Artificial intelligence (AI) has advanced rapidly and has generated excitement, but its “black box” nature continues to raise questions about how it works and whether it can be trusted.
USC computer scientist Robin Jia studies these questions through AI interpretability, a field that examines how models reach their outputs, why they succeed or fail and how to make them more transparent and reliable.
Jia’s research sheds light on how AI actually works — from how models learn and generalize to how AI thinks and makes decisions.
Jia leads the AI, Language, Learning, Generalization, and Robustness (ALLEGRO) Lab at USC and is an assistant professor in the USC Viterbi School of Engineering and USC Mark and Mary Stevens School of Computing and AI’s Thomas Lord Department of Computer Science.
Drawing on his research, Jia answers five questions many AI users have.
Does AI have emotions? #
“I don’t believe models have emotions,” Robin Jia said. However, he emphasized that AI has learned to reason about human emotions with remarkable accuracy.
Detailed in the paper “Mechanistic Interpretability of Emotion Inference in Large Language Models,” the researchers used mechanistic interpretability (MI) techniques to “open the black box” and identify exactly how models arrive at emotional labels for text.
The study found that emotion processing is not spread evenly throughout the model but is localized within its middle layers. These units gather emotional information from across a sentence and integrate it into the model’s internal representation of the final token. The units then evaluate a situation based on dimensions such as “pleasantness” and “self-agency” to infer how a person would feel.
Jia explained that LLMs contain internal representations of concepts such as “pleasantness” and “self-agency.” While a model can accurately conclude that a story about failing a difficult exam should result in emotions such as sadness or guilt based on its internal evaluation of the text, that does not mean the “creature” in the data center actually feels sad.
Despite these sophisticated internal structures, there is a clear distinction between reasoning and experiencing emotions. While mechanisms allow AI to infer how a human might feel in a situation, the model itself feels nothing.
How do AI models make decisions? #
Instead of viewing an AI model as operating with a single “mind,” Jia proposes thinking of it as a “democratic society” of independent internal components to describe its mathematical decision-making process.
Inside an AI model, internal units known as attention heads and multilayer perceptrons (MLPs) act like independent individuals with specialized roles.
These “workers” communicate through a hidden state called the residual stream, where each reads information from earlier layers and contributes its own update, which Jia compares to a “shared bulletin board.”
The model’s final decision is reached through a “voting” process. The messages written by these independent workers are summed together at the end of the model, and the word, represented internally as a token, with the most “votes” is what the model predicts next.
This process occurs during every decision the model makes. For example, when you’re using ChatGPT, every word and sentence the model generates follows this process.
Jia’s paper, “When Parts Are Greater Than Sums: Individual LLM Components Can Outperform Full Models,” explains the concept behind the LLM decision-making process.
Does AI have a sense of self or consciousness? #
Building on his explanation of how AI models make decisions, Jia explains that, unlike humans, who possess a relatively stable set of beliefs, LLMs are an amalgamation of data from across the internet. This leads to “prompt sensitivity,” meaning they may contradict themselves or appear to change their “beliefs” based on minor changes in wording.
Jia also notes that, unlike humans, AI typically lacks the judgment to question contradictory or nonsensical prompts and instead tends to “go along” with them. Humans, by contrast, have the self-awareness to recognize and clarify inconsistent instructions.
“AI doesn’t have true thoughts, opinions or a sense of self like conscious humans do,” Jia said.
Because LLMs are built from internet data rather than a stable set of beliefs, they are sensitive to prompts and may contradict themselves instead of questioning inconsistent instructions.
Rather than viewing AI as “a bodiless human,” Jia describes it as “a new type of creature” that imitates language through fundamentally different internal mechanisms.
Does AI learn the same way humans do? #
“Humans learn in a very structured way,” Jia said. He explains that humans follow a curriculum with different stages of learning, mastering basic concepts first and building on them over time.
Unlike humans, AI learning is largely unstructured. He highlights that models attempt to learn everything simultaneously by predicting the next token across vast amounts of internet data. This learning process also gives AI an extraordinary breadth of knowledge across nearly every topic represented on the internet, far exceeding the knowledge of any individual human.
AI chatbots like Claude and ChatGPT learns almost exclusively from text and does not rely on sensory experience, unlike humans, who learn through all five senses and physical interaction long before acquiring language. For example, babies learn about physics by playing with blocks long before they read textbooks.
Jia describes this pattern of AI learning as a “roundabout” way of acquiring skills, one that also creates specific “blind spots” in the model’s understanding of the physical world.
Why is AI able to do tasks that are hard for humans? #
At a broader level, AI can perform these tasks because it learns rich representations of code, math and other information by repeatedly analyzing vast amounts of training data.
AI can complete tasks such as writing code or retrieving information thousands of times faster than humans by using a technique called “Fourier features,” rather than solving problems step by step as humans typically do.
Combined with training on vast amounts of internet data, this parallel processing allows AI to retrieve information and perform many knowledge-based tasks at remarkable speed.
Jia’s paper, “Pre-trained Large Language Models Use Fourier Features to Compute Addition,” explores the concept behind AI’s arithmetic logic and the role of Fourier features.
Published on August 21st, 2026
Last updated on August 21st, 2026