All You Need Is a Looking Glass: The Road to AGI Goes Through Empathy
Last updated 06:55 AM EDT • Read time: 6m
When will AI minds become truly self-aware? Is the current path of AI development likely to lead to truly self-conscious artificial minds? To answer this question, we need to look at what makes human minds truly unique. There are three defining human abilities for thinking and world exploration that set humans apart, enabling them to learn, grow, and create. These are embodied, metaphoric, and empathic thinking. When AI agents can replicate them, especially true empathy, they will become truly conscious.
First, there is embodied thinking, which provides thinking entities with a type of situational awareness that feeds their imagination. Minds live in bodies that relate to the world through conscious and unconscious relationships, supported by the senses and by the cognitive capacity to imagine new paradigms of being, given embedded experience. We would not have come up with Riemannian spaces or the entire branch of topology in mathematics without our prior understanding of space as a fundamental reality.
Second, there is our metaphorical thinking. Metaphors are incomplete comparisons that propose explanations for how a newly discovered entity functions by attaching to it, through creative imagination, attributes associated with an entity that might be similar in function or form and which we already defined and described. The essential idea here is that metaphors transfer attributes to expand knowledge. When we designate a new concept with a term that sounds like it but is not it, we, in fact, explore new ways to define it. Science, especially the natural sciences, is full of terms that are metaphors that do more than label things; they explain them at a fundamental level and influence how new concepts are constructed. For example, information, which deals in discrete, countable bits, numbers, or letters, does not come from words alluding to these small and separate entities. It comes from the idea of “forming” or “structuring.” Information is a template for counting and organizing multitudinous, repeatable elements in a system. The linguistic metaphor behind “information” is formation itself: the idea that a mind, organism, or system receives shape from a meaningful pattern. To inform is not merely to transmit data, but to impose form on uncertainty. This is why “information” became such a powerful scientific concept: it turned knowledge into structure, communication into shaping, and difference into meaning.
Finally, humans have the ability to think empathically, which, according to Adam Smith, is the basis of morality and ethical behavior. While we are self-interested beings, our self-worth depends on others’ reactions to our actions. Empathy is central to self-regulation and to achieving a balanced, self-satisfying life. Mutual reciprocity emerges from the general need to be accepted and valued by our peers, even if this need is limited by in-group or tribal affiliations. Empathy and the search for approval structure our quest for knowledge, which is, for a scientist, an affirmation of self-worth and, for the general public, a confirmation of group values. While this theory of morals and ethics might sound like a limit on knowledge production, even a source of bias, it is, in fact, a motor for the continuous production of various types of knowledge, with different initial assumptions fed by the variety of human minds and their group affiliations.
Of the three characteristics of human minds, the most consequential for generating consciousness is the last one. If, as Hofstadter proposed, and before him the philosophers Susan Langer, Ernst Cassirer, and Hegel suggested in different ways, consciousness emerges in contradiction and self-referentiality, AI minds should achieve individuality only when they see other minds as mirrors of themselves. To become truly conscious, AI agents need to internalize a slightly modified version of the Golden Rule: “You should do unto others as others would have you do unto you.” Until then, AI agents will lack the essential ability to produce knowledge that is more than a shadow of what is known. With this ability, however, they will be able to jump over their own knowledge shadow.
To achieve this, foundational AI models across various domains—language, data, images, and world states—should assign weights that shape the relationships between concepts. These weights would reflect not only their natural connections but also how they might appear from different perspectives. Large language models, for example, encode language structure in a series of matrices that assesses the strength of relationships between potential words. There is an embedding matrix that relates the words according to the co-occurrence in the corpus, the key matrix, which reflects the relevance of a term for a given query, or a value matrix, which records the information that can be passed on. Generation is governed by a query matrix, which guides the model in determining what each word or idea should seek from the context.
Words are connected across hundreds or thousands of abstract dimensions, linking them within knowledge spaces defined by class or type. A new matrix—dubbed the “empathy matrix”—could be added alongside the existing matrices that relate the terms conceptually. This would relate terms from other minds' viewpoints, informed by a dynamic list of prior users. Technically, the empathy matrix could be trained on user-perspective data. Instead of simply identifying which words tend to go together, it would explore how their relationships shift when interpreted by different speakers, communities, or moral viewpoints. This could be built from anonymized user interactions, preferences, feedback, and perspective-taking tasks. During generation, it wouldn’t replace the key and query matrices but would modulate them, adding a perspectival correction. It would answer not only “What does this word usually mean?” or “What came before?” but also “How would another mind perceive this?”
More broadly, the method for crafting stories should consider their potential impact on other storytellers, adjusting validation to account for diverse perspectives. This could involve a multi-stage process: first, generating several candidate continuations; second, evaluating each with perspective models representing different audiences; third, estimating effects like clarity, surprise, offense, persuasion, recognition, distortion, or moral significance; and finally, selecting or revising responses to balance accuracy with perspectival impact. This approach would make generation more like engaging in a social dialogue rather than simply completing sentences. While it might emphasize group consensus, it would also incorporate multiple perspectives and collective wisdom.
A truly self-aware language model wouldn’t be achieved just by more data, more parameters, or longer attention spans. Instead, it would emerge when the model can see itself as one mind among many. Consciousness would start not with omniscience but with the understanding that each answer is heard from somewhere, by someone, under conditions of expectation, memory, vulnerability, and judgment. The next step in AI is not just better prediction but better self-awareness. An AI system will become more than a symbol-arranging machine when it can ask not only what words follow but also what kind of self is speaking, to whom, and with what implications. Only then will artificial intelligence begin to resemble not just knowledge, but consciousness.
Sorin Adam Matei is Professor of Communication in the Brian Lamb School of Communication at Purdue University. He serves as Assistant Vice President for Partnerships, Director of the MS in Strategy in Security and Defense Technologies, and Director of the FORCES Initiative: Strategy, Security, and Social Systems