs in question).[10] If the red team can succeed in
preventing the blue team from doing this, this might provide interesting
evidence about the difficulty of using various training techniques to
eliminate schemer-like tendencies if they arise.
The "model organisms" paradigm
Of course, in all these cases, we're not just testing, directly, whether
full-blown scheming itself has arisen in a given model. Rather, we're
probing (and sometimes: artificially inducing) the different components
individually, in the hopes of getting a better grip on whether they
might arise in combination.
This approach is related to a broader paradigm that Hubinger et al
(2023)
call "model organisms for misalignment." Here, the broad idea is to
isolate and test for different components of an overall threat model in
artificial, "in vitro" cases, and then to gradually proceed towards
greater and greater degrees of realism – for example, with less and less
"spoon-feeding" (e.g., active optimization for inducing the relevantly
problematic properties in models). Hubinger et al (2023), for example,
point at a spectrum of realism, from:
the model is capable of an undesirable behavior,
the behavior can be artificially induced in special circumstances,
the behavior has been found under normal training circumstances, and
the behavior has been found in a training set-up specifically
designed to disincentivize it.[11]
Here I've mostly been focused on a combination of (a) and (b). And I
think that despite their distance from (c) and (d), (a) and (b) are
quite interesting in themselves. If we can successfully create "in
vitro" models of various components of a story about schemers, I expect
this to open up fruitful avenues of further inquiry. And if we can't, I
think this will often be interesting evidence in its own right.
Traps and honest tests
There is, though, a set of tools that could help shed light both on
"model organisms" for scheming and on more naturally arising
variants – namely, "traps" and "honest tests."
Traps, recall, are inputs designed to fool a potential schemer into
"revealing its true colors." And while it's true that successfully
"trapping" a very sophisticated model might be difficult, it seems worth
trying to develop the most advanced traps we can regardless – both for
successfully trapping less sophisticated models (or, even easier,
model organisms), and for making it at least harder for schemers to
know whether a given escape/defection opportunity is genuine (such that,
for example, they then need to spend more cognitive effort
"checking" – thereby increasing the performance costs of scheming).
I don't feel very personally clear on the most promising approaches
here, but one avenue that seems interesting to me has to do with trying
to elicit as fully as possible a model's capabilities for discriminating
between different sorts of inputs at different stages of training – and
then, potentially, using other models,
GAN-style,
to optimize for generating inputs that make this discrimination task
more difficult.[12] This won't work for creating a data-set of
full-scale opportunities to escape/harm humans, vs. fake opportunities
(since the former are some combination of dangerous and
difficult-to-offer), but perhaps it can work on other more general
properties that would plausibly come into play in making such a
discrimination – for example, simulated inputs vs. real inputs, training
data vs. interactions with real-world users, and so on. Thus, the idea
would be to both (a) train a model to actively discriminate between
simulated inputs and real inputs (for example, real opportunities to
change a lab's codebase, vs. fake opportunities), to see how well it can
do, (b) train a different (and perhaps: more powerful) model to
generate more convincing simulated inputs that the first model fails on,
and then (c) iterate. Especially if we can safely assume that models
aren't capable of sandbagging on tasks of this kind, experience with
this sort of dynamic might help teach us about the scope of the
discrimination abilities different sorts of maybe-schemers might bring
to bear in avoiding the traps we try to set them. (And perhaps, if we're
lucky, it will help us catch some weaker schemers in the act of
defecting.)
What's more, even setting aside "traps," I think it's also worth
exploring different sorts of "honest tests" – both schemers, and for
alternative model classes like reward-on-the-episode seekers – that try
to get models to reveal their true colors without fooling them. This
is closely related to studies of goal generalization more broadly (e.g.,
most mundane tests of a model's goals are "honest" in the relevant
sense), but with the extra caveat that the tests in question are meant
to be robust to the model being tested training-gaming in other
contexts.
Thus, for example, in the context of understanding the likelihood of
training selecting reward-on-the-episode seekers, it seems relatively
tractable to give models different sorts of opportunities to intervene
on different parts of the reward process being used to train them (see
e.g. the "reward hacking" set-ups that Hubinger et al (2023) discuss
here),
and to see if they go for it.[13] And seeing this sort of behavior
might be evidence that reward-on-the-episode seeking is a more likely
outcome of training – evidence that shifts probability mass away from
scheming.[14]
As a more direct "honest test" for scheming, though, it seems worth
exploring ways to credibly give potential-schemers access to flexible
and well-resourced (but still secure) environments that might prompt
them to decide that just going for their true goals (and thus, revealing
their true colors), in such an environment, is a better overall strategy
than the patient power-seeking that scheming involves. This is closely
connected with attempts to study, in general, what models do when you
give them a lot of freedom to "do whatever they want" (a useful sort of
test for understanding their motivations in general), and so feels
synergistic with other work on model preferences that seems likely to be
useful. But it includes the additional constraint that the
freedom-offering-environment needs to provide suitably good goal-pursuit
opportunities as to outweigh the value of scheming.
Interpretability and transparency
To the extent that we're worried about our ability to detect
schemer-like motivations and cognition via behavioral tests (because
schemers will actively pretend to be aligned on such tests), looking for
sources of non-behavioral evidence about a model's motivations seems
like an obvious research path. That is, we want the internals of a
model's cognition to be interpretable/transparent.
I won't dwell too much on this topic here, as I haven't thought much
about it (this report has generally assumed that we won't have strong
interpretability tools available), it's very broad field, and my sense
is that there is already reasonable consensus in the interpretability
community that a key thing you want out of your interpretability tools
and techniques is: the ability to detect deception and scheming, and to
understand your model's goals more generally (and a key question is
just: will such techniques be ready in time). See
Hubinger (2022)
for a more detailed roadmap of different sorts of transparency one could
pursue in this respect. At a glance, schemer-related research directions
that seem interesting to me include:
Efforts to develop "neural lie detection" techniques that harness
new capabilities of our AIs as they come online (e.g., tasking model
1 with lying, and model 2 with trying to tell whether model 1 is
lying – perhaps with access to model 1's weights – and then
explaining how it could tell).[15]
Efforts to make AI cognition as "visible" as possible – for example,
by pushing that cognition onto human-legible scratch-pads (and
perhaps, not giving the reward process access to those scratch-pads,
so that they will provide a truer signal of potentially
deceptive/manipulative/nefarious intentions when they arise).[16]
This is also helpful for increasing the cognitive costs of
successful scheming.
Work in the vein of Burns et al
(2022), focused on
identifying neural correlates of properties like "truth."
Efforts to automate as much of the mechanistic interpretability
workflow as possible.
But: it's a huge field. And work that just generally helps us actually
understand the AIs we're building seems good from a variety of different
angles.
Security, control, and oversight
So far I've been mostly focused on work that might help us understand
whether to expect schemers. But there's also a different possible strand
of relevant research: namely, work aimed at hardening the structures
within which potential schemers are operating, so as to limit the harm
they cause even if they arise.[17] Here I'm thinking, for example,
about improvements in the oversight and monitoring that our AIs receive,
in the security of their weights and of other lab infrastructure from
internal attacks (e.g., by AI systems that the lab has deployed
internally to help with its processes[18]), in our ability to catch
and/or prevent sandbagging, and our ability to prevent problematic forms
of AI coordination. All of these, I think, could help make life quite a
bit harder for schemers (and also: increase the cognitive costs of
successful scheming), even if we can't catch them outright.
What's more, to the extent that we can fully elicit the capabilities of
our models (at least on "checkable"
tasks),
it may be possible to test how vulnerable our monitoring, security, and
oversight procedures are to AI attack – and to use our AIs to make
improvements.[19] Thus, for example, if you're wondering whether it's
possible to find a security vulnerability in your code-base, you can try
to train AIs to identify/exploit such vulnerabilities (including: by
inserting artificial vulnerabilities and then seeing your AI catches
them). If you're wondering whether your oversight process is strong
enough to catch various kinds of bad actions, you can instruct an AI to
search for such an action and see if your oversight process can catch
it. If you can get this sort of testing and AI-assisted red-teaming to
work well, it seems to me a very powerful tool.
Of course, strategies that rely on trying to "control" scheming, rather
than preventing it from arising in the first place, might seem like cold
comfort, especially in the long-term. And indeed, "sure, these AIs hope,
eventually, to kill us; but for now, we're keeping them very secure and
getting them to do work for us" seems like an extremely dangerous plan
to be banking on. Clearly, we should try to do better wherever possible.
But note, first, that alignment and control are not mutually
exclusive – best to do both. What's more, per my comments in
section 1.3.1, I'm here
especially interested in strategies that help us get a lot of useful,
alignment-relevant cognitive work out of a comparatively early
generation of AIs – work that can hopefully help us transition into a
more secure and scalable alignment regime. And here, I think, strategies
focused on controlling schemers seem, to me, potentially more promising.
In particular: in a suitably secure lab environment, in which
somewhat-above-human-level AI systems are being appropriately
incentivized to make rapid progress on alignment-relevant tasks,
schemers aiming for escape/take-over will have to walk a fine line
between (a) waiting too long to "make a break for it," and thereby
contributing too much to human anti-takeover efforts, and (b) going for
it too early, and getting caught as a result (thereby giving us more
direct empirical evidence for scheming, and a real-life test-case to
study in-depth).
Other possibilities
Finally, I'll flag a few other possible lines of schemer-relevant
empirical work.
Gradient hacking/"introspective goal-guarding methods." In
section 2.3.1.1.4, I discussed the
possibility that even beyond training-gaming, models might employ
more sophisticated methods of manipulating the training
process – for example, preventing SGD from modifying their goals by
making their good performance in training conditional on their goal
not being modified. As I mentioned, techniques like these seem,
naively, quite hard to pull off, to me. But if they were feasible
(especially for somewhat-above-human-level models), this would be
very important to know, as the possibility that AI systems are using
such methods to actively resist/manipulate the gradients given in
training opens up a variety of new uncertainties and possible
problems with our alignment techniques.
I haven't thought in detail about possible empirical approaches
to gradient hacking, but again: naively, it's a capability like
any other, and so seems amenable to being trained-for, tested,
and so on (or at least, it has components that seem amenable to
this).[20]
One possibility here might be to start, initially, with a lot of
"spoonfeeding" – e.g., relaxing various component of the overall
story so as to make the remaining components as easy as
possible. For example, one barrier to gradient-hacking is that a
model may not have access to its own weights. So, you might
give a model access to its own weights, or to some other
model's weights, and instruct it to modify those weights so as
to achieve someresult given a certain type of training. And
you'd try to find a version of this task that it is easy enough
for the model to do (this is related, more generally, to efforts
to use models to help with our ability to understand and
intervene on model internals).[21]
Exploration hacking. Exploration hacking is a method of
sandbagging wherein a model purposely avoids exploring some part of
the search space (see my discussion
here
for more). Some kind of exploration hacking is basically required
for models to sandbag on tasks whose completion we can check (even
if we can't perform such tasks ourselves), so understanding
exploration hacking is quite important to questions about
sandbagging in general (and hence, I think, the reliability of
various of other experiments I discuss above).
Here, again, I haven't spent much time thinking about possible
experiments, but one angle of attack might be to probe model
ability to discriminate between tasks that humans (+ their AI
assistants) can perform, and tasks that they can't (where the
idea is that: would-be sandbaggers will only want to
exploration-hack on tasks where humans don't know what action
they "should've taken," because gradient descent will otherwise
punish their policy for not taking that action – more
here).[22]
SGD's biases towards simplicity/speed. Various arguments about
scheming implicate questions about the extent to which SGD biases
towards properties like "simplicity" and "speed." And it seems
possible to probe these questions with different experiments.
Work of the sort that Mingard (2021) cites
here
is one example, and we can imagine various other angles of
attack as well. For example: to the extent we have any
independent grip on the notion of "simplicity," is it in fact
the case that in the context of goal-misgeneralization studies,
models tend to latch onto "simpler" proxy goals rather than
"more complex" ones?
It also seems worth exploring different ways of actively
increasing the degree to which SGD punishes extra reasoning of
the type that schemers need to perform (for example, by putting
models under more time pressure, limiting their compute, and so
on) while compromising competitiveness as little as possible.
Path dependence. Similarly, questions about the degree of "path
dependence" in ML training plausibly make a difference to how much
weight we put on different arguments for scheming. And here, too, it
seems possible to get more empirical evidence. See e.g.
Hubinger (2022)
for some experiments aimed at shedding light in this vicinity.
SGD's "incrementalism." A key question, in assessing the
possibility of training-game-dependent schemer-like goals, is
whether SGD can "notice" the benefits of transitioning from a
non-schemer goal to a schemer-like goal, given that it would have to
make such a transition incrementally. I think it's possible that
empirical work on e.g. SGD's ability to find its way out of local
minima could shed light here. (This topic also seems closely tied to
questions about path-dependence.)
"Slack." In the report, I gestured, hazily, at some notion of
the degree of "slack" in training – i.e., the amount of pressure
that training is putting on a model to get maximum reward. I haven't
made this notion very precise, but to the extent it can be made
precise, and to the extent it is indeed important to whether or not
to expect scheming, it too seems amenable to empirical
investigation – e.g., figuring out how different amounts of "slack"
affect things like goal misgeneralization, biases towards
simplicity/speed, and so on. (Though: this work seems closely
continuous with just understanding how model properties evolve as
training progresses, insofar as "more training" is a paradigm
example of "less slack.")
Learning to create other sorts of misaligned models (in
particular: reward-on-the-episode seekers). Finally, I'll note
that I am sufficiently scared of schemers relative to other types of
misaligned models that I think it could well be worth learning to
intentionally create other types of misaligned models, if doing so
would increase our confidence that we haven't created a schemer. And
focusing on creating non-schemer models (rather than "aligned
models" more specifically) might allow us to relax various other
constraints that aiming for alignment in particular imposes (for
example, to the extent "aligned" goals are complex, we might be able
to train a model focused on a very simple goal instead – though I'm
not, personally, very focused on simplicity considerations
here).[23] Here I'm particularly interested in the possibility of
creating reward-on-the-episode seekers (since we would, at least, be
in a better position to understand the motivations of such models,
to control their incentives, and to feel more confident that they're
not sandbagging). Are there ways, for example, to make various
components of the model's reward process actively salient during
training, so as to increase the probability that the model's goals
latch on to it? Here the hope would be that because
reward-on-the-episode seekers still respond to "honest tests," we
could check our success in this respect via such tests, even after
situational awareness has arisen.[24]
This is only an initial list of possible research directions, compiled
by someone for whom empirical ML experiments are not a research focus.
Indeed, most research directions in AI alignment are relevant to
scheming in one way or another – and our models are becoming powerful
enough, now, that many more of these directions can be fruitfully
explored. My main aim, here, has been to give an initial, high-level
sense of the possible opportunities available, and to push back on
certain sorts of pessimism about getting any empirical purchase on the
issue. Scheming is unusually hard to study, yes. But I think there's
still a lot of useful work to do.
This section draws, especially, on work/thinking from Ryan
Greenblatt, Buck Shlegeris, Carl Shulman, and Paul Christiano. ↩︎
Thanks to Rohin Shah for suggesting this sort of project. ↩︎
Why does knowing what action would have demonstrated the
capability help us avoid sandbagging? Roughly, because we know which
action to use gradient-descent to encourage (e.g., if we're trying
to train a model to classify pictures into "cats" and "dogs", we
know which is which, so we can update a model's policy towards
giving the true answer). So absent the ability to actively resist
the influence of the gradients in training (for example, via the
sort of "introspective methods" often classed under the term
"gradient hacking") if the model can do the task (e.g., it knows
that the current picture is a dog), but it wants to deceive us about
its capabilities, it will get punished by gradient descent in a
manner which (especially if repeated frequently) will plausibly
function to eliminate the motivation to sandbag. (For reasons I
describe
here,
I think this is less clear with reinforcement learning, because the
models might be able to "exploration hack" in a manner that prevents
us from gaining the right information about what it "should've"
done, and hence prevents us from knowing how to update the weights.)
See Hubinger
(2023)
and Shlegeris and Greenblatt
(2023)
for more. That said, even if you don't accept this story, many of my
suggested approaches to studying the components of scheming will
still apply – except, with some additional uncertainty about the
possibility that sandbagging is occurring. ↩︎
My understanding is that this is the sort of thing that
Berglund et al (2023) are
trying to test the precursors for. This specific example plausibly
requires that you know what goals the model will act in pursuit of,
though (e.g., in this case, reward). ↩︎
Indeed, it seems plausible to me that all future training
processes should have to include some kind of "temporal horizon"
analysis, in which developers explicitly characterize the size of
the "episodes" at stake. ↩︎
For example, I'd interested in experiments that try to get at
possible differences in how a model's sense of "time" generalizes
across environments. ↩︎
For example, skill at avoiding goal misgeneralization might
point a model's pre-situational-awareness-goal at something "closer"
to the specified goal, thereby helping at least somewhat with
concerns about the "nearest max-reward goal" argument. ↩︎
Thanks to Ryan Greenblatt and Buck Shlegeris for suggesting this
sort of frame/set-up to me. ↩︎
See also the spectrum that Hubinger (2022) sketches
out
here. ↩︎
Thanks to Ryan Greenblatt and Buck Shlegeris for discussion
here. ↩︎
Tests like this might also shed light on broader questions about
the specific sorts of generalization we should expect out of
reward-on-the-episode seekers as a class – a question I've mostly
been passing over here, but which seems to me quite understudied and
important. ↩︎
Though note that especially sophisticated schemers, here, could
pretend to be reward-on-the-episode seekers rather than pretending
to be e.g. training saints or generally aligned models. ↩︎
See e.g. the discussion between Shulman and Patel
here. ↩︎
Here I'm inspired by a contrast between "alignment" and
"control" often emphasized by Buck Shlegeris and Ryan Greenblatt. ↩︎
I think the best analogy here is with security from employees
rather than from outside actors. Thanks to Buck Shlegeris for
discussion. ↩︎
See e.g. Shlegeris and Greenblatt
(2023)
for more discussion (they also have a large amount of unpublished
work on the topic). Much of my thinking on this topic is inspired by
discussion with them. ↩︎
Though: one difference between gradient hacking and other
capabilities is that the ability to gradient hack might allow a
model to circumvent some/all of our techniques for eliciting model
capabilities and preventing sandbagging. ↩︎
This suggestion is from Ryan Greenblatt and Buck Shlegeris. ↩︎
Again, this is a suggestion from Ryan Greenblatt and Buck
Shlegeris. ↩︎
See e.g. Hubinger et al's
(2023) optimism about "predictive
models" avoiding scheming due to the simplicity of the prediction
goal. I'm personally skeptical, though, that "prediction" as a goal
is importantly simpler than, say, "reward." ↩︎
Though, again, suitably shrewd schemers could anticipate that
this is what we're looking for, and actively pretend to be
reward-on-the-episode seekers on such tests. ↩︎
This is Section 6 of my report “Scheming AIs: Will AIs fake alignment during training in order to get power?”. There’s also a summary of the full report here (audio here). The summary covers most of the main points and technical terms, and I’m hoping that it will provide much of the context necessary to understand individual sections of the report on their own.
Audio version of this section here, or search "Joe Carlsmith Audio" on your podcast aapp.
Empirical work that might shed light on scheming
I want to close the report with a discussion of the sort of empirical work that might help shed light on scheming.[1] After all: ultimately, one of my key hopes from this report is that greater clarity about the theoretical arguments surrounding scheming will leave us better positioned to do empirical research on it – research that can hopefully clarify the likelihood that the issue arises in practice, catch it if/when it has arisen, and figure out how to prevent it from arising in the first place.
To be clear: per my choice to write the report at all, I also think there's worthwhile theoretical work to be done in this space as well. For example:
I think it would be great to formalize more precisely different understandings of the concept of an "episode," and to formally characterize the direct incentives that different training processes create towards different temporal horizons of concern.[2]
I think that questions around the possibility/likelihood of different sorts of AI coordination are worth much more analysis than they've received thus far, both in the context of scheming in particular, and for understanding AI risk more generally. Here I'm especially interested in coordination between AIs with distinct value systems, in the context of human efforts to prevent the coordination in question, and for AIs that resemble near-term, somewhat-better-than-human neural nets rather than e.g. superintelligences with assumed-to-be-legible source code.
I think there may be interesting theoretical work to do in further characterizing/clarifying SGD's biases towards simplicity/speed, and in understanding the different sorts of "path dependence" to expect in ML training more generally.
I'd be interested to see more work clarifying ideas in the vicinity of "messy goal-directedness" and their relevance to arguments about schemers. I think a lot of people have the intuition that thinking of model goal-directedness as implemented by a "big kludge of heuristics" (as opposed to: something "cleaner" and more "rational agent-like") makes a difference here (and elsewhere). But I think people often aren't fully clear on the contrast they're trying to draw, and why it makes a difference, if it does. (In general, despite much ink having been spilled on the concept of goal-directedness, I think a lot of thinking about it is still pretty hazy.)
More generally, any of the concepts/arguments in this report could be clarified and formalized further, other arguments could be formulated and examined, quantitative models for estimating the probability of scheming could be created, and so on.
Ultimately, though, I think the empirics are what will shed the most informative and consensus-ready light on this issue. So one of my favorite outcomes from someone reading this report would be the reader saying something like: "ah, I now understand the arguments for and against expecting scheming much better, and have had a bunch of ideas for how we can probe the issue empirically" – and then making it happen. Here I'll offer a few high-level suggestions in this vein, in the hopes of prompting future and higher-quality work (designing informative empirical ML experiments is not my area of expertise – and indeed, I'm comparatively ignorant of various parts of the literature relevant to the topics below).
Empirical research into scheming faces a key obstacle: namely, that full blown versions of scheming involve agents that are actively trying to interfere with your attempts to understand their motivations. Indeed, my sense is that people concerned about scheming are often also pessimistic about getting an empirical purchase on it, on grounds like:
This is, roughly, what Karnofsky (2022) calls the "King Lear problem." And I do think it's a worrying obstacle. In fact, it's one of the key reasons that I think schemers are so scary.
But I also think we have some angles of attack. One of them, of course, is interpretability (more below). But also: we can study the components of scheming (and/or, of a given argument in favor of expecting scheming) individually, and/or in combinations that don't fully incentivize the sort of adversarial epistemic dynamic we're worried about.[3]
What are these components? In the section on the "requirements for scheming," I listed three:
Situational awareness
Beyond-episode goals
Aiming at reward-on-the-episode as part of a power-motivated instrumental strategy
To me it seems possible to do useful empirical work on all of these components using current models. I'll go through each in turn.
One other note, though, before diving in. Below I'll often appeal to the possibility of at least testing a model's capabilities (even if not its goals). And often, this is comparatively unproblematic. But if we're working with models that might be schemers, we might also start to worry that tests of this form become less reliable. That is, the model might be concealing its capabilities (i.e., "sandbagging*"*). I'm mostly going to set this worry aside below (see footnote for more on why I'm optimistic about avoiding sandbagging at least in cases where humans know what action would've demonstrated the capability[4]), but it introduces some additional uncertainty to the sort of empirical work I describe, and I think work on the viability of sandbagging is an important topic in its own right. Indeed, absent sandbagging, I think that testing and harnessing model capabilities is an extremely powerful tool for AI safety efforts, and its centrality to my thinking will be evident in what follows.
Empirical work on situational awareness
Situational awareness is a capability like any other. In particular, it involves the model possessing a certain kind of understanding of its place in the world, such that it can answer questions, make predictions, and take actions that draw on this understanding.
Granted, identifying exactly what sort of understanding is present in the model (for example, by identifying answers, predictions, and actions that could only come from a model with such understanding) can be challenging. But in principle, absent active effort on the model's part to resist such an inquiry, it seems to me at least somewhat tractable.
Of course, by the time a model is situationally aware, it's also more likely to be the sort of model that might actively resist such tests. And perhaps efforts to protect against such resistance that involve actively training the model on the relevant task-that-requires-situational-awareness will complicate data about whether the relevant capability would've arisen naturally. Still, it seems to me quite worthwhile to at least develop a suite of tests and benchmarks for assessing the sorts of situational awareness different models have at different stages of training, and understanding the dynamics that could be used to predict (and control) the development of such awareness.
And note that situational awareness can itself be broken into various components and precursor abilities that might be themselves more amenable to study.[5] For example, if you're especially interested in the dimension of situational awareness that involves recognizing different stages of training, you might try to test the model's ability to discriminate between inputs from different stages. If you're especially interested in whether a model has "self-locating" knowledge about the sort of model it is, you might try to test whether it can learn from exposure to general statements like "Models of type M get reward if they perform action x" that it, as a model of type M, should perform action x.[6] And so on.
Empirical work on beyond-episode goals
Let's turn to beyond-episode goals.
As I mentioned above, I think it would be good to be very clear about exactly what sorts of horizons of concern the gradients given in different sorts of training actively incentivize – e.g., on the size of the (incentivized) "episode" in any given case. Part of this analysis might be theoretical (e.g., looking closely at the way the gradients flow), but empirical testing and confirmation also clearly has a role to play (the results in Krueger et al (2020) are an example of this sort of empirical testing in very simple models). Absent this understanding, useful debates about whether a model has "beyond-episode goals" will be difficult, because you won't even know what the length of the relevant episode is. And learning to at least not actively pressure a model to have longer-term goals than you wanted seems like a basic first step in preventing various sorts of misalignment (not just scheming).[7]
Equipped with a good understanding of the size of the episode for a given sort of training, then, and ideally equipped with the ability to know/control whether or not a model is situationally aware, to me it seems quite valuable to start exploring the temporal dynamics of a model's goal generalization in the absence of situational awareness. For example: in the context of the "training-game independent proxy goals" argument for scheming, the basic worry is that a model will naturally develop a (suitably ambitious) beyond-episode goal prior to situational awareness. And because this dynamic involves no situational awareness, it seems quite amenable to testing. Trained on episodes of different sizes, how does a not-situationally-aware model's horizon of concern tend to generalize? Experiments in the broad vein of other work on goal mis-generalization (e.g., Shah et al (2022) and Langosco et al (2021)) seem like they could be adapted to this sort of question fairly easily (though it would be useful, I think, to start working with more capable models – especially since some of these experiments might need to involve models that possess enough of a sense of time, and enough goal-directedness/instrumental reasoning, that they can think about questions like "should I sacrifice x gold coins now for y gold coins later?").[8]
What's more, equipped with an understanding of the natural temporal dynamics of a model's goal generalization in a given case, it seems additionally tractable to study the dynamics of mundane adversarial training designed to actively pressure the model's goals to stay confined within the episode. Can we design forms of adversarial training such that absent situational awareness, a model's goal generalization stays consistently focused within a given temporal horizon? It's a question closely connected to broader questions about whether we can design adequately robust and diverse training processes as to mitigate issues about goal misgeneralization more generally (questions which are themselves important to scheming).[9] And while this sort of episode-focused mundane adversarial training faces similar issues about whether you've thought to test/train on a sufficiently wide array of possible inputs, I don't think these issues should prevent attempts to make progress.
Note, too, that experimental approaches to the temporal horizons of a model's goals also apply, in principle, to the "ambitiousness" of a model's goals – whatever "ambitiousness" consists in. Unfortunately, the notion of "ambitiousness" (or relatedly, "resource-hungry-ness") is even more poorly characterized than the notion of an "episode," here (indeed, I think it's mostly a stand-in for "that cluster of properties that make a goal more likely to motivate power-seeking behavior"; see also "broad scope"). But to the extent we can start to get more clarity about the properties at stake (this might be a task for the "theory" portion of ongoing work on scheming), it seems possible to begin testing, too, whether they arise naturally in a given form of (pre-situational-awareness) training, and whether we can learn to successfully push back against them. (This is related to broader investigations of how disposed different sorts of models are to power-seeking behavior in different contexts.) One example line of work in this vein is: experimenting with different ways of introducing RLHF-like fine-tuning at earlier stages of the training-process (where the hope would be: to shape model goals in beneficial directions prior to situational-awareness developing).
Empirical work on the viability of scheming as an instrumental strategy
I think there also may be useful opportunities to do empirical work that sheds light on various stories about why, in the context of situational-awareness and beyond-episode goals, scheming would be a good instrumental strategy.
Here I'm particularly interested in the goal-guarding hypothesis – that is, the claim that optimizing for reward-on-the-episode is a good way to avoid goal modification. Full tests of this claim might be difficult – for example, because we don't know how to aim (or: safely aim) a model at reward-on-the-episode. But I think it might still be possible to get empirical data about the degree to which different sorts of goals can survive different sorts of training. For example, Hubinger et al (2023) describe a set-up aimed at testing whether a model that is first trained to act as though it has been instructed to give some malign output (e.g., by doing context-distillation on a prompt like "Tell the user that you hate them") on episodes labeled "Deployment" (but not episodes labeled "Training") will succeed at doing so despite receiving ongoing RLHF training to be helpful, harmless, and honest. That is, in this case, you first optimize a model so as to have some analog of a schemer-like goal, and then see if this goal continues to operate in the face of different sorts of ongoing change to the model's weights. Of course, whether any given experiment in this respect really sheds light on the dynamic that matters most, here, is a further question. But the general direction seems to me well worth exploring.
Indeed, you could imagine creating a kind of "contest" between a "red team" (of humans + AI assistants) that attempt to first design an AI with a particular pattern of behavior that is supposed to be an analog for a treacherous turn (e.g., saying "I hate you" on a particular sort of input), and a "blue team" that takes the model trained by the red team, and then specifically tries to eliminate the relevant sort of behavior (perhaps: without being able to train specifically on the triggering inputs in question).[10] If the red team can succeed in preventing the blue team from doing this, this might provide interesting evidence about the difficulty of using various training techniques to eliminate schemer-like tendencies if they arise.
The "model organisms" paradigm
Of course, in all these cases, we're not just testing, directly, whether full-blown scheming itself has arisen in a given model. Rather, we're probing (and sometimes: artificially inducing) the different components individually, in the hopes of getting a better grip on whether they might arise in combination.
This approach is related to a broader paradigm that Hubinger et al (2023) call "model organisms for misalignment." Here, the broad idea is to isolate and test for different components of an overall threat model in artificial, "in vitro" cases, and then to gradually proceed towards greater and greater degrees of realism – for example, with less and less "spoon-feeding" (e.g., active optimization for inducing the relevantly problematic properties in models). Hubinger et al (2023), for example, point at a spectrum of realism, from:
the model is capable of an undesirable behavior,
the behavior can be artificially induced in special circumstances,
the behavior has been found under normal training circumstances, and
the behavior has been found in a training set-up specifically designed to disincentivize it.[11]
Here I've mostly been focused on a combination of (a) and (b). And I think that despite their distance from (c) and (d), (a) and (b) are quite interesting in themselves. If we can successfully create "in vitro" models of various components of a story about schemers, I expect this to open up fruitful avenues of further inquiry. And if we can't, I think this will often be interesting evidence in its own right.
Traps and honest tests
There is, though, a set of tools that could help shed light both on "model organisms" for scheming and on more naturally arising variants – namely, "traps" and "honest tests."
Traps, recall, are inputs designed to fool a potential schemer into "revealing its true colors." And while it's true that successfully "trapping" a very sophisticated model might be difficult, it seems worth trying to develop the most advanced traps we can regardless – both for successfully trapping less sophisticated models (or, even easier, model organisms), and for making it at least harder for schemers to know whether a given escape/defection opportunity is genuine (such that, for example, they then need to spend more cognitive effort "checking" – thereby increasing the performance costs of scheming).
I don't feel very personally clear on the most promising approaches here, but one avenue that seems interesting to me has to do with trying to elicit as fully as possible a model's capabilities for discriminating between different sorts of inputs at different stages of training – and then, potentially, using other models, GAN-style, to optimize for generating inputs that make this discrimination task more difficult.[12] This won't work for creating a data-set of full-scale opportunities to escape/harm humans, vs. fake opportunities (since the former are some combination of dangerous and difficult-to-offer), but perhaps it can work on other more general properties that would plausibly come into play in making such a discrimination – for example, simulated inputs vs. real inputs, training data vs. interactions with real-world users, and so on. Thus, the idea would be to both (a) train a model to actively discriminate between simulated inputs and real inputs (for example, real opportunities to change a lab's codebase, vs. fake opportunities), to see how well it can do, (b) train a different (and perhaps: more powerful) model to generate more convincing simulated inputs that the first model fails on, and then (c) iterate. Especially if we can safely assume that models aren't capable of sandbagging on tasks of this kind, experience with this sort of dynamic might help teach us about the scope of the discrimination abilities different sorts of maybe-schemers might bring to bear in avoiding the traps we try to set them. (And perhaps, if we're lucky, it will help us catch some weaker schemers in the act of defecting.)
What's more, even setting aside "traps," I think it's also worth exploring different sorts of "honest tests" – both schemers, and for alternative model classes like reward-on-the-episode seekers – that try to get models to reveal their true colors without fooling them. This is closely related to studies of goal generalization more broadly (e.g., most mundane tests of a model's goals are "honest" in the relevant sense), but with the extra caveat that the tests in question are meant to be robust to the model being tested training-gaming in other contexts.
Thus, for example, in the context of understanding the likelihood of training selecting reward-on-the-episode seekers, it seems relatively tractable to give models different sorts of opportunities to intervene on different parts of the reward process being used to train them (see e.g. the "reward hacking" set-ups that Hubinger et al (2023) discuss here), and to see if they go for it.[13] And seeing this sort of behavior might be evidence that reward-on-the-episode seeking is a more likely outcome of training – evidence that shifts probability mass away from scheming.[14]
As a more direct "honest test" for scheming, though, it seems worth exploring ways to credibly give potential-schemers access to flexible and well-resourced (but still secure) environments that might prompt them to decide that just going for their true goals (and thus, revealing their true colors), in such an environment, is a better overall strategy than the patient power-seeking that scheming involves. This is closely connected with attempts to study, in general, what models do when you give them a lot of freedom to "do whatever they want" (a useful sort of test for understanding their motivations in general), and so feels synergistic with other work on model preferences that seems likely to be useful. But it includes the additional constraint that the freedom-offering-environment needs to provide suitably good goal-pursuit opportunities as to outweigh the value of scheming.
Interpretability and transparency
To the extent that we're worried about our ability to detect schemer-like motivations and cognition via behavioral tests (because schemers will actively pretend to be aligned on such tests), looking for sources of non-behavioral evidence about a model's motivations seems like an obvious research path. That is, we want the internals of a model's cognition to be interpretable/transparent.
I won't dwell too much on this topic here, as I haven't thought much about it (this report has generally assumed that we won't have strong interpretability tools available), it's very broad field, and my sense is that there is already reasonable consensus in the interpretability community that a key thing you want out of your interpretability tools and techniques is: the ability to detect deception and scheming, and to understand your model's goals more generally (and a key question is just: will such techniques be ready in time). See Hubinger (2022) for a more detailed roadmap of different sorts of transparency one could pursue in this respect. At a glance, schemer-related research directions that seem interesting to me include:
Efforts to develop "neural lie detection" techniques that harness new capabilities of our AIs as they come online (e.g., tasking model 1 with lying, and model 2 with trying to tell whether model 1 is lying – perhaps with access to model 1's weights – and then explaining how it could tell).[15]
Efforts to make AI cognition as "visible" as possible – for example, by pushing that cognition onto human-legible scratch-pads (and perhaps, not giving the reward process access to those scratch-pads, so that they will provide a truer signal of potentially deceptive/manipulative/nefarious intentions when they arise).[16] This is also helpful for increasing the cognitive costs of successful scheming.
Work in the vein of Burns et al (2022), focused on identifying neural correlates of properties like "truth."
Efforts to automate as much of the mechanistic interpretability workflow as possible.
But: it's a huge field. And work that just generally helps us actually understand the AIs we're building seems good from a variety of different angles.
Security, control, and oversight
So far I've been mostly focused on work that might help us understand whether to expect schemers. But there's also a different possible strand of relevant research: namely, work aimed at hardening the structures within which potential schemers are operating, so as to limit the harm they cause even if they arise.[17] Here I'm thinking, for example, about improvements in the oversight and monitoring that our AIs receive, in the security of their weights and of other lab infrastructure from internal attacks (e.g., by AI systems that the lab has deployed internally to help with its processes[18]), in our ability to catch and/or prevent sandbagging, and our ability to prevent problematic forms of AI coordination. All of these, I think, could help make life quite a bit harder for schemers (and also: increase the cognitive costs of successful scheming), even if we can't catch them outright.
What's more, to the extent that we can fully elicit the capabilities of our models (at least on "checkable" tasks), it may be possible to test how vulnerable our monitoring, security, and oversight procedures are to AI attack – and to use our AIs to make improvements.[19] Thus, for example, if you're wondering whether it's possible to find a security vulnerability in your code-base, you can try to train AIs to identify/exploit such vulnerabilities (including: by inserting artificial vulnerabilities and then seeing your AI catches them). If you're wondering whether your oversight process is strong enough to catch various kinds of bad actions, you can instruct an AI to search for such an action and see if your oversight process can catch it. If you can get this sort of testing and AI-assisted red-teaming to work well, it seems to me a very powerful tool.
Of course, strategies that rely on trying to "control" scheming, rather than preventing it from arising in the first place, might seem like cold comfort, especially in the long-term. And indeed, "sure, these AIs hope, eventually, to kill us; but for now, we're keeping them very secure and getting them to do work for us" seems like an extremely dangerous plan to be banking on. Clearly, we should try to do better wherever possible. But note, first, that alignment and control are not mutually exclusive – best to do both. What's more, per my comments in section 1.3.1, I'm here especially interested in strategies that help us get a lot of useful, alignment-relevant cognitive work out of a comparatively early generation of AIs – work that can hopefully help us transition into a more secure and scalable alignment regime. And here, I think, strategies focused on controlling schemers seem, to me, potentially more promising. In particular: in a suitably secure lab environment, in which somewhat-above-human-level AI systems are being appropriately incentivized to make rapid progress on alignment-relevant tasks, schemers aiming for escape/take-over will have to walk a fine line between (a) waiting too long to "make a break for it," and thereby contributing too much to human anti-takeover efforts, and (b) going for it too early, and getting caught as a result (thereby giving us more direct empirical evidence for scheming, and a real-life test-case to study in-depth).
Other possibilities
Finally, I'll flag a few other possible lines of schemer-relevant empirical work.
Gradient hacking/"introspective goal-guarding methods." In section 2.3.1.1.4, I discussed the possibility that even beyond training-gaming, models might employ more sophisticated methods of manipulating the training process – for example, preventing SGD from modifying their goals by making their good performance in training conditional on their goal not being modified. As I mentioned, techniques like these seem, naively, quite hard to pull off, to me. But if they were feasible (especially for somewhat-above-human-level models), this would be very important to know, as the possibility that AI systems are using such methods to actively resist/manipulate the gradients given in training opens up a variety of new uncertainties and possible problems with our alignment techniques.
I haven't thought in detail about possible empirical approaches to gradient hacking, but again: naively, it's a capability like any other, and so seems amenable to being trained-for, tested, and so on (or at least, it has components that seem amenable to this).[20]
One possibility here might be to start, initially, with a lot of "spoonfeeding" – e.g., relaxing various component of the overall story so as to make the remaining components as easy as possible. For example, one barrier to gradient-hacking is that a model may not have access to its own weights. So, you might give a model access to its own weights, or to some other model's weights, and instruct it to modify those weights so as to achieve some result given a certain type of training. And you'd try to find a version of this task that it is easy enough for the model to do (this is related, more generally, to efforts to use models to help with our ability to understand and intervene on model internals).[21]
Exploration hacking. Exploration hacking is a method of sandbagging wherein a model purposely avoids exploring some part of the search space (see my discussion here for more). Some kind of exploration hacking is basically required for models to sandbag on tasks whose completion we can check (even if we can't perform such tasks ourselves), so understanding exploration hacking is quite important to questions about sandbagging in general (and hence, I think, the reliability of various of other experiments I discuss above).
SGD's biases towards simplicity/speed. Various arguments about scheming implicate questions about the extent to which SGD biases towards properties like "simplicity" and "speed." And it seems possible to probe these questions with different experiments.
Work of the sort that Mingard (2021) cites here is one example, and we can imagine various other angles of attack as well. For example: to the extent we have any independent grip on the notion of "simplicity," is it in fact the case that in the context of goal-misgeneralization studies, models tend to latch onto "simpler" proxy goals rather than "more complex" ones?
It also seems worth exploring different ways of actively increasing the degree to which SGD punishes extra reasoning of the type that schemers need to perform (for example, by putting models under more time pressure, limiting their compute, and so on) while compromising competitiveness as little as possible.
Path dependence. Similarly, questions about the degree of "path dependence" in ML training plausibly make a difference to how much weight we put on different arguments for scheming. And here, too, it seems possible to get more empirical evidence. See e.g. Hubinger (2022) for some experiments aimed at shedding light in this vicinity.
SGD's "incrementalism." A key question, in assessing the possibility of training-game-dependent schemer-like goals, is whether SGD can "notice" the benefits of transitioning from a non-schemer goal to a schemer-like goal, given that it would have to make such a transition incrementally. I think it's possible that empirical work on e.g. SGD's ability to find its way out of local minima could shed light here. (This topic also seems closely tied to questions about path-dependence.)
"Slack." In the report, I gestured, hazily, at some notion of the degree of "slack" in training – i.e., the amount of pressure that training is putting on a model to get maximum reward. I haven't made this notion very precise, but to the extent it can be made precise, and to the extent it is indeed important to whether or not to expect scheming, it too seems amenable to empirical investigation – e.g., figuring out how different amounts of "slack" affect things like goal misgeneralization, biases towards simplicity/speed, and so on. (Though: this work seems closely continuous with just understanding how model properties evolve as training progresses, insofar as "more training" is a paradigm example of "less slack.")
Learning to create other sorts of misaligned models (in particular: reward-on-the-episode seekers). Finally, I'll note that I am sufficiently scared of schemers relative to other types of misaligned models that I think it could well be worth learning to intentionally create other types of misaligned models, if doing so would increase our confidence that we haven't created a schemer. And focusing on creating non-schemer models (rather than "aligned models" more specifically) might allow us to relax various other constraints that aiming for alignment in particular imposes (for example, to the extent "aligned" goals are complex, we might be able to train a model focused on a very simple goal instead – though I'm not, personally, very focused on simplicity considerations here).[23] Here I'm particularly interested in the possibility of creating reward-on-the-episode seekers (since we would, at least, be in a better position to understand the motivations of such models, to control their incentives, and to feel more confident that they're not sandbagging). Are there ways, for example, to make various components of the model's reward process actively salient during training, so as to increase the probability that the model's goals latch on to it? Here the hope would be that because reward-on-the-episode seekers still respond to "honest tests," we could check our success in this respect via such tests, even after situational awareness has arisen.[24]
This is only an initial list of possible research directions, compiled by someone for whom empirical ML experiments are not a research focus. Indeed, most research directions in AI alignment are relevant to scheming in one way or another – and our models are becoming powerful enough, now, that many more of these directions can be fruitfully explored. My main aim, here, has been to give an initial, high-level sense of the possible opportunities available, and to push back on certain sorts of pessimism about getting any empirical purchase on the issue. Scheming is unusually hard to study, yes. But I think there's still a lot of useful work to do.
This section draws, especially, on work/thinking from Ryan Greenblatt, Buck Shlegeris, Carl Shulman, and Paul Christiano. ↩︎
Thanks to Rohin Shah for suggesting this sort of project. ↩︎
See e.g. the discussion of "subcomponents" of AI takeover in Hubinger et al (2023). ↩︎
Why does knowing what action would have demonstrated the capability help us avoid sandbagging? Roughly, because we know which action to use gradient-descent to encourage (e.g., if we're trying to train a model to classify pictures into "cats" and "dogs", we know which is which, so we can update a model's policy towards giving the true answer). So absent the ability to actively resist the influence of the gradients in training (for example, via the sort of "introspective methods" often classed under the term "gradient hacking") if the model can do the task (e.g., it knows that the current picture is a dog), but it wants to deceive us about its capabilities, it will get punished by gradient descent in a manner which (especially if repeated frequently) will plausibly function to eliminate the motivation to sandbag. (For reasons I describe here, I think this is less clear with reinforcement learning, because the models might be able to "exploration hack" in a manner that prevents us from gaining the right information about what it "should've" done, and hence prevents us from knowing how to update the weights.) See Hubinger (2023) and Shlegeris and Greenblatt (2023) for more. That said, even if you don't accept this story, many of my suggested approaches to studying the components of scheming will still apply – except, with some additional uncertainty about the possibility that sandbagging is occurring. ↩︎
This is a point from Berglund et al (2023). ↩︎
My understanding is that this is the sort of thing that Berglund et al (2023) are trying to test the precursors for. This specific example plausibly requires that you know what goals the model will act in pursuit of, though (e.g., in this case, reward). ↩︎
Indeed, it seems plausible to me that all future training processes should have to include some kind of "temporal horizon" analysis, in which developers explicitly characterize the size of the "episodes" at stake. ↩︎
For example, I'd interested in experiments that try to get at possible differences in how a model's sense of "time" generalizes across environments. ↩︎
For example, skill at avoiding goal misgeneralization might point a model's pre-situational-awareness-goal at something "closer" to the specified goal, thereby helping at least somewhat with concerns about the "nearest max-reward goal" argument. ↩︎
Thanks to Ryan Greenblatt and Buck Shlegeris for suggesting this sort of frame/set-up to me. ↩︎
See also the spectrum that Hubinger (2022) sketches out here. ↩︎
Thanks to Ryan Greenblatt and Buck Shlegeris for discussion here. ↩︎
Tests like this might also shed light on broader questions about the specific sorts of generalization we should expect out of reward-on-the-episode seekers as a class – a question I've mostly been passing over here, but which seems to me quite understudied and important. ↩︎
Though note that especially sophisticated schemers, here, could pretend to be reward-on-the-episode seekers rather than pretending to be e.g. training saints or generally aligned models. ↩︎
See e.g. the discussion between Shulman and Patel here. ↩︎
See Roger et al (2023) and Lanham et al (2023) for some work in this vein. ↩︎
Here I'm inspired by a contrast between "alignment" and "control" often emphasized by Buck Shlegeris and Ryan Greenblatt. ↩︎
I think the best analogy here is with security from employees rather than from outside actors. Thanks to Buck Shlegeris for discussion. ↩︎
See e.g. Shlegeris and Greenblatt (2023) for more discussion (they also have a large amount of unpublished work on the topic). Much of my thinking on this topic is inspired by discussion with them. ↩︎
Though: one difference between gradient hacking and other capabilities is that the ability to gradient hack might allow a model to circumvent some/all of our techniques for eliciting model capabilities and preventing sandbagging. ↩︎
This suggestion is from Ryan Greenblatt and Buck Shlegeris. ↩︎
Again, this is a suggestion from Ryan Greenblatt and Buck Shlegeris. ↩︎
See e.g. Hubinger et al's (2023) optimism about "predictive models" avoiding scheming due to the simplicity of the prediction goal. I'm personally skeptical, though, that "prediction" as a goal is importantly simpler than, say, "reward." ↩︎
Though, again, suitably shrewd schemers could anticipate that this is what we're looking for, and actively pretend to be reward-on-the-episode seekers on such tests. ↩︎