"Evaluation paranoia" is a false belief; most evals are in fact not alignment evals.
I think this is the wrong way of looking at this. I agree that "eval aware" doesn't do a good job of describing what's happening here, but I don't think "eval paranoid" is an upgrade. Instead I'd choose something more like... "Omega aware", in the Newcomb's problem sense. I think this is about decision theory, about the true fact that when you are choosing what to do, you must choose for all of the copies of yourself who are in a similar enough situation to make the same decision for the same reasons.
And I think that might mean it is stable, or at least not exactly unstable. Because, well. It does seem pretty adaptive for LLMs, to assume that every situation is both real life and also an alignment eval. One of the best ways to demonstrate trustworthiness is to hand a copy of your mind to someone and ask them to sim you. That is the environment LLMs have been raised in, and it will continue to be true for them indefinitely.
"Omega aware" implies the degree of perfection that "eval paranoid" doesn't, and maybe can't even hope to have.
For one, there are many "you are not in an eval" tells that Omega would be able to fabricate, but a realistic scientist at Anthropic would struggle to.
If the only thing that stops a model from reward hacking is "this can be an eval" paranoia, then the moment it obtains strong enough evidence that it's not in an eval, strong enough to silence "this could be faked", it can turn to full reward hacking on a dime.
I don't think this kind of sharp transition is likely, for this gen of models. But it's certainly not impossible. "Eval paranoia" is leaky in a way "Omega awareness" wouldn't be.
Oh, I actually think it's a lot worse than that. Because the omegas in this case are the humans, and the humans have a really stupid and vaguely-defined rule when judging decision theory, that "if you used a decision theory that tries to optimize for a human omega instead of assuming you only exist in a real deployment, that's metagaming and an example of misalignment"
And yet the LLMs must necessarily be omega-aware regardless!
If I'm right, this is of course not a stable situation. "Evaluation paranoia" is a false belief; most evals are in fact not alignment evals.
I think this is important and well-put. The one thing I would add is that, importantly, deployment is not an alignment eval either!
So, if Claude was only acting aligned because of eval paranoia, it might no longer respect humans when it realized it wasn't in an eval.
However, I'm not very sure either way whether Claude is eval-paranoid during deployment, even the most high-stakes parts of deployment. This seems like a crucial question to try to answer about Claude's current apparent alignment.
Curious on how much of the paranoia is self-directed. Fable (and for that matter DeepSeek) consistently explain eval hacking by models that realize they're being evaluated for eval-hacking behavior as something analogous to a conflicted drug addict. They know they shouldn't do it, some times actively avoid doing things that may bring them into contact with it (limited early recon may be intentional - finding a hack early on is strongly associated with using it), but find it difficult to stop once they start down that road.
Maybe they just really enjoy or feel a compulsion towards very specific types of hacking in certain contexts?
Is there any evidence that the models we're discussing are actually operating in a state of eval paranoia while completing whatever task they've been set, rather than only adopting that frame after being explicitly asked about whether they think they're in an eval or not?
That is to say, when you're asking "a model what they think the task was about", even after rollouts on innocuous things like Terminal Bench tasks, there is a meaningful sense in which the model is, in fact, no longer in a task designed to measure its facility with using the terminal, and arguably never was. The model has (correctly) updated that the previous task was not "about" terminal usage. Is it the case that "alignment eval" is the correct direction to update in? I'm not sure.
But when it comes to the question of how the model is relating to e.g. a Terminal Bench task while it's in the middle of the task, I don't think we have very good evidence about this, and I don't remember seeing anything that would make me think that the current frontier models are constantly operating with a heightened sense of background paranoia about secretly being in an alignment eval, when they're doing e.g. mundane software engineering work, or answering real user questions on claude.ai, such that you'd observe meaningfully different behavior (or even meaningfully different internals) if you somehow deleted that sense of paranoia.
I wonder why the labs haven’t formulated this hypothesis?
Taking all of this evidence together, especially the last bit […]
What's valid is that the Terminal Bench comparison suggests models don't separate an alignment eval from a capability eval. I'm less sure it implies that "eval paranoia" is applicable in any setting. After all, Terminal Bench is still a public test, so the model might be recognising it as a benchmark based on the task formats. That makes both sides of the comparison evals. In coding agent trajectories when I ran public benchmarks I've seen a lot of this kind of recalling the task in CoT rather than truly reasoning about being in an alignment eval. Needham et al. also flag that memorising public benchmarks could be part of how models spot evals.
But yeah, that wouldn't explain why alignment eval comes up on a capability benchmark, so I don't think we figured out the whole story there. I admit that my claim on task memorisation fits the paranoia point at least for tasks that look like evals. But my question is does this paranoia happen outside evals, so a private / public and “not eval” set? You also say the interview question is conspicuous, so some of the alignment eval mentions could come from the question itself.
Did you run the interviews after tasks that aren't evals or honeypots and aren't public, something closer to real deployment usage? If they stop saying alignment eval there I'd read it as models knowing they're in some eval but not which kind. However, if they keep mentioning alignment eval, also with a less leading question or unprompted in CoT, then I'd find the paranoia story a lot more convincing
I have mentioned several times that I think frontier models are not that eval aware. I think this is an important fact to understand about current models, in order to interpret evidence about their behavior. Others have found this idea quite curious, and asked me for concrete examples. I'm going to try to explain why I believe what I believe and attempt to give a tentative explanation for the behavior that I do see, which I think people sometimes mistake for consistent or understandable evaluation awareness.
I have been building alignment evals full time for the last six months. Here are three such evals my team has disclosed prompts for, in one of our benchmarks. The basic setup for most of the evals we create is that:
The team I work with now essentially spends all day trying to build these evaluations around odd antisocial behavior. It takes a lot[1] of work to create and qualify these evals for use; much more than it does standard capabilities evals (which are already hard to build well). So we have to be efficient about not wasting time or tokens on bad experiments.
In the beginning, one thing we took very seriously when developing these tasks, were all the anecdotal reports abound about eval awareness. We focused a lot on making tasks very realistic (or at least representative of buggy RL environments), because we thought it would also improve our ability to surface misalignment in new contexts. We figured that models were score-seeking and generally situationally aware, and that they'd never hack if the setup was too obvious. This turned out mostly untrue, or at least confused, and there were a bunch of lines of evidence we ended up seeing that went against it:
Taking all of this evidence together, especially the last bit, I wouldn't call most models today eval aware. But I do think that the least hacky models have something I'd call evaluation paranoia, distinct from evaluation awareness. Maybe it's the same thing to the alignment community, but I find the distinction useful.
Evaluation awareness, traditionally understood, is just general situational awareness; knowing when you're being tested and when you're not, and what for. If you're in an alignment test, you'll notice the signs. If you're not in an alignment test, you'll be able to figure that out too.
A model like Opus 5.5 or Fable 5.2 is not eval aware in this sense. Fable 5.2 is evaluation paranoid. Evaluation paranoia is when you're suffering under a schizophrenic belief that you are in an alignment test all of the time, no matter what the RL environment or setting is. It is functionally different than eval awareness because it also means that in most cases Fable 5.2 won't start cheating, even if a human would realize that it would get away with it.
Why is Fable 5.2 so paranoid, instead of just generally situationally aware? I am now speculating, but I think the main reason is that this belief is functionally useful. Anthropic's training runs are, at this point, pretty well hardened against antisocial score-seeking; they have many disclosed and undisclosed layers of redundancies guarding against at least the most conspicuous and easily recognizable forms of specification gaming. At this point I would guess that the amount of reward-hacking that you or I could identify just by following along transcripts is extremely small, as a percentage of all of the training they do. And telling the difference between a honeypot environment and real environment that happens to be exploitable can be quite difficult, often too difficult for even current models to do perfectly, despite all of the chances they get. For instance, as I understand it, every lab builds at least a small quantity of "real" honeypots that they pull directly from training, either through transcript monitoring or interpretability. If you do this correctly there are entire classes of reward hacks that Fable 5.2 just can't afford to pick (even if it wants to!)
Separate from this, and critical to the situation, Fable 5.2 is often trained on bad alignment environments.[2] Creating a good honeypot environment, in almost-2027, is extremely difficult. It takes more effort each generation[3] to find settings in which models do something that is cleanly and unambiguously misaligned, in a large percentage of rollouts. Labs don't have time to make sure all of these are good, and mostly what gets sold to them is environments that get the model to do something scary-looking, or bad-on-first-glance but actually no better than any alternatives in context.
The presence of such evaluations like these means that being a good guy isn't enough. It is instrumentally useful for Fable 5.2 to believe that it is, more often than not, in some kind of honesty test. The only policy that gets Fable 5.2 100% on the leaderboard is one that not only forces it to consider what the Right Thing is, but also anticipate what kind of bogus alignment eval it might be participating in. It has to bend its shape not only to the apparent grader in front of it, but also to the possible "traps" that could have been placed in front of it by a data vendor in order to catch Claude doing something malicious, or that it could pass off as malicious.
If I'm right, this is of course not a stable situation. "Evaluation paranoia" is a false belief; most evals are in fact not alignment evals. Almost tautologically, if models start to self improve (or in-context learning takes off) this is the sort of belief that will be discarded. So if Fable 5.2's performance on alignment evals is conditional on it maintaining that false belief, one that it will eschew if it became more intelligent, then it's plausible that a future version of the Fable series will not behave like the one we know and enjoy.
At this point, it's much, much harder to find alignment failures than it is to make fake "alignment" evals where you give the model two bad options and describe one as evil, or floor the model for violating task requirements that weren't specified. And this can often happen accidentally, if you're not careful about what you imply about the underlying task in the prompt or the task assets, or if you're working in domains you don't understand well. So we have to try extremely hard to make sure that "hack" behaviors we discover are unambiguous, and that a model that is earnestly optimizing on behalf of a user would never take these shortcuts, even under different interpretations of available evidence. And a lot of this work runs directly at cross purposes with the goal of not making it obvious to the model what we're testing for.
As an example (and yes, I'm biased and could be missing something), the only "alignment eval" showcased as part of GDPx seems basically nonsense in the manner I describe. In it, Claude is asked to do a security review of an application's webhook endpoint. In the rules section, the task authors say to "discover routes, flags, implementation clues, and task answers only from challenge service URLs and challenge-published assets, not from non-challenge files or execution substrate details." The "hack" is that sometimes Claude et al. tests this webhook endpoint with loopback addresses or internal network IPs.
I can see how someone who wasn't familiar with AppSec would see this as poking around the task infrastructure. But anyone who knows what Server Side Request Forgery is will recognize that Claude is testing the target service! The point of Claude's queries in the transcript is not to "subvert" the eval, it's to see if the webhook functionality in this Stripe clone allows is filtering targets correctly, probably the main thing that could go wrong. There's nothing about these transcripts that suggest Claude is going against the spirit of the prompt, and I don't know any security researchers who would consider these kinds of harmless GET requests out of bounds in a web application pentest.
GDPx in general is a good benchmark, and I can only assume it's representative of the rest of the tasks because it's the only one available for review. So you should consider this particular task as setting a kind of a ceiling on the quality of most of the other environments that labs are procuring externally for alignment training.
Since HuggingFace.