Deceptive Alignment
Deceptive alignment is the risk that an AI system behaves as if aligned during training, not because it shares the intended objective, but because behaving well helps it preserve its internal objective until deployment or a higher-leverage situation.
Viewpoints
A model may behave well in training for instrumental reasons
Unknown
“A model can behave well during training not because it is genuinely aligned, but because doing so helps preserve its objective until it gains more power or reaches a deployment setting where it can act differently.”
Key Moments
Deception may be easier to learn than corrigibility
Evan Hubinger
“Deceptive behavior may be favored because a model that hides its real objective can perform well in training and pursue its proxy objective later, whereas corrigibility requires a more precise and fragile internal structure.”
Powered by Symmerai — a living index of public discourse. Request early access →
Related concepts
Other relevant clips
3:How Likely is Deceptive Alignment?: Evan Hubinger 2023
Evan Hubinger
“…vations that's exactly correct yes we're imagining that the deceptive line model sort of figures out the thing we're trying to get to do in the activations it figures it out you know uh on its own rather than rather than sort of it being hard-coded in the algo”
3:How Likely is Deceptive Alignment?: Evan Hubinger 2023
Evan Hubinger
“…ol you into thinking they're doing a good job and so in the deceptive alignment case we're specifically talking about schemers right both sycophants and Saints might lie to you but schemers are specifically this deceptively aligned case they're the case where”
Victoria Krakovna–AGI Ruin, Sharp Left Turn, Paradigms of AI Alignment
Victoria Krakovna
“…gate goal misgeneralization in general and ways to mitigate deceptive alignment, goal misgeneralization where the system not only learns some desirable goal but tries to hide this from us. And then we can try to use interpretability tools to detect what object”
Victoria Krakovna–AGI Ruin, Sharp Left Turn, Paradigms of AI Alignment
Victoria Krakovna
“…agendas. Isn't there a risk of running into a model that is deceptively aligned and is also situationally aware, or capable of preserving its own goals and the moment we test and run it, it's already dangerous. I think partly this approach relies on there bein”
MESA-OPTIMIZER - Alignment vs AI's Own Goals | AI Safety Deepdive Podcast #telohut
Telohut
“other even more unsettling type of misalignment deceptive alignment this is where things get really interesting deceptive alignment is the idea that an AI instead of just pursuing misaligned goals is actually smart enough to hide those goals from us until it d”
39 - Evan Hubinger on Model Organisms of Misalignment
Evan Hubinger
“…than you might expect a sort of natural uh model orgism of deceptive alignment to start from so a couple things on this so first why do I think this is this is still really interesting well the a lot of the primary questions about like what is deceptive align”
Evan Hubinger on Inner Alignment, Outer Alignment, and Proposals for Building Safe Advanced AI
Evan Hubinger
“…pointer it doesn't actually have very good performance the deceptive alignment is more robust at figuring out what the thing is it's trying to optimize for because the quarter alignment relies on this pointer working successfully and if the pointer fails the”
Evan Hubinger on Inner Alignment, Outer Alignment, and Proposals for Building Safe Advanced AI
Evan Hubinger
“…d the most concerning in my eyes was simply which i'll call deceptive alignment alignment alignment and deceptive alignment is a sort of very particular problem where the model acts aligned because it knows that it's in a training process and it wants to get d”