Crossposted from the AI Alignment Forum. May contain more technical jargon than usual.
This is a linkpost for

TLDR: In a new paper, we explore whether we could train future LLMs to accurately answer questions about themselves. If this works, LLM self-reports may help us test them for morally relevant states like consciousness.

We think it's possible to start preliminary experiments testing for moral status in language models now, so if you're interested in working with us, please reach out and/or apply to the Astra fellowship or SERI MATS (deadline November 17).

Tweet thread paper summary: Link


As AI systems become more advanced and widely deployed, there will likely be increasing debate over whether AI systems could have conscious experiences, desires, or other states of potential moral significance. It is important to inform these discussions with empirical evidence to the extent possible. We argue that under the right circumstances, self-reports, or an AI system’s statements about its own internal states, could provide an avenue for investigating whether AI systems have states of moral significance. Self-reports are the main way such states are assessed in humans ("Are you in pain?"), but self-reports from current systems like large language models are spurious for many reasons (e.g. often just reflecting what humans would say). To make self-reports more appropriate for this purpose, we propose to train models to answer many kinds of questions about themselves with known answers, while avoiding or limiting training incentives that bias self-reports. The hope of this approach is that models will develop introspection-like capabilities, and that these capabilities will generalize to questions about states of moral significance. We then propose methods for assessing the extent to which these techniques have succeeded: evaluating self-report consistency across contexts and between similar models, measuring the confidence and resilience of models’ self-reports, and using interpretability to corroborate self-reports. We also discuss challenges for our approach, from philosophical difficulties in interpreting self-reports to technical reasons why our proposal might fail. We hope our discussion inspires philosophers and AI researchers to criticize and improve our proposed methodology, as well as to run experiments to test whether self-reports can be made reliable enough to provide information about states of moral significance.

See also this earlier post for more discussion on the relevance to alignment (AI systems that are suffering might be more likely to take catastrophic actions), as well as some initial criticisms of a preliminary version of our proposed test in the comments. We've added a significant amount of content to our updated proposal to help to address several reservations people had on why the initial proposal might not work, so we're excited to get additional feedback and criticism on the latest version of our proposal as well.

New Comment
3 comments, sorted by Click to highlight new comments since:

What you are doing is training the AI to have an accurate model of itself, used with language like "I" and "you". You can use your brain to figure out what will happen if you ask "are you conscious?" without having previously trained in any position on similarly nebulous questions. Training text was written overwhelmingly by conscious things, so maybe it says yes because that's so favored by the training distribution. Or maybe you trained it to answer "you" questions as about nonfiction computer hardware and it makes the association that nonfiction computer hardware is rarely conscious.

Basically, I don't think you can start out confused about consciousness and cheat by "just asking it." You'll still be confused about consciousness and the answer won't be useful.

I'm worried this is going to lead, either directly or indirectly, to training foundation models to have situational awareness, which we shouldn't be doing.

And perhaps you should be worried that having an accurate model of onesself, associated with language like "I" and "you", is in fact one of the ingredients in human consciousness, and maybe we shouldn't be making AIs more conscious.

Some food for thought:

A -> Nature of Self-Reports in Cognitive Science: In cognitive science, self-reports are a widely used tool for understanding human cognition and consciousness. Training AI models to use self-reports (in an ideal scenario, this is analogous to giving a microphone, not giving the singer) does not inherently imply they become conscious. Instead, it provides a framework to study how AI systems represent and process information about themselves, which is crucial for understanding their limitations and capabilities.

B -> Significance of Linguistic Cues: The use of personal pronouns like "I" and "you" in AI responses is more about exploring AI's ability to model relational and subjective experiences than about inducing consciousness. These linguistic cues are essential in cognitive science (if we were to view LLMs as a semi-accurate model of how intelligence works) for understanding perspective-taking and self-other differentiation, which are key areas of study in human cognition.  Also, considering that some research points to that a certain degree of self-other overlap is necessary for truly altruistic behaviors, tackling this self-other issue can be an important stepping stone to developing an altruistic AGI. In the end, what we want is AGI, not some statistical, language-spitting, automatic writer.

C -> Ethical Implications and Safety Measures: The concerns about situational awareness and inadvertently creating consciousness in AI are valid. However, the paper's proposal involves numerous safety measures and ethical considerations. The focus is on controlled experiments to understand AI's self-modeling capabilities, not on indiscriminately pushing the boundaries of AI consciousness.

Sorry for commenting twice, and I think this second one might be a little out of context (but I think it makes a constructive contribution to this discussion). 

I think we must make sure that we are working on the "easy problems" of consciousness. This portion of consciousness has a relatively well-established philosophical explanation. For example, the Global Workspace Theory provides a good high-level interpretation of human consciousness. It proposes a cognitive architecture to explain consciousness. It suggests that consciousness operates like a "global workspace" in the brain, where various neural processes compete for attention. The information that wins this competition is broadcast globally, becoming accessible to multiple cognitive processes and entering conscious awareness. This theory addresses the question of how and why certain neural processes become part of conscious experience while others remain subconscious. The theory posits that through competitive and integrative mechanisms, specific information dominates our conscious experience, integrating different neural processes into a unified conscious experience. 

However, the Global Workspace Theory primarily addresses the functional and mechanistic aspects of consciousness, often referred to as the "Easy Problems" of consciousness. These include understanding how cognitive functions like perception, memory, and decision-making become conscious experiences. However, the Hard Problem of Consciousness, which asks why and how these processes give rise to subjective experiences or qualia, remains largely unaddressed by GWT. The Hard Problem delves into the intrinsic nature of consciousness, questioning why certain brain processes are accompanied by subjective experiences. While GWT offers insights into the dissemination and integration of information in the brain, it doesn't explain why these processes lead to subjective experience, leaving the Hard Problem essentially unresolved.

Until we have a good high-level philosophical foundation for this Hard Problem, it might be a good approach to draw the line between the two and work on the easy problems first.

Bringing home the point: 1. That is, for now, it will be extremely difficult for us to figure out whether LLMs (or any other human beings for the matter) have “phenomenal” or “subjective” dimensions consciousness or not. 2. Rather, focus on the easy, reductively-explainable dimensions of consciousness first. 3. We should make clear distinctions of these two categories when talking about consciousness