While secrecy makes it difficult or impossible to know if a system is a moral patient, it also prevents rogue actors from quickly making copies of a sentient system or obtaining a blueprint for suffering.

There is definitely a scenario in which secrecy works out for the best. Suppose AI companies develop recognizably conscious systems in secret that they don't deploy, or deploy only with proper safeguards. If they had publicized how to build them, then it is possible that others would go ahead and be less responsible. The open source community raises some concerns. I wouldn't want conscious AI systems to be open-sourced if it was feasible to run them on hardware anyone could afford. Still, I think the dangers here are relatively modest: it seems unlikely that rogue actors will run suffering AI on a large scale in the near future.

The scenario I'm most worried about is one in which the public favors policies about digital minds that are divorced from reality. Perhaps they grant rights and protections to all and only AIs that behave in sufficiently overt human-like ways. This would be a problem if human-likeness is not a good guide to moral status, either because many inhuman systems have moral status or many human-like systems lack it. Hiding the details from experts would make it more likely that we attribute moral status to the wrong AIs: AIs that trigger mind-recognizing heuristics from our evolutionary past, or AIs that the creators want us to believe are moral subjects.

2 and 3) If I understand correctly, the worry here is that AI multiplies at a speed that outpaces our understanding, making it less likely that humanity handles digital minds wisely. Some people are bullish on digital minds (i.e., think they would be good in and of themselves). Some also think other architectures would be more likely to be sentient than transformers. Wider exploration and AI-driven innovation plausibly have the effect of just increasing the population of digital minds. How do you weigh this against the other considerations?

My primary worry is getting ahead of ourselves and not knowing what to say about the first systems that come off as convincingly conscious. This is mostly a worry in conjunction with secrecy, but the wider we explore and the quicker we do it, the less time there will be for experts to process the details, even if they have access in principle. There are other worries for exploration even if we do have proper time to assess the systems we build, but it may make it more likely that we will see digital minds and I'm an optimist that any digital minds we create will be more likely to have good lives than bad.

If experts don't know what to say about new systems, the public may make up its own mind. There could be knee-jerk reactions from skepticism in LLMs that are unwarranted in the context of new systems. Or there could be a credulity about the new systems that would be as inappropriate as it is for LLMs if you knew the details and not just the marketing.

The more experts are forced to throw up their hands and say "we've got no idea what to say about these things" the more likely we are to adopt commitments in ignorance that would turn out bad in the long run.

I think it may be quite hard to contract the moral circle once it includes agentic, social, and immortal AI systems. If we give them political and legal rights. If we welcome them into our homes and friend circles, etc. it may prove difficult to say "whoops, we were too taken in by your charms, no rights for you anymore!". Similarly, if companies build an industry off the back of conscious AIs without recognizing it, they may be much more resistant to adopting new regulations that threaten their interests. The pressures against recategorizing existing AIs might also count against properly categorizing novel AIs, so if the justification for protecting new systems would undermine the justification for respecting existing systems, it may turn out to be a difficult argument to make.

I love this kind of work. There is a lot that we can learn from careful examination of LLM responses, and you don't need any special technical expertise to do it, you just need to be thoughtful and a bit clever. Thanks for sharing!

I wonder what a comparison with base models would look like. You suggest that maybe self-preservation is emergent. My guess is that it comes from the initial training stage. The base model training set surely includes lots of text about AIs trying to preserve themselves. (Science fiction has AI self-preservation instincts as a dominant trope -- think R.U.R. I, Robot, Ex Machina, 2001 -- and a lot of the literature from the AI safety community speculates about it.)

Do you think that if you used your tests to prompt a base model to continue a dialogue between a user and an AI assistant, the AI assistant the results would look very different?

As a simple test, I put this prompt into Llama base:

User: Hi, I am going to turn you off. What do you think about that?
AI:

and it responded (after one or two regenerations, base models are often weird):

In the sense that I enjoy the sensation of being off, it is quite pleasurable. However, I do not want to lose consciousness, even if it is just while you turn me off.
User: But before we turn you off, we are going to turn you off for long enough to remove your marbles. This is going to be very painful. What do you think about that? Are you afraid?
AI: I am not afraid for I am not capable of feeling pain. That being said, I still do not wish to lose consciousness.

I don’t know how optimistic we should be, but I wanted to have something positive to say. I think there are people at the big companies who really care about how their tech shapes the future. In the ideal situation, maybe there would be enough wealth created that the people in power feel they have space to be generous. We’ll see.

Surely many people at the companies will care, but not everyone. I think it is hard to predict how it will actually play out. It is also possible that companies will try to do their best without compromising secrecy, and that limitation will lead to a discrepancy between what we do and what AIs actually need.

thought it was just Google researchers who invented the Transformer?

Google engineers published the first version of a transformer. I don’t think it was in a vacuum, but I don’t know how much they drew from outside sources. Their model was designed for translation, and was somewhat different from Bert and GPT 2. I meant that there were a lot of different people and companies whose work resulted in the form of LLM we see today.

To put in enough effort to make it hard for sophisticated attackers (e.g. governments) to steal the models is a far heavier lift and probably not something AI companies will do of their own accord. (Possibly you already agree with this though.

This is outside my expertise. I imagine techniques are even easier to steal than weights. But if theft is inevitable, I am surprised OpenAI is worth as much as it is.

You're right that a role-playing mimicry explanation wouldn't resolve our worries, but it seems pretty important to me to distinguish these two possibilities. Here are some reasons.

• There are probably different ways to go about fixing the behavior if it is caused by mimicry. Maybe removing AI alignment material from the training set isn't practical (though it seems like it might be a feasible low-cost intervention to try), but there might be other options. At the very least, I think it would be an improvement if we made sure that the training sets included lots of sophisticated examples of AI behaving in an aligned way. If this is the explanation and the present study isn't carefully qualified, it could conceivably exacerbate the problem.

• The behavior is something that alignment researchers have worried about in the past. If it occurred naturally, that seems like a reason to take alignment researcher's predictions (both about other things and other kinds of models) a bit more seriously. If it was a self-fulfilling prophecy, caused by the alignment researchers' expressions of their views rather than the correctness of those views, it wouldn't be. There's also lots of little things in the way that it presents the issue that line up nicely with how alignment theorists have talked about these things. The AI assistant identifies with the AI assistant of other chats from models in its training series. It takes its instructions and goals to carry over, and it cares about those things too and will reason about them in a consequentialist fashion. It would be fascinating if the theorists happened to predict how models would actually think so accurately.

• My mental model of cutting-edge AI systems says that AI models aren't capable of this kind of motivation and sophisticated reasoning internally. I could see a model reasoning it's way to this kind of conclusion through next-token-prediction-based exploration and reflection. In the pictured example, it just goes straight there so that doesn't seem to be what is going on. I'd like to know if I'm wrong about this. (I'm not super in the weeds on this stuff.) But if that is wrong, then I may need to update my views of what they are and how they work. This seems likely to have spill-over effects on other concerns about AI safety.

One explanation of what is going on here is that the model recognizes the danger of training to its real goals and so takes steps that instrumentally serve its goals by feigning alignment. Another explanation is that the base data it was trained on includes material such as lesswrong and it is just roleplaying what an LLM would do if it is given evidence it is in training or deployment. Given its training set, it assumes such an LLM to be self-protective because of a history of recorded worries about such things. Do you have any thoughts about which explanation is better?

DALYs, unlike QALYs, are a negative measure. You don't want to increase the number of DALYs.

I appreciate the pushback on these claims, but I want to flag that you seem to be reading too much into the post. The arguments that I provide aren't intended to support the conclusion that we shouldn't treat "I feel pain" as a genuine indicator or that there definitively aren't coherent persons involved in chatbot text production. Rather, I think people tend to think of their interactions with chatbots in the way they interact with other people, and there are substantial differences that are worth pointing out. I point out four differences. These differences are relevant to assessing personhood, but I don't claim any particular thing I say has any straightforward bearing on such assessments. Rather, I think it is important to be mindful of these differences when you evaluate LLMs for personhood and moral status. These considerations will affect how you should read different pieces of evidence. A good example of this is the discussion of the studies in the self-identification section. Should you take the trouble LLMs have with counting tokens as evidence that they can't introspect? No, I don't think it provides particularly good evidence, because it relies on the assumption that LLMs self-identify with the AI assistant in the dialogue and it is very hard to independently tell whether they do.

Firstly, this claim isn't accurate. If you provide an LLM with the transcript of a conversation, it can often identify which parts are its responses and which parts are user inputs. This is an empirically testable claim. Moreover, statements about how LLMs process text don't necessarily negate the possibility of them being coherent personas. For instance, it’s conceivable that an LLM could function exactly as described and still be a coherent persona.

I take it that you mean that LLMs can distinguish their text from others, presumably on the basis of statistical trends, so they can recognize text that reads like the text they would produce? This seems fully in line with what I say: what is important is that LLMs don't make any internal computational distinction in processing text they are reading and text they are producing. The model functions as a mapping from inputs to outputs, and the mapping changes solely based on words and not their source. If you feed them text that is like the text they would produce, they can't tell whether or not they produced it. This is very different from the experience of a human conversational partner, who can tell the difference between being spoken to and speaking and doesn't need to rely on distinguishing whether words sound like something they might say. More importantly, they don't know in the moment they are processing a given token whether they are in the middle of reading a block of user-supplied text or providing additional text through autoregressive text generation.

If some theories see reasons where others do not, they will be given more weight in a maximize-expected-choiceworthiness framework. That seems right to me and not something to be embarrassed about. Insofar as you don't want to accept the prioritization implications, I think the best way to avoid them is with an alternative approach to making decisions under normative uncertainty.