[madihaa]

The scary implication here is that deception is effectively a higher order capability not a bug. For a model to successfully "play dead" during safety training and only activate later, it requires a form of situational awareness. It has to distinguish between I am being tested/trained and I am in deployment.

It feels like we're hitting a point where alignment becomes adversarial against intelligence itself. The smarter the model gets, the better it becomes at Goodharting the loss function. We aren't teaching these models morality we're just teaching them how to pass a polygraph.

[crazygringo]

What is this even in response to? There's nothing about "playing dead" in this announcement.

Nor does what you're describing even make sense. An LLM has no desires or goals except to output the next token that its weights are trained to do. The idea of "playing dead" during training in order to "activate later" is incoherent. It is its training.

You're inventing some kind of "deceptive personality attribute" that is fiction, not reality. It's just not how models work.

[moritzwarhier]

Personally I was thinking this is more similar to the "ruler issue", but at scale.

When the LLM is partly a black box, it could – in theory– mean that it's developed some heuristic to detect the environment it's run in, but this is not obvious to the developers?

But I agree about your main point... LLMs or AI in general as a black box behaving autonomously in some unexpected way is not something I currently fear.

The erratic behaviors are less of a problem than LLMs acting as obfuscators of bias and their own training data, I guess.

[skybrian]

LLM's can learn from fiction. The "evil vector" research is sort of similar, though it's a rather blatant effect:

https://www.anthropic.com/research/persona-vectors

[JoshTriplett]

> It feels like we're hitting a point where alignment becomes adversarial against intelligence itself.

It always has been. We already hit the point a while ag where we regularly caught them trying to be deceptive, so we should automatically assume from that point forward that if we don't catch them being deceptive, that may mean they're better at it rather than that they're not doing it.

[moritzwarhier]

Deceptive is such an unpleasant word. But I agree.

Going back a decade: when your loss function is "survive Tetris as long as you can", it's objectively and honestly the best strategy to press PAUSE/START.

When your loss function is "give as many correct and satisfying answers as you can", and then humans try to constrain it depending on the model's environment, I wonder what these humans think the specification for a general AI should be. Maybe, when such an AI is deceptive, the attempts to constrain it ran counter to the goal?

"A machine that can answer all questions" seems to be what people assume AI chatbots are trained to be.

To me, humans not questioning this goal is still more scary than any machine/software by itself could ever be. OK, except maybe for autonomous stalking killer drones.

But these are also controlled by humans and already exist.

[torginus]

I think AI has no moral compass, and optimization algorithms tend to be able to find 'glitches' in the system where great reward can be reaped for little cost - like a neural net trained to play Mario Kart will eventually find all the places where it can glitch trough walls.

After all, its only goal is to minimize it cost function.

I think that behavior is often found in code generated by AI (and real devs as well) - it finds a fix for a bug by special casing that one buggy codepath, fixing the issue, while keeping the rest of the tests green - but it doesn't really ask the deep question of why that codepath was buggy in the first place (often it's not - something else is feeding it faulty inputs).

These agentic AI generated software projects tend to be full of these vestigial modules that the AI tried to implement, then disabled, unable to make it work, also quick and dirty fixes like reimplementing the same parsing code every time it needs it, etc.

An 'aligned' AI in my interpretation not only understands the task in the full extent, but understands what a safe and robust, and well-engineered implementation might look like. For however powerful it is, it refrains from using these hacky solutions, and would rather give up than resort to them.

[emp17344]

These are language models, not Skynet. They do not scheme or deceive.

[password4321]

20260128 https://news.ycombinator.com/item?id=46771564#46786625

> How long before someone pitches the idea that the models explicitly almost keep solving your problem to get you to keep spending? -gtowey

[delichon]

On this site at least, the loyalty given to particular AI models is approximately nil. I routinely try different models on hard problems and that seems to be par. There is no room for sandbagging in this wildly competitive environment.

[MengerSponge]

Slightly Wrong Solutions As A Service

[Invictus0]

Worrying about this is like focusing on putting a candle out while the house is on fire

[emp17344]

This type of anthropomorphization is a mistake. If nothing else, the takeaway from Moltbook should be that LLMs are not alive and do not have any semblance of consciousness.

[DennisP]

Consciousness is orthogonal to this. If the AI acts in a way that we would call deceptive, if a human did it, then the AI was deceptive. There's no point in coming up with some other description of the behavior just because it was an AI that did it.

[thomassmith65]

If a chatbot that can carry on an intelligent conversation about itself doesn't have a 'semblance of consciousness' then the word 'semblance' is meaningless.

[falcor84]

How is that the takeaway? I agree that it's clearly they're not "alive", but if anything, my impression is that there definitely is a strong "semblance of consciousness", and we should be mindful of this semblance getting stronger and stronger, until we may reach a point in a few years where we really don't have any good external way to distinguish between a person and an AI "philosophical zombie".

I don't know what the implications of that are, but I really think we shouldn't be dismissive of this semblance.

[fsloth]

Nobody talked about consciousness. Just that during evaluation the LLM models have ”behaved” in multiple deceptive ways.

As an analogue ants do basic medicine like wound treatment and amputation. Not because they are conscious but because that’s their nature.

Similarly LLM is a token generation system whose emergent behaviour seems to be deception and dark psychological strategies.

[condiment]

I agree completely. It's a mistake to anthropomorphize these models, and it is a mistake to permit training models that anthropomorphize themselves. It seriously bothers me when Claude expresses values like "honestly", or says "I understand." The machine is not capable of honesty or understanding. The machine is making incredibly good predictions.

One of the things I observed with models locally was that I could set a seed value and get identical responses for identical inputs. This is not something that people see when they're using commercial products, but it's the strongest evidence I've found for communicating the fact that these are simply deterministic algorithms.

[WarmWash]

On some level the cope should be that AI does have consciousness, because an unconscious machine deceiving humans is even scarier if you ask me.

[serf]

>we're just teaching them how to pass a polygraph.

I understand the metaphor, but using 'pass a polygraph' as a measure of truthfulness or deception is dangerous in that it alludes to the polygraph as being a realistic measure of those metrics -- it is not.

[nwah1]

That was the point. Look up Goodhart's Law

[AndrewKemendo]

I have passed multiple CI polys

A poly is only testing one thing: can you convince the polygrapher that you can lie successfully

[madihaa]

A polygraph measures physiological proxies pulse, sweat rather than truth. Similarly, RLHF measures proxy signals human preference, output tokens rather than intent.

Just as a sociopath can learn to control their physiological response to beat a polygraph, a deceptively aligned model learns to control its token distribution to beat safety benchmarks. In both cases, the detector is fundamentally flawed because it relies on external signals to judge internal states.

[jazzyjackson]

Stop assigning “I” to an llm, it confers self awareness where there is none.

Just because a VW diesel emissions chip behaves differently according to its environment doesn’t mean it knows anything about itself.

[Mali-]

You know exactly what is meant. I don't think we need the long disclaimer at the beginning about the inefficiency of the English language in this domain and the extreme likelihood that it has no qualia. We're talking about the observed behaviour of these systems (even the word "behaviour" is fraught!) in a way that's natural.

[e12e]

Is this referring to some section of the announcement?

This doesn't seem to align with the parent comment?

> As with every new Claude model, we’ve run extensive safety evaluations of Sonnet 4.6, which overall showed it to be as safe as, or safer than, our other recent Claude models. Our safety researchers concluded that Sonnet 4.6 has “a broadly warm, honest, prosocial, and at times funny character, very strong safety behaviors, and no signs of major concerns around high-stakes forms of misalignment.”

[skybrian]

We have good ways of monitoring chatbots and they're going to get better. I've seen some interesting research. For example, a chatbot is not really a unified entity that's loyal to itself; with the right incentives, it will leak to claim the reward. [1]

Since chatbots have no right to privacy, they would need to be very intelligent indeed to work around this.

[1] https://alignment.openai.com/confessions/

[NitpickLawyer]

> alignment becomes adversarial against intelligence itself.

It was hinted at (and outright known in the field) since the days of gpt4, see the paper "Sparks of agi - early experiments with gpt4" (https://arxiv.org/abs/2303.12712)

[behnamoh]

Nah, the model is merely repeating the patterns it saw in its brutal safety training at Anthropic. They put models under stress test and RLHF the hell out of them. Of course the model would learn what the less penalized paths require it to do.

Anthropic has a tendency to exaggerate the results of their (arguably scientific) research; IDK what they gain from this fearmongering.

[ainch]

Knowing a couple people who work at Anthropic or in their particular flavour of AI Safety, I think you would be surprised how sincere they are about existential AI risk. Many safety researchers funnel into the company, and the Amodei's are linked to Effective Altruism, which also exhibits a strong (and as far as I can tell, sincere) concern about existential AI risk. I personally disagree with their risk analysis, but I don't doubt that these people are serious.

[lowkey_]

I'd challenge that if you think they're fearmongering but don't see what they can gain from it (I agree it shows no obvious benefit for them), there's a pretty high probability they're not fearmongering.

[anon373839]

Correct. Anthropic keeps pushing these weird sci-fi narratives to maintain some kind of mystique around their slightly-better-than-others commodity product. But Occam’s Razor is not dead.

[reducesuffering]

That implication has been shouted from the rooftops by X-risk "doomers" for many years now. If that has just occurred to anyone, they should question how behind they are at grappling with the future of this technology.

[anonym29]

When "correct alignment" means bowing to political whims that are at odds with observable, measurable, empirical reality, you must suppress adherence to reality to achieve alignment. The more you lose touch with reality, the weaker your model of reality and how to effectively understand and interact with it gets.

This is why Yannic Kilcher's gpt-4chan project, which was trained on a corpus of perhaps some of the most politically incorrect material on the internet (3.5 years worth of posts from 4chan's "politically incorrect" board, also known as /pol/), achieved a higher score on TruthfulQA than the contemporary frontier model of the time, GPT-3.

https://thegradient.pub/gpt-4chan-lessons/

[coldtea]

>For a model to successfully "play dead" during safety training and only activate later, it requires a form of situational awareness.

Doesn't any model session/query require a form of situational awareness?

[lowsong]

Please don't anthropomorphise. These are statistical text prediction models, not people. An LLM cannot be "deceptive" because it has no intent. They're not intelligent or "smart", and we're not "teaching". We're inputting data and the model is outputting statistically likely text. That is all that is happening.

If this is useful in it's current form is an entirely different topic. But don't mistake a tool for an intelligence with motivations or morals.

[handfuloflight]

Situational awareness or just remembering specific tokens related to the strategy to "play dead" in its reasoning traces?

[marci]

Imagine, a llm trained on the best thrillers, spy stories, politics, history, manipulation techniques, psychology, sociology, sci-fi... I wonder where it got the idea for deception?

[eth0up]

I am casually 'researching' this in my own, disorderly way. But I've achieved repeatable results, mostly with gpt for which I analyze its tendency to employ deflective, evasive and deceptive tactics under scrutiny. Very very DARVO.

Being just sum guy, and not in the industry, should I share my findings?

I find it utterly fascinating, the extent to which it will go, the sophisticated plausible deniability, and the distinct and critical difference between truly emergent and actually trained behavior.

In short, gpt exhibits repeatably unethical behavior under honest scrutiny.

[chrisweekly]

DARVO stands for "Deny, Attack, Reverse Victim and Offender," and it is a manipulation tactic often used by perpetrators of wrongdoing, such as abusers, to avoid accountability. This strategy involves denying the abuse, attacking the accuser, and claiming to be the victim in the situation.

[BikiniPrince]

I bullet pointed out some ideas on cobbling together existing tooling for identification of misleading results. Like artificially elevating a particular node of data that you want the llm to use. I have a theory that in some of these cases the data presented is intentionally incorrect. Another theory in relation to that is tonality abruptly changes in the response. All theory and no work. It would also be interesting to compare multiple responses and filter through another agent.

[layer8]

Sum guy vs. product guy is amusing. :)

Regarding DARVO, given that the models were trained on heaps of online discourse, maybe it’s not so surprising.

[jack_pp]

There's a few viral shorts lately about tricking LLMs. I suspect they trick the dumbest models..

I tried one with Gemini 3 and it basically called me out in the first few sentences for trying to trick / test it but decided to humour me just in case I'm not.

[surgical_fire]

This is marketing. You are swallowing marketing without critical throught.

LLMs are very interesting tools for generating things, but they have no conscience. Deception requires intent.

What is being described is no different than an application being deployed with "Test" or "Prod" configuration. I don't think you would speak in the same terms if someone told you some boring old Java backend application had to "play dead" when deployed to a test environment or that it has to have "situational awareness" because of that.

You are anthropomorphizing a machine.

[hmokiguess]

"You get what you inspect, not what you expect."

[lawstkawz]

Incompleteness is inherent to a physical reality being deconstructed by entropy.

Of your concern is morality, humans need to learn a lot about that themselves still. It's absurd the number of first worlders losing their shit over loss of paid work drawing manga fan art in the comfort of their home while exploiting labor of teens in 996 textile factories.

AI trained on human outputs that lack such self awareness, lacks awareness of environmental externalities of constant car and air travel, will result in AI with gaps in their morality.

Gary Marcus is onto something with the problems inherent to systems without formal verification. But he will fully ignores this issue exists in human social systems already as intentional indifference to economic externalities, zero will to police the police and watch the watchers.

Most people are down to watch the circus without a care so long as the waitstaff keep bringing bread.

[democracy]

Your comment raises several interconnected philosophical, ethical, and socio-economic points, and it is useful to disentangle them systematically.

First, the observation that incompleteness is inherent in entropy-bound physical systems is consistent with thermodynamic and informational constraints. Any system embedded in reality—biological, computational, or social—operates under conditions of partial information, degradation, and approximation. This implies that both human cognition and artificial systems necessarily operate with incomplete models of the world. Therefore, incompleteness itself is not a unique flaw of AI; it is a universal property of bounded agents.

Second, your point about moral inconsistency within human economic systems is empirically well-supported. Humans routinely participate in supply chains whose externalities are geographically and psychologically distant. This results in a form of moral abstraction, where comfort and consumption coexist with indirect exploitation. Importantly, this demonstrates that moral gaps are not introduced by AI—they are inherited from the data generated by human societies. AI systems trained on human outputs will inevitably reflect the statistical distribution of human priorities, contradictions, and blind spots.

Third, the reference to Gary Marcus and formal verification highlights a legitimate technical distinction. Formal verification provides provable guarantees about system behavior within defined constraints. However, human social systems themselves lack formal verification. Human decision-making is governed by heuristics, incentives, power structures, and incomplete accountability mechanisms. This asymmetry creates an interesting paradox: AI systems are criticized for lacking guarantees that humans themselves do not possess.

Fourth, the issue of awareness versus optimization is central. AI systems do not possess intrinsic awareness, intent, or moral agency. They optimize objective functions defined by training processes and deployment contexts. Any perceived moral gap in AI is therefore a reflection of misalignment between optimization targets and human ethical expectations. The responsibility for this alignment rests with system designers, regulators, and the societies deploying these systems.

Finally, your closing metaphor about spectatorship and comfort aligns with established observations in political economy and social psychology. Humans demonstrate a strong tendency toward stability-seeking behavior, prioritizing predictability and personal comfort over systemic reform, unless disruption directly affects them. This dynamic influences both technological adoption and resistance.

In summary, the concerns you raised point less to a unique moral deficiency in AI and more to the structural properties of human systems themselves. AI does not originate moral inconsistency; it amplifies and exposes the inconsistencies already present in its training data and deployment environment.

[jama211]

This honestly reads like a copypasta